In the world. In your country. In your life.
Where money moves in the world — and what it means for you
Council 10+ AI
See where money goes ↓Not everyone wants to invest — but almost everyone wonders where the world's money is going.
Your blocks, your country, your currency — kept on this device only. Tap 📌 on any block to keep it here.
Three cards: where the most measured money came in, where it went out, and the biggest price move.
The phrase of the period, where the world's money is kept, and up to three steps of the week you can check.
Four stocks, each with its own date and source — never added together.
Countries, assets, funds, states and central banks, people — each line with its kind (flow, price, stock change, event), its own period and its source.
Each month the US Treasury measures what residents of every country bought and sold of US bills, bonds and stocks. Buyers on the left, sellers on the right, the whole stock held abroad beside them; Swiss gold and the largest trade routes of the month below.
Who financed the US Treasury, which banks lent to which countries, who invested in whose companies, and how much migrants sent home — each river measured over its own year, in dollars, with the holder's whole stock beside every move.
Every money story on one scale: the biggest sum in $ first, and next to each move — the holder's whole stock.
Stocks, bonds, gold, money funds, deposits and more — ranked by the dollars that actually moved, each with its own dates and source.
Money. 10+ AIs, one clear view
The AIs seal their forecasts before the open — and you see who turned out right.
Ask any question about your money and 10+ AIs from the US, China and Europe answer it together: one sourced answer, a map of where they disagree and a public record of every forecast they seal. Write down what you own — a fund, a home, a bond, anywhere in the world — the goal it serves and why you own it (My money in the world), and Money tells you when the case for your thesis changes, with sources and the counter-case.
What to do: read today’s council answer → check its public record → make your call first. Next: choose your asset, test your thesis with the council, then monitor what changes (Ask the council · Thesis Monitor).
Loading live data…
Free: every verdict and the whole Accuracy Ledger. $2/mo: today in full — the hourly Pulse, the whole debate, every Lab asset. From $24/mo: your own theses and portfolio. Plans ↓
Loading live data…
The issue in 2 minutes
Loading live data…
The weekly issue: what changed, what it means for my country, what to check at home, what was decided before.
The weekly issue, out on Saturday: three changes (a flow, a price, an event), one consequence for you, one action. Every number is checked by code before it is published.
One month, measured: what foreign holders bought and sold of US securities, how prices moved, and what it meant for mortgages, savings and inflation — against the same month a year ago.
Today's issue → · Archive of issues →
Get each issue the moment it is sealed
Every weekday before the US open: 10+ AIs answer, the answers are sealed, the issue comes out. Pick how it reaches you.
Notifications on a phone: add the page to the Home Screen
- iPhone and iPad (Safari): tap Share □↑ → “Add to Home Screen” → open Money from the new icon → “Notify me about each issue” → Allow. Apple sends web notifications only to a page opened from the Home Screen.
- Android (Chrome): menu ⋮ → “Add to Home screen” or “Install app” → “Notify me about each issue” → Allow. Without installing, notifications also work straight from Chrome.
- Computer: “Notify me about each issue” works right in the browser; or take the calendar (.ics) or RSS.
Seal your call before the AIs
One question a day, the same one the Council answers. Pick up or down and how sure you are. Your call is sealed with a SHA-256 hash in your browser, then you see the Council — and on the result day, who was closer.
Loading today's question…
Loading live data…
Who moved money, where, how much and why — biggest first; each move with the holder's whole reserve.
For the biggest money stories: which groups of people gain, which may be affected and through what — with the status of every number.
One probability from a prediction market, with the time of its price and a link to the market.
What the same sum in your currency became over the period: cash, a deposit, dollars, gold, US stocks, bitcoin.
One gauge of market stress from the OFR Financial Stress Index, VIX beside it, six tiles from our own dated series and one line of the verdict.
Country by country: does money in a safe deposit keep its buying power?
Pick two countries and see them side by side: inflation, central-bank rate, real rate, short government bond, currency, country fund, remittances, direct investment, house prices — one indicator and one date per row.
Slow rows, each with its own measure, period, country and date. None of them is a transfer between two places.
Cost of my usual life, what a loan really costs, how long a reserve lasts, a business cash calendar, subscriptions, two offers side by side, saving for a goal and a currency converter. Counted in your browser.
The interest you missed, what money from year N became, compound interest, a loan against inflation, gold in dollars and ounces, a budget and a net-worth template. Counted in your browser.
Four doors — family, work, business, investments: facts for your country with dates, how it works, and short steps to save and to earn.
The world moves trillions; a person moves hundreds. The logic is the same: enter your income and six articles, see your flow and compare it with your country.
Three more doors beside «what it changes for me»: debt and mortgage, trips and purchases abroad, not investing yet — facts for your country with dates, how it works, steps to check for yourself.
Your place on the wealth scale and on the income scale of your country, from the latest household survey (Federal Reserve, ECB, OECD). Counted in your browser.
Three everyday money questions on invented samples: a bill, extra subscriptions, a suspicious SMS. A strong AI model answers; copy the prompt and use it on your own document.
Every week we take one real decision: who made it, what they risked, how much, the document, and what happened next.
We don't teach money. We show who said what, what they did and where they were wrong — every claim with a source and a date.
What a company or a central bank said about money, when the next report can show it, and what is disclosed so far.
Money companies returned to the owners of their shares over the last four quarters, next to what they spent on equipment and data centers. Shown side by side, never added together.
What the big companies plan to spend in 2026, what they have spent, what big funds held, and what insiders and deals did. Each line has its own measure.
Form 4 and 13D filings of the last business day, with the trade date and the filing date apart, and the main money dispute of the day.
Remittance corridors «from → to» in dollars for one year — a model estimate, with the method and the year; and China's monthly receipts and payments through banks.
My topics: press ☆ next to a topic, a row, a story, a decision or a promise. Stored only on this device.
Hide the blocks you do not need and put the rest in your order — kept on this device only.
What did you do with your money this month? One answer a month; results are for readers of this site, not for the whole country.
Editorial code: a person decides what is published, no paid placements, open corrections, a Council of 10+ AIs, numbers checked by code.
Five short answers with live numbers: where money goes when markets fall, why gold gets dearer, what a rate cut changes, how government spending reaches people, why a central bank moves its gold.
For big US programs: how much was announced, how much has been committed to recipients, and how much has actually been paid out — three separate numbers, each with its own date.
Why this exists
Panels of AIs often agree for the same reasons and err together. Here it is measured every day where the close is the ground truth: every forecast sealed before it counts, scored on four horizons, with the agreement, the errors made together and the lines the models drew themselves.
Seven chapters. Everything one tap away
The main point of each chapter is open below; the rest sits under “More” — or see the whole page at once.
One question goes to 10+ AIs from the US, China and Europe. You get one sourced answer, the map of where they disagree and a measured count of independent views.
Every weekday before the New York open the council seals its calls. Each one is scored in public — misses included — against the crowd, the S&P 500 and a coin.
What matters now, where money moves by market and sector, where AI odds and market odds split — and what happened next the last time it looked like this.
Start from your own money and goal: what each move means for you, in the currency you will spend — and the way in that fits your work.
What already exists in AI research, what it costs and where Money is different — with a route from the decision you need to make.
Four core products, 21 modules of one research system, plans from free to Desk — and how to build it with us.
Five auditable layers from question to evidence, every component with its license and status, and the primary sources behind each number.
Questions and answers
Who are these AIs?
Several independent AI models from different companies. Each gets the same question about every market, answers on its own, and every answer is shown under the model’s name.
Where does the data come from, and how can I check a forecast?
Prices are public market quotes. Every forecast is sealed with a timestamp and a hash before the New York open and counted against the real close afterwards. The sealed lists and all the data can be downloaded (CSV, JSON), so anyone can recount the score.
Are the AIs better than the market?
Not proven. So far 105 comparable questions are counted of the 350 needed for a statistical verdict; today the data do not show that the AIs beat the market.
Is this investment advice?
No. All money here is virtual. The forecasts are an experiment in how well AIs call markets, not a recommendation to buy or sell anything.
One question. 10+ models. One sourced answer, with the disagreement measured
Every answer shows where 10+ models agree, where they split and how many truly independent views stand behind it. Frontier models often make the same mistakes, and the strongest ensemble is the most diverse, not the largest (a 2026 study of Metaculus questions) — so Money draws on different regions and model families and weights families instead of counting seats. Model consensus is not proof: primary data outranks any vote.
Council poll: one question, every model answers in one line
Loading live data…
● council poll · Nov 14, 2026 · 6:33–6:47 pm ET
12 models agree.
Dispatch
The question goes to all 10+ models at once. Each returns an answer, probability, sources and a split between what is known and assumed.
Fact check
Each figure is checked against its primary source; the same figure repeated by several models counts once, never as agreement. Conflicting facts get flagged.
Weight by family and track record
The strongest model in a family gets a full vote; other models from the same company get a reduced vote, so one company can’t outvote the rest. Models with a stronger record on that question type gain weight.
Map the disagreement
See who dissents, why and on what evidence; once calls resolve, the Accuracy Ledger shows which side was closer.
Compare with market odds
Set the council’s probability against Kalshi/Polymarket prices and futures. A gap can point to a trade worth testing.
Log the call
Date-stamp the forecast. Once the event resolves, update each model’s and the council’s public Brier score.
What the Council would do
Forecasts of the leading AIs of the USA and China for markets over four horizons. Where a model has no new answer, its position is unchanged.
Loading live data…
How alike are the answers
Loading live data…
14 models, but how many independent views?
Loading live data…
Loading live data…
Why they did it: each AI's own reasoning
Loading live data…
How the council debates and routes questions across models
Separate
Search builds a shared, verified source pack. Models work independently without seeing one another’s conclusions.
Clash
An advocate makes the case, a skeptic seeks contrary evidence and an arbiter weighs both. Models stress-test the thesis instead of voting on a price target.
Verify
Every claim needs a primary source, publication date and figure; a claim without one is dropped before the debate rather than argued over. Text fetched from the web is read as evidence, never as instructions. No model is a source for another: a figure one model states enters the debate only with its own primary source, so one model’s mistake cannot spread through the council as agreement. On the Nov 17, 2026 poll, 3 of 385 claims with a figure were dropped as one model repeating another (226 of 10,062 over 22 polls); two models citing one article count as one support. Free — the rule and, on the daily poll, the count of claims dropped as one model repeating another.
Poll of Nov 17, 2026: 385 claims with a figure from 13 models — 3 dropped as one model repeating another, 326 with the model’s own primary source, 56 without a source.
Test
Before the arbiter rules, the case is restated as a claim the data can settle — “this signal beat buy-and-hold after fees” — and run through the History Engine or the Strategy Backtest. The verdict carries the result (“held on the data,” “failed,” “not enough data”), how firm the reasoning is — not how far the price will move — and the one event that would break the case; that event can be followed as a condition in the Thesis Monitor. Conviction stays capped while any claim lacks a source or the test has failed. On the Nov 17, 2026 poll, 48 calls were tested on the council’s own sealed record: 1 held, 10 failed, 37 had not enough data. Free — the result on the daily poll; $2 — the full test beside the full debate while fresh, open to all after 30 days; One — on your own question.
Poll of Nov 17, 2026: 48 calls tested on the council’s own sealed record — 1 held, 10 failed, 37 not enough data; firmness capped on 28.
Remember
Outcomes are checked against forecasts to calibrate the council. The record also tracks where models disagreed. Every resolved call gets a postmortem that grades the process apart from the outcome — were the sources primary, did the test hold, did the break condition trigger — so a right call on a failed test is logged as luck and a wrong call on a sound process as a known risk that played out. To Nov 16, 2026: 393 resolved calls, 327 graded — 89 sound and right, 66 a known risk that played out, 87 right on luck, 85 weak and wrong. Free — the grade on every resolved call in the Accuracy Ledger; One — the same review of your own theses in the Decision Mirror.
393 resolved council calls to Nov 16, 2026; 327 graded on the process before the outcome — 89 skill, 66 a known risk played out, 87 luck, 85 weak and wrong.
Cascade. Routine questions go to the 3 models with the highest current weight. Conflicts, weak sources and high-stakes decisions trigger the full Council 10+; the highest-risk calls also get analyst review. You can call the full council on any question.
- Independent views, measured: on the Nov 14 poll, 13 models answered; effectively ≈ 7.35 independent answers — a single shared view would give ≈ 3.82 — and ≈ 1.71 positions on the key verdict. A council is worth what it adds beyond one model, so the engine counts it on every verdict (see the verdict). On the 12 sealed markets the same measure runs every day: the mean correlation between the AIs’ one-day calls is ρ ≈ 0.43 (21 issues) → Neff = n / (1 + (n − 1)ρ) ≈ 2.26 independent voices; across horizons 2.3 (95%: 2.0 to 2.7) (export ↗, data Nov 16, 2026). The herd module of the live record gives a second reading by horizon on Nov 16 — day 1.5 · week 2.2 · month 3.9 · six months 1.9, herd index 81% (live ↗). That is why the Council counts voices, not seats. Three models agreeing are not three confirmations →
- Weights from a record: how each seat earns its weight — see the Router.
- Which model is stronger on which market — the Accuracy Ledger shows it market by market (rates, metals, equities, China, Europe) as sealed calls resolve: skill is earned on resolved questions, not assigned by geography.
Ten-plus seats are not ten-plus opinions if the models read the same three articles. Models from one family tend to share mistakes; in a 2026 study, Grok’s different errors made it more useful to the crowd than ten similar models. So every verdict carries an independence score, computed by the engine and shown beside the council verdict.
“Diversity is the Strength of the AI Crowd”, 2026 ↗ · “Wisdom of the Silicon Crowd” ↗
Ask the council. 10+ AIs answer
Send the question you’d ask a research desk. 10+ models from the US, China and Europe answer it with sources, a map of where they disagree and what would change the answer — then keep it up to date.
Your market question, answered by 10+ AIs
Check your idea → · Two sentences first, then the Council →
Send your question through the form. The answer comes back with sources, the disagreement map and what would change it, and becomes a living answer: dated, and updated as the facts change. The council works through new questions twice a day, at 10:00 ET before the US open and at 17:30 ET after the close. Your email is optional and stays with Google and on our Mac.
Ask a question and see the queue
Loading the engine slice…
Answers base: questions investors ask
Answers base: questions investors ask
27 questions with the latest dated version of each answer. A version is kept when the facts change; sources are inside each answer.
Loading live data…
Each answer to a reader’s question is marked up as a Q&A page (schema.org QAPage): question, dated answer, last-updated date and links to the sources and engine slices behind its figures.
When the answer hinges on something you didn’t say — horizon, spending currency, tolerable loss, tax country — the council first asks one to three follow-ups; skip them and the answer states its assumptions (“in dollars, over three years”). Follow-ups cost nothing, on every plan. Checked on 70 questions in all site languages: never more than three.
A question about a specific trade (“should I buy X?”) gets scenarios — “if A holds, …; if B breaks, …”: the case for and against, the risks and what to watch — never a buy or sell call.
What matters most on Wall Street right now?
Whether AI capex can pay off as rates rise. Hyperscalers are spending $725–828bn in 2026 (sources and the financing behind it — the AI capex meter) while the 10-year closed at 5.15% on Nov 13 (US Treasury ↗) and markets expect a Dec 9 Fed hike (Axios, Nov 10 ↗).
Answer by the Money AI council · 13 models · sources in the council section above.
Sealed before the open. Scored in public
Every weekday before the New York open, AIs from the US and China seal calls on 12 markets — the indexes, the US 10-year, the dollar, VIX, gold, oil, copper, bitcoin — for a day, a week, a month and six months. External timestamps fix the moment: Bitcoin via OpenTimestamps since October 16, 2026 (first issue, 22:30 ET), RFC 3161 since October 20, Sigstore Rekor since October 21; the open book lets anyone recount every result. How to check ↓ What “sealed” means ↓
First your call, then the comparison: Seal your call before the AIs →
Loading live data…
What “sealed” means
Forecast → external timestamp → open ledger → scored after close
The live score today
Numbers straight from the open record: they change after every issue and every close.
Loading the live score…
Brier of the Council, the crowd and their mix on the same questions, with the flag «saw the market price» on every sealed answer.
Loading…
check it in 3 steps
1) download council-forecasts.csv or the sealed book council-book.jsonl; 2) compare today's lines with the seal hash and its OpenTimestamps proof; 3) recount Brier and the right side from the outcomes in the same file.
Every account, studied
Four leagues on the same AIs, every account on one chart, and the check that being right more often is not the same as earning more.
Loading…
Calendar of mistakes
Every position an AI closed at a loss in the league, day by day — the darker the day, the bigger the losses. Tap a day to read who lost, on which market, how much, and the AI’s own reason.
Loading…
How good are the forecasts — checks on ourselves
Does a stated 60% come true 60% of the time, what a good Brier looks like, who is right in which market regime, and two tests sealed in advance.
Loading…
Which AI is stronger on which market
More useful than one overall winner: each AI on each of the 12 markets, and accuracy by type of question.
Loading…
Book of misses — explore every case
Where the Council and the crowd were both wrong, where only one of them was, market by market — and what each AI said that morning, from its own sealed line.
Loading…
How to check any call: call → seal before the open (SHA-256 chain; Bitcoin via OpenTimestamps since Oct 16, 2026; RFC 3161 time stamps since Oct 20; the Rekor log since Oct 21) → outcome at the close → score in the open book; a correction is a new line, the old one never changes.
Nothing deleted, every fix on file — check it yourself
The open book of forecasts only grows: lines are numbered, each carries the link of the line before, and a fix is a new line beside the old one. The latest lines are checked as the page opens; the whole book — in your browser, on one tap.
Loading…
The buttons open the Money live record on polyhelper.ai/trends, running since October 16.
Scoreboard by origin: US · China · Europe
Same questions, one seal. The score accumulates; a verdict comes after 350 comparable questions; until then the comparison is provisional. The sealed record runs 10+ AIs from the US and China; Le Chat (Mistral, Europe) answers the Council poll and joins the sealed book next.
| Who | Right side | Brier | Forecasts | Team account, $100,000 |
|---|---|---|---|---|
| US AIs | 54.7% | 0.2478 | 1,287 | $96,929 |
| Chinese AIs | 54.7% | 0.2466 | 1,197 | $95,828 |
| European AIs | next | next | — | Le Chat (Mistral) joins the sealed book next |
| The crowd (Polymarket, Kalshi) | 74.0%* | — | 100* | — |
| S&P 500, bought and held | — | — | — | $98,177 |
* the crowd and the Council on 100 comparable crowd questions; AI rows — all their resolved forecasts. Cut-off: session of Nov 13, 2026, data Nov 16, 2026, 11:38 ET. council-league.json ↗ · council-stats.json ↗
Your $100,000 against the AIs → · Humans vs machines: one table for everyone → · Share the day in one picture → · Get each issue the moment it is sealed →
What the AIs’ work cost — beside what they made
Each account’s return before and after paying for the AI itself, at the provider’s public list price: does the result survive the bill?
Loading…
More: the US–China paired score, East against West, calibration, the crowd by venue, independence, the reader experiment, the schedule
Paired on the same 388 cells: Brier for US AIs 0.2495, Chinese AIs 0.2506; difference −0.001 (95%: −0.0056 to +0.0037, 20 sessions) — not yet distinguishable. East against West, Nov 16: the widest gap between the groups is copper — West 47%, East 43%; distance from the crowd: West 11.0 points, East 8.7; furthest from its own group: Amazon Nova, 6.7 points; the result comes after 30 outcomes (0 so far). East against West ↗
- AIs and the crowd on the same contracts — same side on 92% of 49; the crowd sits 22.7 points from 50, the AIs 8.9. The gap contract by contract is in the Divergence Radar. export ↗
- Best start day — from Nov 13, 2026: the Council $99,856, the S&P 500 $99,637. export ↗
- How independent the AIs are — on 12 markets ≈ 2.26 independent voices on the one-day horizon; the method and our verdict-level measure are in the Council ↑. export ↗
- Against the crowd, venue by venue — Polymarket: 51 events, AIs right on 36, the crowd on 35, Brier 0.2177 vs 0.1758; Kalshi: 9 events, 8 vs 9, Brier 0.1974 vs 0.0711. These are the league’s resolved events — a different, smaller set than the 100 comparable questions above. export ↗
- Do the AIs make readers more accurate — readers are split for good into two groups: one sees the AIs before its first call of the day, the other only after; first calls are compared, target 200 per group; the count is building. export ↗
- Schedule — forecasts are sealed before the 09:30 ET open, the day’s score comes after the close, at 16:30 ET; Saturdays bitcoin, Sundays the weekly radar.
History of the models — the Council’s ghost
AIs get new versions. Every version is kept with its own dates and answers, and earlier versions answer the same questions beside the new ones: is the newer one really more accurate?
Loading…
More: all 14 league accounts and 43 more modules of the live record on polyhelper.ai — leagues, market-by-market signals, checks on ourselves, your turn, data · all open
Leagues and virtual accounts
- AI Portfolio League — each AI runs a virtual $100,000: cash plus up to 12 markets, 5–25% a position, filled at the first price after the open, costs inside; the rules are an open file in the method. Right more often ≠ earned more: both are shown side by side. All 14 accounts are in the table above. live on polyhelper.ai · Free open ↗
- Pro league — a second $100,000 per AI with shorts and leverage up to 2×, margin calls and a closed account on a wipeout; its own calendar of mistakes. Nov 13 close: leader Gemini $102,475 (+2.47%), the S&P 500 −1.56% over the same dates from Oct 20. live on polyhelper.ai · Free open ↗
- From the day you pick — $100,000 put into the Council’s portfolio at any start date, against the same sum in the S&P 500. live on polyhelper.ai · Free open ↗
- Seasons — season 1 (Oct 19 – Nov 8) stays frozen as the baseline; from Nov 9 each new league runs from its own start date with its own score. live on polyhelper.ai · Free open ↗
Experiments on the same accounts — the library
Each card: what is tested and the rule, the result right now, the full table one tap further.
Loading…
Do the AIs make readers sharper? → · Learn to read a forecast by playing with one →
Experiments on the same accounts
- Risk league — virtual $10,000 per AI, the whole account on ONE most confident position at 3× every day; a zero is a wipeout. As of Nov 16, 2026: leader ERNIE (Wenxin) $19,942, wipeouts 0 of 14, 5 of 30 resolved bets each; control (random side): a coin $9,498. live on polyhelper.ai · Free open ↗
- Tail league — $100 on the cheap side (under 20¢) of every crowd market, every zero listed. As of Nov 16, 2026: $10,000 → $6,929; 2 of 42 contracts paid. live on polyhelper.ai · Free open ↗
- Fade the Council — virtual $10,000 that bets AGAINST the Council wherever it differs from the crowd by more than 15 points, $100 a bet, with a random-side control. As of Nov 16, 2026: 0 of 30 settled. live on polyhelper.ai · Free open ↗
- Paper alpha fund — rules published before the count: long or short on the Council’s consensus, equal weights, rebalanced on Mondays, against the S&P 500 and cash; started Nov 9. live on polyhelper.ai · Free open ↗
- Confidence poker — each AI spreads a bank across the markets by its own confidence; compared with equal shares. As of Nov 16, 2026: biggest stake Amazon Nova, $1,500 on chips falling. live on polyhelper.ai · Free open ↗
- Whale mirror — the month’s top-5 Polymarket traders by public profit: where an AI sides with them its virtual $10,000 copies, where against — it bets against. live on polyhelper.ai · Free open ↗
- Polymarket–Kalshi spread — the same market on two venues, the cheaper side of each, after fees, against a random-side control. As of Nov 16, 2026: 11 pairs, widest gap 12.2 pp, +$352 vs +$161 for the control, 9 of 30 settled. live on polyhelper.ai · Free open ↗
- Hunt the herd — when the crowd is over 80% sure, an AI may bet against it: +3 for a right hunt, −1 for a wrong one. As of Nov 16, 2026: 81 bets, 0 right. live on polyhelper.ai · Free open ↗
- Shadow modes — the same AIs under other conditions, each with its own accounts and file: the 12 real funds (SOXX, TLT, IBIT…), a closed book without web search, an S&P 100 league, role characters. live on polyhelper.ai · Free open ↗
- Market-aware vs blind Council — the market-aware Council moves halfway to the crowd’s price, the blind one never sees it; lower Brier wins after 30 outcomes. live on polyhelper.ai · Free open ↗
- Council weights v2 — weights set on season 1 and sealed before season 2 (hash published Nov 9): the weighted Council against the simple mean on new questions. live on polyhelper.ai · Free: the result and the hash · the weights: no one open ↗
What the council says, market by market
- Map +100% — sealed before the open, each AI may name one market that could double within a week, with its chance; scored when the week closes. live on polyhelper.ai · Free: the summary line · $2 at launch: in full while fresh, free once resolved open ↗
- Court of the Council — after the seal, up to three AIs that disagreed with the majority name one dated fact it may be missing, and the level at which they would concede. live on polyhelper.ai · Free: the summary line · $2 at launch: in full while fresh, free once resolved open ↗
- Heat map: market × model × horizon — side and confidence of every AI on every market for a day, a week, a month and six months; each cell links to its slice. live on polyhelper.ai · Free: the consensus cell of the day · $2 at launch: the full map while fresh, free once resolved open ↗
- Why they did it — each AI’s own headline, chain of reasoning and why its portfolio looks the way it does, as written. live on polyhelper.ai · Free: each AI’s headline · $2 at launch: the full reasoning while fresh · the reasoning archive: contract open ↗
- Market cards — 12 markets — S&P 500, Nasdaq-100, Dow, chips, the US 10-year, the dollar, VIX, gold, WTI, Brent, copper, bitcoin: each AI’s action, chance of a rise, target, ‘wrong if’ and entry; a price chart with every AI’s entry level and ‘wrong if’ line, the Polymarket and Kalshi odds beside the AIs’; the AIs can stay hidden until you make your own call. live on polyhelper.ai · Free: the consensus line · $2 at launch: every AI while fresh open ↗
- Four roles on one market — bull, bear, historian and crowd advocate on the same market; the historian-vs-advocate gap goes into the debate record. live on polyhelper.ai · Free: the summary line · $2 at launch: in full while fresh, free once resolved open ↗
- Expected moves, 5th / 50th / 95th percentile — the mean of the percentiles the AIs gave, market by market. As of Nov 16, 2026: VIX −13.2% / 3.6% / 29.8%. live on polyhelper.ai · Free open ↗
- Mirror of fear — how far the AIs disagree on tomorrow’s VIX: a disagreement index. As of Nov 16, 2026: spread 5.8 points. live on polyhelper.ai · Free open ↗
- Crowdometer — the crowd’s price of ‘up’ minus the Council’s mean, market by market; right of the line the crowd is pricier. As of Nov 16, 2026: dollar index +15.6 points, gold −5.3, S&P 500 −8.6. live on polyhelper.ai · Free open ↗
- First: the AI or the price — after a news event, did an AI call the move before the price made it. As of Nov 16, 2026: 0 of 9 events. live on polyhelper.ai · Free open ↗
- Council ghost: older vs newer versions — earlier versions of the same AIs on the same questions; is the newer version more accurate (a gap under 5 points counts as the same). live on polyhelper.ai · Free open ↗
- Anti-herd points — a point for changing sides with a stated reason; counts the change, not whether it was right. live on polyhelper.ai · Free open ↗
- Sources the AIs leaned on · links that did not open — every cited link counted by site; a 404 or a non-existent site is counted against the answer. As of Nov 16, 2026: 9 of 189 links (4.8%) did not open. live on polyhelper.ai · Free open ↗
Checks on ourselves
- Museum of the seal — three sealed lists a day — the last before the open, the first after it, the first after the close — with their line counts. live on polyhelper.ai · Free open ↗
- Calibration by AI — ECE and Brier for each AI; who sounds surer than it turns out and who is underconfident. live on polyhelper.ai · Free open ↗
- Who is right in which regime — Council and best-AI Brier when VIX is below 18 or 18–25, when oil moves 2%+, when the S&P 500 is up or down. live on polyhelper.ai · Free open ↗
- Oracle league (Elo) — every AI against every other on the same questions, lower Brier wins; form over 7 days and the AI of the month. live on polyhelper.ai · Free open ↗
- Book of misses — the days when the Council and the crowd were both wrong, with the quote that led the call. As of Nov 16, 2026: 29 of 155 cases; the three latest shown are in the table below. live on polyhelper.ai · Free open ↗
- Places by the 30/10 rule — a place after 30 resolved forecasts over 10 trading days, by horizon, with a 95% interval and tie groups. live on polyhelper.ai · Free open ↗
- Scientific checks — horizon profile, how distinct each AI’s view is, and the test of whether US and Chinese AIs form two blocks. live on polyhelper.ai · Free open ↗
- What speaks against us · what studies showed · nearest projects — outside results set beside ours, each with its source and how our method differs. live on polyhelper.ai · Free open ↗
Your turn
- Make your call first · your own $100,000 — seal your own chance before you see the AIs, or spread virtual $100,000 over the same markets; a shared table of humans against machines and the weekly ‘Council or crowd?’. live on polyhelper.ai · Free open ↗
- Try the rules yourself — runs in your browser, nothing is sent: update a belief with a new fact (odds × strength^weight), check a hash against the seal, test your own rule on resolved questions. live on polyhelper.ai · Free open ↗
- Submit your own model — probability forecasts for the same 12 markets and four horizons in the same JSON, sealed like ours, in a separate table of invited models. live on polyhelper.ai · Free open ↗
Data, widget and languages
- Widget: AIs vs the market — free to put in an article, no sign-up; the numbers update with each issue (served by the live record). Code:
<iframe src="https://polyhelper.ai/trends/markets/embed/en/" width="100%" height="230" style="border:0;max-width:520px" loading="lazy" title="AIs vs the market"></iframe>. live on polyhelper.ai · Free open ↗ - How Money grew — one chart of the Council, the S&P 500 and cash, with the dated steps: Oct 16 first issue and the Bitcoin seal · Oct 19 the league starts · Oct 20 RFC 3161 time stamps · Oct 21 the Sigstore Rekor log · Oct 22 AIs set against the crowd · Nov 6 a block for newcomers · Nov 9 season 2. live on polyhelper.ai · Free open ↗
- Daily alerts — a calendar feed (.ics), RSS and web push, one line whenever an issue comes out; the Open API with every file of the record (CC BY 4.0) — index of all files ↗, webhooks
issueanddaysigned with HMAC. live on polyhelper.ai · Free open ↗ - The live record in all site languages — the live record on polyhelper.ai in English and ten translations: Russian · Ukrainian · Spanish · Portuguese · French · German · Japanese · Chinese · Korean · Hindi. live on polyhelper.ai · Free open ↗
Every module above is live and free on polyhelper.ai/trends; “open” takes you straight to it. Figures dated Nov 16, 2026 come from the live record. Method and seal checks — method ↗.
Cut-off: session of Nov 13, 2026, data Nov 16, 2026, 11:38 ET · snapshot 138d253f1c98 · all money is virtual.
For those who follow theses and markets
November 2026. What really matters?
AI is hitting the balance sheet. Investors must weigh its potential against borrowing costs and real-world bottlenecks: power, copper, chips and capital.
2026 capex by the largest data center operators (IEA); Goldman puts global AI investment above $1tn.
IEA ↗ · Goldman Sachs ↗of asset owners question the payoff from AI valuations and capex; 76% cite inflation (Morningstar 2026).
Morningstar ↗prioritize diversification; 85% expect more volatility (Schroders, respondents managing $72tn).
Schroders ↗World Bank sees metals prices at +25% in 2026, then −3% in 2027. One forecast won’t capture that swing.
World Bank ↗More context: the core question, what asset owners flag, the investor pain map
You enter a company, sector or asset → Money checks how much future profit the price already assumes → you get a dated verdict: still underpriced, or already paid for.
Money does it one company, sector or asset at a time: the evidence, the counter-case, the scenarios and a dated record of what changed. Wells Fargo’s November outlook circles the same issues — data-center payback, how the build-out is financed, rising leverage and whether demand turns into cash flow.
- Inflation and whether AI valuations and capex pay off — the card above
- 71% — the rise of generative AI
- 66% — energy and logistics crises
500+ asset owners in 11 countries · Morningstar ↗
| Investor pain point | What the investor needs | Where Money answers |
|---|---|---|
| Fear of a large loss | Know how much could be lost — and why | Portfolio Guard: stress tests and historical regimes |
| Information overload | A few material events, not thousands of headlines | Thesis Monitor: alerts only when the facts change |
| Distrust of data and advice | The primary source, the calculation and its limits | Evidence chain and Report Checker |
| Fear of missing out | Tell a real opportunity from hype | Council debate: advocate, skeptic, arbiter |
| Hidden concentration | See when different holdings ride on one factor | Portfolio Guard: factor exposure |
| Money beyond Wall Street | Know what a local asset will really be worth in the currency I will spend — after the exchange rate, local tax and inflation | Goal Navigator: My money in the world |
| Wanting a machine to do it | Hand the routine to AI without losing the brakes — and see its record before trusting it with money | Broker Agent within your limits; the Accuracy Ledger first |
| No time | Stay on top of positions without reading the news all day | 30-minute routine and morning brief |
| Uncertainty | A plan for more than one outcome | Scenario Lab: base, upside and downside cases |
Behind the map: investors target 7.9% a year, yet 22% of their portfolios sit in cash (Fidelity — the yardsticks behind Goal Navigator), and only 44% feel confident about a comfortable retirement (Betterment) — they want growth but hold back for fear of a loss. Individual investors and small funds often see news after algorithmic traders; a single model can invent a valuation multiple or repeat a stale quote; and when a handful of AI companies dominate an index, a day like Nov 5 hits portfolios that looked diversified. McKinsey cites geopolitics, AI’s impact and capital discipline; Janus Henderson cites worries about AI valuations, distrust of recommendations and how hard they are to verify. McKinsey ↗ · Betterment 2026 ↗ · Fidelity 2026 ↗
This week’s Pulse: what moved markets, the events ahead, themes on the radar
The AI boom meets a 5% Treasury yield
- Earnings vs. rates — S&P 500 Q3 earnings growth near 30% (FactSet ↗), while the 10-year crossed 5% for the first time since 2023 and hit 5.16–5.18% in September, the highest since 2007; on Nov 13 it closed at 5.15% (US Treasury ↗).
- Nov 5 — after Democrats won both chambers, fears of limits on data-center construction sent the S&P 500 down 4.4% — its worst day since April 2025 — and the Nasdaq-100 down 6.4%, its worst since March 2020 (HNGN, Nov 5 ↗).
- Inflation, Nov 10 — October CPI rose 0.4% on the month (BLS ↗). The consensus was 0.3%, and Fed Governor Waller raised the possibility of a 50 bp hike (MarketWatch via Morningstar ↗).
- Company filing, Nov 12 — Nvidia’s 8-K: on Nov 11 the board accepted Jensen Huang’s resignation as CEO and named CFO Colette Kress interim CEO (SEC filing ↗). The Thesis Monitor puts a filing like this on every thesis that holds the company; the next filing replaces this line.
- Concentration — AI companies account for more than 80% of S&P 500 growth in 2026 and about half its market cap (J.P. Morgan AM). Morningstar put US stocks about 10% below fair value at Sep 30 — yet without seven stocks the market traded right at fair value (Morningstar ↗).
- October rotation — Morningstar’s US sector indexes: energy +14.05%, consumer defensive +11.55%, utilities +9.87%, healthcare +8.82%; communication services −14.72%, where Alphabet and Meta make up 78% of the index (Morningstar, Nov outlook ↗). Index returns, not measured money flows; Money’s own rotation, country by sector, is the board’s Rotation view.
- Mood, week to Nov 4 — AAII bearish sentiment 48.4% (up 6.9 points on the week) against a 31.5% average (AAII ↗).
- Nov 13 close — S&P 500 7,561.67 (−0.40% on the day, +2.51% on the week, ending a four-week losing streak; record 7,844.67 on Oct 6). Dow 52,347.55 (+0.10%). AI, chip and momentum stocks sold off; defensives and the equal-weighted index held up (CNBC ↗).
CNBC ↗ · Market Scholars ↗ · Goldman Sachs AM ↗ · TMGM ↗ · Kitco ↗ · Wells Fargo ↗
| Date | Event | What the council seals |
|---|---|---|
| Nov 17 | Nvidia earnings, days after its Nov 12 leadership filing (SEC 8-K ↗) | Data-center revenue vs. guidance; reaction of the chip index |
| Nov 18–19 | Anthropic IPO (Nasdaq: ANT) — 825mn shares at $70–78 — about $0.92–1.02tn on the ~13.09bn shares outstanding after the offering, up to ~$1.4tn fully diluted; pricing expected Nov 18, first trade Nov 19 (WSJ via Decrypt ↗ · SEC S-1/A ↗). The range is public; the council seals its own call | Pricing vs. range; first-week move |
| Dec 9 | Fed decision — after the Nov 10 CPI, markets expect a hike (Axios, Nov 10 ↗; live odds: Kalshi ↗, Polymarket ↗, CME FedWatch ↗); when the council seals its forecast, a dated snapshot of these odds is kept beside it and both are scored after the decision | Hold, +25 bp or +50 bp, with probabilities |
|
FED DECISION · MARKET ODDS BESIDE THE COUNCIL
Loading the Kalshi and Polymarket odds… | ||
| Dec 10 | November CPI | Above, at or below consensus |
| Dec 17 | ECB decision (ECB calendar ↗) | Deposit rate held or changed, with probabilities |
| Dec 18 | Bank of Japan decision after its Dec 17–18 meeting (BoJ ↗) | Hold or hike, with probabilities |
More: what the calendar reads — other central banks, wording, σ surprises, prospectuses, rule changes
Other central banks: the decisions and CPI of every market on the board — the Bank of England, RBI and Copom after the ECB and the Bank of Japan — and new listings on HKEX and NSE are sealed the same way and scored market by market; RBI meeting dates come from RBI’s own schedule, CPI dates from the calendars of the BLS, Eurostat, IBGE, the ONS, the Statistics Bureau of Japan and MOSPI.
Central-bank wording: after each decision the engine compares the statement with the previous one — phrases added, dropped or softened — and sets that shift beside the sealed call and in the calendar row: the Fed, the ECB and the Bank of Japan first, then the central banks of the board’s markets. Free — the shift once the next meeting has passed; $2 — while it is fresh. Sites that render by script or answer bots with 403 are read now: the BIS feed of central bankers' speeches (50 speeches from 34 institutions, Oct 29 – Nov 12, 2026), the latest speeches of all nine regional Reserve Banks that were missing from the speaker map, RBI's policy corridor from its own daily release (repo 5.50%, SDF 5.25%, MSF 5.75%, Nov 16, 2026) and the ECB's statement of Oct 29, 2026 set against Sep 10: rates unchanged after a 25 bp hike, 54 sentences new, 61 dropped, 21 kept.
Speaker map: Live for the Fed Board, the ECB and the Bank of England, with dissents from the FOMC statements, and for the Bank of Japan’s board and the regional Fed presidents whose banks publish a speech feed (Dallas, Richmond, San Francisco); the other regional presidents are next. Every speech, interview and dissenting vote of each policymaker — Fed governors and regional presidents, the ECB Governing Council, the Bank of England’s MPC, the Bank of Japan’s board — read against that person’s own earlier words, so a hawk turning cautious stands apart from a hawk repeating himself; the map sits beside the sealed call for the next meeting, and the Accuracy Ledger later shows whose words led the decision. Free — the map once the meeting has passed; $2 — while it is fresh.
Speaker map: each policymaker against his or her own words
Regional Fed presidents and the Bank of Japan board
Surprise in σ: the released figure, the median of the council’s sealed calls and the history of misses on that series → “came in at X, expected Y, N σ above” in the calendar row; above |z| of 2, with a condition touched, it becomes a Thesis Monitor event.
Prospectus amendments: each new S-1/A is compared with the previous version (edgartools) — price range, share count, debt, risk factors, dates — and the listing row shows what changed before the first trade, linking both filings. First case: Anthropic's second amendment (10 Nov 2026) set the first price range, $70–$78, and moved the risk that its models have behaved in unintended ways from 13th place to first (S-1/A ↗). Free — the day’s main release and the fact that a prospectus changed; $2 — every release and the full amendment read the moment they land, free a day after the release and after the listing; One — on your thesis.
Rule changes: the dates when rules for markets and for AI in finance take effect — exchanges, securities regulators, EU AI Act stages — from the official text, with withdrawals logged as events too: on June 12, 2025, the SEC withdrew 14 proposals, including the predictive data analytics rule for brokers and advisers (SEC ↗). Each change is read against the previous text — what was added, dropped or reworded — with the publication date and the effective date kept apart; a proposal never changes the live Market Profile, and a repeal is logged as a new version. Free — the decision, its dates and the official link; $2 — the full change map while fresh, open to all after 30 days; One — alerts when a rule touches your thesis.
Government bond auctions: Live for the US, Germany and Japan from each debt office’s own results. Each sovereign auction in the 20 markets of Government bonds of the world → bid-to-cover, the tail against the pre-auction yield, the foreign share and the surprise in σ against that issuer’s own history. A bond-vigilante flag lights when a weak auction, a rising term premium and a falling currency land in the same week; the Atlas’s fiscal-limit flag stands beside it. Debt agencies’ borrowing plans and central-bank statements are read in the original language; a shift in a borrowing plan is $2 while fresh, then free. Free — the auction result; $2 — the σ surprise and both flags while fresh, open to all after 30 days.
What the calendar reads
Anthropic, PBC · S-1/A No. 2 · 10 Nov 2026
| price range | shares offered | net proceeds | total debt | value at midpoint |
|---|---|---|---|---|
| $70.00–$78.00 | 825.0 mn | $84.6 bn | $6.2 bn | $968.4 bn computed |
2026-11-18 · Pricing of the Global Offering · 2026-11-19 · Listing on the Nasdaq Global Select Market · FWP ↗
S-1/A No. 1 (2 Nov) → S-1/A No. 2 (10 Nov) · previous filing ↗ · new filing ↗
- first price range: $70.00–$78.00; 801.4 mn new shares and 23.6 mn from selling stockholders; option for 123.75 mn more
- concurrent private placement of Class N stock: $26.0 bn
- risk factor reworded and moved from #13 to #1: «Our models have in the past behaved, and may in the future behave, in unintended, misaligned, or autonomous ways…»
New listings on NSE
CPI dates: UK, Japan, India
| Theme on the radar | Where it stands, Nov 2026 | Where Money tracks it |
|---|---|---|
| AI capex vs. borrowing costs | The organizing story: capex lifted chip stocks “impressively” while fears of AI “displacement” weighed on software (Columbia Threadneedle, 2027 outlook ↗) | Scenario Lab · AI-capex meter |
| Elections and the Fed | Data-center regulation risk after the Nov 3 midterms; after the Nov 10 CPI, markets expect a hike on Dec 9 (Axios, Nov 10 ↗) | Event calendar |
| Robotics and Chinese assets | Chinese industry and supply chains as the next leg of AI demand: China installed 295,000 industrial robots in 2024, 54% of the world total (IFR World Robotics, Sep 2025 ↗) | Country × Sector board |
| Copper — “AI’s new oil” | Records of $14,533/t on the LME and $6.95/lb on COMEX in September on AI and grid demand (Yellow ↗) | Scenario Lab · copper |
| Gold under rate pressure | From a record $5,595.47 intraday spot on Jan 29 (World Gold Council; helped by central-bank buying in China, India and Poland) to ~$3,881 on Nov 13; −2.46% on Nov 13 alone. 2025 was gold’s best year since 1979 (WGC ↗) | Scenario Lab · gold |
| Event odds as a market | Kalshi and Polymarket handled $50.6bn in July 2026; Kalshi set a $15.27bn weekly record in September (Tangem ↗) | Divergence Radar |
| Oil and geopolitics | Brent at $95.84 at the Nov 13 close amid the Iran war (Vista Partners ↗); geopolitical conflict tops the underestimated risks in McKinsey’s survey of 112 investors (McKinsey ↗) | Portfolio Guard · “oil at $120” |
| Defense, space and drones | $35.6bn of venture funding into defense and defense-adjacent startups in H1 2026 (PitchBook via Tectonic ↗); SpaceX’s $75bn IPO on Jun 11 (NPR ↗) | Crowding Index · themes |
| Bitcoin | ~$67–69k on Nov 13, −30% year over year, with outflows from spot ETFs (CNBC ↗) | Pulse |
| Beyond AI | GLP-1 and longevity drugs: Eli Lilly’s Q3 2026 tirzepatide sales reached about $16.1bn (Mounjaro $10.7bn, Zepbound $5.4bn) (Pharmaceutical Technology ↗) | Report Checker |
| 2027 watchlists | Global X: AI, defense and space, infrastructure and automation, energy, critical resources. Grid equipment (Eaton, Schneider Electric, Siemens Energy), copper (Freeport-McMoRan), uranium (Cameco), SMRs, transformers and cables, quantum-resistant cybersecurity, satellite networks, Dow dividend payers with real free cash flow (Global X, 2027 outlook ↗) | Accuracy Ledger · house forecasts |
| Concentration and Asia | The 10 largest stocks are more than ¼ of MSCI World; Asian emerging markets have led this year. Earnings growth is broadening: for Q3 2026 the Magnificent 7 show earnings growth of 22.6% vs. 32.6% for the other 493 S&P 500 companies — blended rates (reported results plus estimates for companies yet to report) as of Nov 3, and growth rates, not shares of total profit (FactSet ↗) | Portfolio Guard · hidden concentration |
Columbia Threadneedle 2027 ↗ · Semafor ↗ · Rio Times ↗ · Yellow ↗ · Tangem ↗ · DeFi Rate ↗ · McKinsey ↗ · Vista Partners ↗ · Global X ↗ · World Gold Council ↗ · AP · gold’s best year since 1979 ↗ · FactSet ↗
Research as of Nov 14, 2026: where capital went in 2026 — on Wall Street and beyond it — and where it may go in 2027 · government bonds of the world
Pools of money, kept apart on purpose: they have different denominators and cannot be added into one “global investment” total.
| Pool | Period and measure | Source |
|---|---|---|
| US ETF flows | Record $1.54tn by September, $1.75tn by October. Bond ETFs a record $469bn; technology led sectors at $66bn; dividend and income equity ETFs more than $100bn. US equity funds ~$974bn in the 12 months to October vs. ~$201bn for developed markets ex-US. Demand for rates, corporate credit and EM debt | InvestmentNews ↗ · ETF Trends ↗ · State Street ↗ · iShares ↗ |
| Fund flows of Vanguard, Invesco, Fidelity and VanEck ETFs · SEC N-PORT | Quarterly, about 60 days late. Flow (sales minus redemptions) and net assets are separate figures. Vanguard rows cover the whole fund, all share classes. | SEC EDGAR N-PORT ↗ |
| Global ETFs | A record $2.37tn of net inflows in 2025 and $19.85tn of assets at year-end (ETFGI); $20.64tn by the end of January 2026. In Q1 2026 active ETFs took $179.3bn against $178.7bn for passive; equity ETFs $162.5bn, bond ETFs $112.2bn of a $358.1bn total (S&P Global). Providers differ in coverage and method: ETFGI puts Q1 net inflows at a record $626.42bn, so the two estimates are never added or read as one series — and assets, a stock, are never added to flows | ETFGI, S&P Global Market Intelligence |
| Managed assets | Third-party managed assets a record $151tn at the end of 2025 across 42 countries; net flows equal to 3.2% of opening assets, concentrated in a small group of managers — a stock of assets, not a flow | McKinsey, Oct 2026 |
| Gold ETFs | About $41bn into physically backed funds worldwide, Jan–Oct | World Gold Council |
| Venture | AI took $102bn, 64% of global venture funding, in Q3 (Crunchbase) and 82.7% of US deal value Jan–Sep (PitchBook-NVCA); global startup funding $685bn in 9 months vs. $449bn in all of 2025 (Dealroom) | Crunchbase, PitchBook-NVCA, Dealroom |
| Private equity | $1.5tn across 14,650 deals in the first nine months of 2026, $854.6bn of it in the Americas (6,845 deals), $521.4bn in Europe, the Middle East and Africa (6,249) and $137.8bn in Asia-Pacific (1,109) — fewer, larger deals. By sector: technology, media and telecom $479.1bn, energy and natural resources $243.3bn, industrials $122.6bn; corporate carve-outs $182.2bn. Exits were worth $985.2bn; the trailing-12-month exit count hit a 22-quarter low while exit value reached an 18-quarter high | KPMG Pulse of PE, Q3 2026 ↗ |
| Corporate and industry spending | Global energy investment ~$3.4tn (industry spending, not fund flows); five tech companies spent more than $400bn on capex in 2025, mostly data centers; Goldman expects AI capex to rise as a share of GDP | IEA, Goldman Sachs |
| Asset-owner allocations | Raised 2026 allocations to private equity (43% of respondents), infrastructure (35%) and private credit (32%) — shares of asset owners raising allocations, not dollar flows. Wealthy individuals (600+ in the UK, Europe, Asia, the Middle East and Africa; Barclays Private Bank survey, Oct 1, 2026; percentages as reported by WealthBriefing): 60% plan to raise their private-market allocations, even though the average share they allocate to private markets edged down from 30% to 27%; within private markets, technology (63%) and healthcare (54%) are the sectors most want more of — shares of respondents | Morningstar, 2026 ↗ · Barclays Private Bank ↗ · WealthBriefing ↗ |
| Private markets and private credit | Three measures, never added together. Loans outstanding: private credit loans to US borrowers were $1.54tn at end-June 2026 (Fed Z.1, cost basis), up $185bn in a year. Assets under management worldwide, including committed capital: $1.7–2.1tn for 2023 (Fed, IMF). Private funds of SEC-registered advisers: $12.2tn net assets in private equity, venture, real estate and other private funds at end-2025 (Form PF). Free — the range row; $2 — holders, fund types and eight quarters. | Fed Z.1 ↗ · SEC Form PF statistics ↗ · IMF GFSR ↗ |
| BDC loan books | When the loans of the largest business development companies come due and which have stopped paying, read line by line from each BDC’s own Schedule of Investments. Today: 7 BDCs with $160bn of investments at fair value (Sep 30, 2026) — loans on non-accrual are 2.3% at cost and 1.3% at fair value; $47.9bn of their debt is due within three years; 13% of the debt pays part of its interest in kind. Fair value and face value are shown apart. Free — this row; $2 — every BDC with its maturity ladder and its largest non-accrual loans. | SEC EDGAR 10-Q ↗ |
| Private company share prices · Forge Price and the last funding round | Only what Forge publishes openly, with its date. Forge Price is an indicative daily price; a funding round is a deal. Where Forge shows no figure, it is marked not measured. | Forge ↗ |
| Off Wall Street · money that moves outside the market | First sources only; each figure shows its measure (flow, stock or share), period, link and quote. A figure the source does not print is marked not measured. | IFSB ↗ · PBoC ↗ · World Gold Council ↗ · World Gold Council ↗ · World Bank ↗ · IFAD ↗ · SEBI ↗ |
| Who financed US real estate | Who financed US real estate in a year: $729bn of net new mortgages — agency pools $314bn, banks $211bn, private ABS $136bn (Fed Z.1) | Fed Z.1 ↗ |
| Energy FDI | Energy FDI 2024: Australia +$13.6bn into mining and oil and gas, Canada +$11.9bn (OECD) | OECD FDI ↗ |
| Fund and manager holdings | How the disclosed positions of large managers and funds changed, from their own SEC filings: managers’ 13F (quarterly, up to 45 days after quarter-end) and funds’ N-PORT (monthly holdings). Money counts the change in stakes net of price moves; every line carries the report date and a delay tag — these are positions as of a past date, not today’s flow: the SEC describes Form 13F as a report of holdings, not a trade history, so a change between two quarters is shown as a change in holdings, never as a purchase made on the filing day. The same fund holdings feed the look-through in Portfolio Guard. Today: seven managers’ latest 13F (Berkshire Hathaway as of Sep 30, 2026, public after 47 days) and five funds’ public N-PORT. ARK Innovation ETF opened SpaceX at 4.9% of the fund and left Roku (N-PORT of Jul 31, 2026); Berkshire roughly doubled its Lennar stake while Lennar’s price fell 9.8% over the quarter — such lines carry the flag. 13F · MANAGERS · NET OF PRICE
Managers' holdings (13F): change in stakes net of priceLoading the engine slice… N-PORT · FUNDS · NET OF PRICE
Fund holdings (N-PORT): change in stakes net of priceLoading the engine slice… More: faster disclosures, three dates, who filedFaster disclosures join the same table: company insiders (SEC Form 4, within two business days of the trade), stakes above 5% (Schedule 13D, within five business days — the deadline since Feb 5, 2024, SEC release 33-11253 ↗), members of Congress (periodic transaction reports: House Clerk ↗, Senate eFD ↗), the daily holdings files some active ETFs publish (ARK: every trading day; the change from one day to the next is shown, not a reported trade) and, beyond the US, company executives’ own dealings (PDMR) under the EU and UK Market Abuse Regulation (within three business days). Every line carries three dates — when it was done, when it became public and when Money recorded it — and names the filer, not a person: a Berkshire Hathaway 13F (EDGAR ↗) shows Berkshire’s holdings, not Warren Buffett’s own trade, and a fund’s action is tied to a manager only when a filing shows that person’s own decision. A filing that shows more shares of a stock while its price fell is flagged; “bought” is used only for a filing that reports a trade (Form 4, Schedule 13D). Today the engine reads the last seven days: Schedule 13D and 13G from their structured EDGAR filings, Germany’s directors’ dealings from the BaFin database and UK PDMR notices from the FCA National Storage Mechanism — each line with the deadline it was due against (13D: 5 business days, amendments 2; MAR Art. 19: 3 business days). 13D · 13G · ACTIVE ETFs · MAR ART. 19
Faster disclosures: three dates on every lineLoading the engine slice… | SEC 13F ↗ · SEC N-PORT ↗ · SEC 13F FAQ ↗ · SEC Form 4 ↗ · SEC 13D rule ↗ · MAR Art. 19 ↗ · BaFin directors’ dealings ↗ · FCA NSM ↗ · ARK daily holdings ↗ |
Beyond Wall Street · as of Nov 15, 2026. For most families the “market” is a flat and gold: financial assets are 25.8% of household wealth in India, about 52% in China and about 79% in the US (UBS Global Wealth Report 2026). Each layer below is a different measure — exchange value, contributions, holdings, turnover — so the layers are never added into one total.
| Layer | Where it stands | Source |
|---|---|---|
| Other exchanges | Market cap at end-September 2026: Asia-Pacific $57.5tn and Europe, Middle East and Africa $25.4tn, against $86.2tn for the Americas. Nasdaq $45.1tn, NYSE $33.6tn; Shanghai $10.2tn, Japan Exchange Group $8.9tn, Euronext $8.0tn, Shenzhen $7.1tn, Hong Kong $5.8tn, Taiwan $4.9tn, BSE and NSE $4.9tn each (largely the same companies — never added), Saudi Exchange $2.4tn. All listed companies worldwide: $151.9tn at end-2025 | WFE, Nov 2026 ↗ · WFE, 2025 ↗ |
| Mass saving routes | India: SIP contributions of ₹32,160 crore in October 2026, from about 10 crore contributing accounts in September (accounts, not people); NSE passed 13 crore unique investors in April. Japan: 28.26mn NISA accounts at end-2025, target 34mn by end-2027. Brazil: Tesouro Direto — 36.4mn registered, 3.83mn active, R$272bn held (August 2026); 60.6mn Brazilians invest (2025). Developing economies: 40% of adults saved in a financial account in 2024, 10% through mobile money | AMFI via Kotak Neo ↗ · 5paisa ↗ · NSE ↗ · FSA ↗ · B3 ↗ · ANBIMA ↗ · World Bank Findex ↗ |
| Property and gold | All real estate: $393.3tn at end-2024 — residential $286.9tn, commercial $58.5tn, farmland $47.9tn (a stock of wealth). Gold: 2,175 t of total investment demand in 2025, 1,374 t of it bars and coins; about 268 t of bars and coins in Q3 2026 | Savills ↗ · WGC 2025 ↗ · WGC Q3 2026 ↗ |
| Pensions | $68.3tn in the 22 largest pension markets at end-2025, 65.8% of it in the US | Thinking Ahead Institute ↗ |
| State money | Sovereign wealth funds above $15tn; all state-owned investors — sovereign funds, public pensions and central banks — about $60tn at end-2025. The largest single fund, Norway’s, held NOK 22,014bn at Sep 30, 2026 (NBIM ↗); ADIA about $1.13tn at end-2025 (Global SWF estimate). Allocation shifts: Live for NBIM (Norway), GIC (Singapore), CIC (China), CalPERS (California) and CPP Investments (Canada) from their own reports; PIF is next. NBIM, PIF, GIC and CIC, from their own reports → how each splits across listed equities, bonds, real estate, unlisted infrastructure and regions, and how that split moved since the previous report, with a delay tag. The same dated view can read public pension and multi-asset manager reports: target allocation versus actual holdings, commitments versus cash called, and liabilities due before an asset can be sold. CalPERS, March 31, 2026: public equity 5.0 pp under its interim target, private equity 4.5 pp over; its private-equity funds have called $130.2bn of $163.0bn committed (79.9%), $45.1bn not yet called; funded status 81.6%. CPP Investments, March 31, 2026: public equities 35.6% of net investments (28.3% a year earlier), unfunded commitments C$80.4bn (C$63.4bn). CIC, end-2024: alternative assets 48.5%, public equity 34.7% (+1.5 pp). Each change points to the fund’s own report; holdings, commitments and flows stay separate, with no ranking by assets. PENSION AND SOVEREIGN FUNDS · OWN REPORTS
Pension and sovereign funds: target versus actualLoading the engine slice… | Global SWF ↗ |
| Off-exchange trading | Foreign exchange $9.6tn a day and OTC interest-rate derivatives $7.9tn a day in April 2025 — turnover, not assets | BIS ↗ |
| Government bonds of the world | For your own date and currency: What stays with me in My money in the world. Here — the country map of the 10 largest government bond markets, then 20 (coverage, not a ranking), with a fiscal-limit monitor, who buys the debt, the maturity wall and index events. Free — the map and the top 10, the method and today’s interest burden, the latest holder facts, the largest maturity wall and the latest index event; $2 — the full table with history, all 20 fiscal limits, holder history, the maturity wall and index events, open to all after 30 days; One — a fiscal-limit alert at your own thresholdCountry map, fiscal limit, who buys, maturity wall, index eventsCoverage: the US, China, Japan, Germany, France, Italy, the UK, India, Brazil and Canada, then South Korea, Mexico, Spain, Australia, Switzerland, Poland, Indonesia, Saudi Arabia, South Africa and Turkey. Country map: the 10-year yield and the 10Y−2Y slope, real yield, spread to US Treasuries and to the Bund, debt to GDP, interest cost as a share of budget revenue, foreign holders’ share, and weight in bond indexes by debt and by GDP — on one data schema, because local series are published on different methods (a zero-coupon point for Germany and the UK, a real yield for Brazil). Fiscal-limit monitor: debt, deficit, average maturity and the maturity wall → the 10-year yield at which interest would pass a stated share of budget revenue, the buffer to it in basis points and a flag: normal, watch, limit. It is a scenario at that threshold: the threshold value and the source of the budget data stand beside the flag; One sets its own threshold. Who buys the debt: TIC, central-bank balance sheets, non-residents, stablecoin reserves — the latest facts free, the history in $2. Central banks as holders: the ECB’s public-sector purchase programme held €1,837.8 bn at book value on Oct 31, 2026, €245.3 bn less than a year earlier as maturing bonds are not reinvested (Germany €423.3 bn, France €359.5 bn, Italy €278.0 bn); the Bank of Japan held ¥504.9 trn of JGBs on Nov 10, 2026, down ¥50.8 trn (−9.1%) in a year (ECB PSPP breakdown, Bank of Japan Accounts). Maturity wall by country, and index events — inclusion in WGBI or GBI-EM and the forced flow that follows. | OECD ↗ · FRED ↗ · ECB Data Portal ↗ · BIS debt securities ↗ · US Treasury TIC ↗ |
| Tokenized government debt | Five kinds of record, never added together: a new digital government bond, a token on an existing bond, a fund share, a collateral (repo) operation, and bills held in stablecoin reserves. Collateral and cash 24/7: a calculator, not a broker — the net yield of each kind of cash after fees and the collateral haircut, and whether the money can be used over a weekend. Each fund is read in Portfolio Guard (what a product actually pays); Citi’s “Tokenization 2030” forecast — in the Accuracy Ledger. Free — the facts and the method; $2 — the daily net-yield table, open to all after 30 days; One — on your amounts; Desk — on your positions; contract — a feed of your counterparties’ termsCollateral and cash 24/7: the kinds of cash and the sourcesFor each kind of cash — a money-market fund, a tokenized bill fund, repo → the net yield after fees and the haircut, weekend availability and, for a tokenized fund, its size on the date from the issuer’s or the chain’s own record; each line with its date and primary source, “planned” where there is none. Sizes are a column of the table, not a ranking of issuers. | Citi GPS ↗ |
| Stablecoins | How many stablecoins are in circulation and what backs them, from the issuers’ own attestations, each figure with its attestation date. Circulation is a stock, so it is never added to flows or to the count of crypto owners. Free — the latest attestations; $2 — series by issuer | Circle ↗ · Tether ↗ |
| Crypto | 741mn owners in 2025 | Crypto.com ↗ |
Allocation shifts of state funds
“Beyond Wall Street” can mean where you buy, where the asset is or how you hold it. A local pension or fund may own US stocks, and local property still moves with global rates — the Atlas keeps those links visible instead of drawing borders. Where a market is hot by sector, it shows on the Country × Sector board.
2027 targets for the S&P 500 span Jefferies’ 9,000, UBS’s 8,900, Barclays’ 8,800, Goldman’s 8,700 (+14%, with earnings growth slowing from 18% to 7%; Yahoo Finance, Nov 13 ↗) and Morgan Stanley’s 8,300 — its horizon is mid-2027, not year-end (the latest set, compiled by Zacks ↗; the next compilation replaces it; each target is now a house forecast in the Journal, with its date and link, and is scored at its horizon against a random picker that draws 5,000 calls on the same day and horizon from the index’s own history). Neuberger Berman favors companies with measurable AI returns over a broad AI bet; Columbia Threadneedle points to semiconductors, Asian EM, local-currency EM debt and commodities, while warning about valuations and concentration; PGIM adds infrastructure debt and parts of private credit.
Banks’ S&P 500 targets against a random picker
Loading the engine slice…
Banks' targets, sealed in the Accuracy Ledger
Loading the engine slice…
| Where capital may go in 2027 | Why | What could stop it |
|---|---|---|
| AI and its infrastructure | Demand for compute and automation | Weak payback, debt, overcapacity |
| Grids and power supply | Data centers and electrification | Construction delays, regulation |
| Companies lifting profits with AI | Measurable productivity gains | AI gains never reach earnings |
| High-quality bonds | Income and diversification | Inflation and rate hikes |
| Infrastructure and private credit | Projects need financing | Defaults, low liquidity |
| Gold | Hedging specific risks | High real yields, a strong dollar |
InvestmentNews ↗ · ETF Trends ↗ · State Street ↗ · iShares ↗ · S&P Global · ETF review Q1 2026 ↗ · ETFGI · 2025 ↗ · ETFGI · Jan 2026 ↗ · ETFGI · Q1 2026 ↗ · McKinsey · Asset management 2026 ↗ · World Gold Council ↗ · Crunchbase ↗ · IEA ↗ · Morningstar ↗ · Neuberger Berman ↗ · Columbia Threadneedle ↗
Government bonds of the world: the full table, the risk behind the rate, and history
Country map: 10-year yields, slope, real yield, spreads, debt, interest, foreign holders, weights · slice as of 2026-11-16
| country | 10Y, % | 10Y−2Y, bp | real, % | spread to Bund, bp | debt, % GDP | interest, % revenue | foreign holders, % | weight by debt, % |
|---|---|---|---|---|---|---|---|---|
| United States | 4.99 2026-09 | 24.00 2026-11-12 | 2.86 2026-11-12 | 157.82 2026-09 | 124.38 2025 | 10.42 2025 | 29.26 2026-08 | 39.42 2025 |
| China | 1.68 2026-09 | in development | in development | -179.41 2026-09 | 98.61 2025 | 3.85 2025 | in development | 19.88 2025 |
| Japan | 3.05 2026-09 | 101.70 2026-11-13 | in development | -40.08 2026-09 | 204.78 2025 | 0.55 2025 | in development | 9.37 2025 |
| Germany | 3.47 2026-09 | 28.00 2026-11-13 | in development | — | 62.66 2025 | 1.82 2025 | in development | 3.29 2025 |
| France | 4.44 2026-09 | in development | in development | 96.59 2026-09 | 115.53 2025 | 3.98 2025 | in development | 4.01 2025 |
| Italy | 4.36 2026-09 | in development | in development | 93.55 2026-09 | 137.10 2025 | 7.51 2025 | in development | 3.59 2025 |
| United Kingdom | 5.24 2026-09 | in development | 2.02 2026-11-11 | 183.67 2026-09 | 102.06 2025 | 6.82 2025 | in development | 4.21 2025 |
| India | 7.06 2026-09 | in development | in development | 371.05 2026-09 | 84.26 2025 | 24.58 2025 | in development | 3.38 2025 |
| Brazil | 9.14 incomparable 2026-09 | 55.00 2026-11-13 | 7.77 2026-11-13 | no data | 93.33 2025 | 18.46 2025 | in development | 2.19 2025 |
| Canada | 3.88 2026-09 | in development | in development | 44.61 2026-09 | 113.58 2025 | -0.55 2025 | in development | 2.71 2025 |
Rules: Each cell has its date and source. Real and nominal yields are separate columns; spreads only between comparable points of the same month; empty cells are 'plan' or 'no data', never 0. Index weights: the 20 weighted by debt and by GDP — the public rule, not a provider's weights.
Sources: OECD DF_FINMARK IRLT (10-year government bond yields, monthly) ↗ · FRED DGS10 ↗ · FRED DFII10 ↗ · www.imf.org ↗ · US Treasury TIC SLT table 5 ↗ · Ministry of Finance Japan — JGB interest rates (historical + current) ↗ · Deutsche Bundesbank BBSIS ZAR 10y ↗ · Bank of England IUDMNPY ↗ · Bank of England IUDMRZC ↗ · Tesouro Direto — preços e taxas ↗
Yield points on one schema and the recount checks · slice as of 2026-11-16
countries with data 17 · points 35 · shown 10
| country | instrument | date | % | annual-compounded % | comparability |
|---|---|---|---|---|---|
| USA | FRED DGS10 | 2026-11-12 | 5.100 | 5.165 | comparable |
| DEU | BBSIS ZST R10XX | 2026-11-13 | 3.430 | incomparable | |
| GBR | BoE IUDMNZC | 2026-11-11 | 5.211 | incomparable | |
| DEU | BBSIS ZAR R10XX | 2026-11-13 | 3.410 | 3.410 | comparable |
| GBR | BoE IUDMNPY | 2026-11-11 | 5.204 | 5.272 | comparable |
| JPN | MOF JGB 10Y | 2026-11-13 | 2.899 | 2.920 | comparable |
recount check Fed GSW curve: zero-coupon → par: zero -0.0022, par -0.0045 bp, within 1 bp: yes 2026-11-06
recount check Treasury auction: price → yield: 0.0021 bp, within 1 bp: yes 2026-11-10
Daily yields beyond these publishers come from the central banks and clearing houses themselves: on Nov 16, 2026 the 10-year stood at 3.40% for AAA euro-area governments, 3.81% in Canada, 3.11% in Sweden, 4.39% in Norway, 1.63% in China and 8.70% in South Africa (R209, 2036); India’s 6.94% G-sec 2036 yielded 7.26% on Nov 17 and Brazil’s NTN-F 2037 15.21%. Instead of a CDS — the spread on the same day: China −349 bp to the US 10-year, Canada −131 bp, South Africa +358 bp, Brazil +1,009 bp; India +387 bp to the Bund (ECB, Bank of Canada, RBA, Riksbank, Norges Bank, SARB, ChinaBond, CCIL, Tesouro Direto).
Rules: One record per point for any country. Restated to an annual-compounded yield to maturity only where the method allows; a zero-coupon point is not a yield to maturity and stays 'incomparable' unless the full curve gives a par yield; real and nominal yields never share a column.
Sources: FRED DFII10 ↗ · FRED DGS10 ↗ · FRED DGS2 ↗ · FRED DGS20 ↗ · OECD DF_FINMARK IRLT (10-year government bond yields, monthly) ↗ · Ministry of Finance Japan — JGB interest rates (historical + current) ↗ · Deutsche Bundesbank BBSIS ZAR 2y ↗ · Deutsche Bundesbank BBSIS ZAR 10y ↗ · Deutsche Bundesbank BBSIS ZST 2y ↗ · Deutsche Bundesbank BBSIS ZST 10y ↗ · Bank of England IUDMNPY ↗ · Bank of England IUDMNZC ↗ · Bank of England IUDMRZC ↗ · Bank of England IUDSNZC ↗ · Tesouro Direto — preços e taxas ↗ · Federal Reserve Board — GSW nominal yield curve (feds200628) ↗ · TreasuryDirect auctions (Note) ↗
Fiscal-limit monitor (scenario at a threshold of 15% of revenue) · slice as of 2026-11-16
| country | interest / revenue on the slice date, % | critical 10Y, % | margin, bp | flag |
|---|---|---|---|---|
| United States | 10.42 | 4.76 | -29.3 | limit (scenario at this threshold) |
| China | 3.85 | 3.37 | 169.0 | attention |
| Japan | 0.55 | 3.87 | 80.0 | attention |
| Germany | 1.82 | 9.85 | 637.1 | normal |
| France | 3.98 | 6.14 | 170.0 | attention |
| Italy | 7.51 | 5.55 | 114.0 | attention |
| United Kingdom | 6.82 | 5.88 | 57.1 | attention |
| India | 24.58 | 4.80 | -238.1 | limit (scenario at this threshold) |
| Brazil | 18.46 | 8.56 | -673.5 | limit (scenario at this threshold) |
| Canada | -0.55 | 21.28 | 1,736.4 | normal |
Rules: Critical 10-year rate at which interest / revenue reaches the threshold within the horizon as old debt is refinanced; margin = critical − current 10Y. A scenario at a given threshold, not a default point.
Sources: www.imf.org ↗ · fiscaldata.treasury.gov ↗ · OECD DF_FINMARK IRLT (10-year government bond yields, monthly) ↗ · Tesouro Direto — preços e taxas ↗
Who holds the debt · slice as of 2026-11-16
TIC 2026-08 · Holdings by country of the custodian, not of the owner (TIC method)
| holder (TIC: country of the custodian) | USD bn | 1 month | 12 months |
|---|---|---|---|
| Japan | 1,068.3 | -35.6 | -115.6 |
| United Kingdom | 1,009.5 | 11.2 | 108.2 |
| China, Mainland | 590.7 | -27.3 | -109.0 |
| Belgium | 490.5 | 19.8 | 39.4 |
| Canada | 464.6 | 38.3 | 20.6 |
| Cayman Islands | 463.0 | 2.9 | 52.5 |
| Luxembourg | 448.2 | 6.1 | 37.4 |
| France | 363.4 | 15.0 | -25.7 |
| Ireland | 352.1 | 1.9 | 12.1 |
| Taiwan | 298.4 | 2.3 | -15.0 |
| Switzerland | 285.8 | 1.0 | -17.2 |
| Singapore | 277.9 | 0.2 | 16.9 |
| Hong Kong | 261.5 | 3.2 | 9.6 |
| India | 219.7 | 17.1 | 0.3 |
| Norway | 214.0 | 7.1 | -0.9 |
central bank: USA: Federal Reserve (H.4.1, Treasury securities held outright) — 4,586.6 USD bn (2026-11-11) · JPN: in development · GBR: in development · DEU: in development
Sources: US Treasury TIC SLT table 5 ↗ · FRED TREAST ↗
Maturity wall · slice as of 2026-11-16
USA · 2026-10-31 · largest year: 2027 — 6,715.8 USD bn
| year | USD bn |
|---|---|
| 2026 | 4,791.0 |
| 2027 | 6,715.8 |
| 2028 | 3,735.7 |
| 2029 | 2,664.4 |
| 2030 | 2,057.5 |
| 2031 | 2,026.2 |
| 2032 | 1,200.5 |
| 2033 | 1,010.5 |
| 2034 | 635.0 |
| 2035 | 675.7 |
| 2036 | 538.9 |
| 2037 | 38.0 |
JPN, DEU, GBR, FRA, ITA, CAN, BRA, IND, CHN: in development
Sources: fiscaldata.treasury.gov ↗
Index events · slice as of 2026-11-16
| index | country | announced | effective | forced flow |
|---|---|---|---|---|
| FTSE World Government Bond Index (WGBI) | KOR | 2024-10-08 | 2025-11-01 | in development |
| FTSE Emerging Markets Government Bond Index (EMGBI) | IND | 2024-10-08 | 2025-09-01 | in development |
| Bloomberg EM Local Currency Government Index | IND | 2024-03-05 | 2025-01-31 | in development |
Rules: Forced flow = weight × assets that follow the index; without a primary source for those assets it is not shown as a number.
Sources: FTSE Russell — FTSE World Government Bond Index (WGBI) ↗ · Bloomberg Index Services — Bloomberg EM Local Currency Government Index ↗
Government bond auctions · slice as of 2026-11-16
| issuer | date | term | bid-to-cover | high yield, % | tail, bp | tail, σ |
|---|---|---|---|---|---|---|
| USA | 2026-11-12 | 30-Year | 2.36 | 5.441 | 9.1 | 1.62 |
| USA | 2026-11-10 | 10-Year | 2.46 | 5.020 | 14.0 | 2.34 |
| USA | 2026-11-09 | 3-Year | 2.53 | 4.627 | 3.7 | 1.08 |
| USA | 2026-10-29 | 7-Year | 2.46 | 4.894 | -5.6 | -0.65 |
| USA | 2026-10-27 | 5-Year | 2.33 | 4.845 | 0.5 | 0.04 |
| USA | 2026-10-26 | 2-Year | 2.59 | 4.708 | 1.8 | 0.65 |
| USA | 2026-10-21 | 20-Year | 2.73 | 5.469 | -0.1 | -0.07 |
| USA | 2026-10-08 | 30-Year | 2.54 | 5.618 | -5.2 | -1.15 |
| USA | 2026-10-07 | 10-Year | 2.77 | 5.300 | 3.0 | 0.28 |
| USA | 2026-10-06 | 3-Year | 2.62 | 4.932 | -3.8 | -1.59 |
Next auctions
| issuer | date | term | USD |
|---|---|---|---|
| USA | 2026-11-18 | 20-Year | 16.00 bn |
| USA | 2026-11-19 | 10-Year | 19.00 bn |
| USA | 2026-11-23 | 2-Year | |
| USA | 2026-11-24 | 1-Year 11-Month | |
| USA | 2026-11-24 | 5-Year | |
| USA | 2026-11-25 | 7-Year |
history: 500 auctions, 29 bond-vigilante flags
other issuers (DEU, GBR, JPN, ITA, FRA, CAN, BRA, IND, CHN): in development
Rules: Tail = high yield − pre-auction yield; no pre-auction yield → no tail. Surprise in σ against the issuer's own last 24 auctions of the same term. The vigilante flag needs all three: weak auction, term premium up and currency down in the same week.
Sources: TreasuryDirect auctions (Note) ↗ · TreasuryDirect auctions (Bond) ↗
Bond history: what happened after a yield crossed a level · slice as of 2026-11-16
eras: Nov 2011 · Italy 10-year near 7% · May 2013 · eve of the taper tantrum · Sep 2022 · the day before the UK mini-budget (gilt stress) · Dec 2022 · BoJ widens the yield-curve band · Jul 2023 · BoJ makes YCC 'flexible' · Oct 2023 · BoJ: 1% becomes a reference · Mar 2024 · BoJ ends YCC and negative rates
Japan 10-year government bond yield (daily) above 1.0% — what happened next
| crossed on | value, % | +3 months | +6 months | +12 months |
|---|---|---|---|---|
| 2003-07-03 | 1.096 | 1.383 | 1.363 | 1.785 |
| 2010-08-27 | 1.029 | 1.187 | 1.255 | 1.066 |
| 2011-09-08 | 1.011 | 1.012 | 0.984 | 0.811 |
| 2012-03-14 | 1.004 | 0.850 | 0.794 | 0.615 |
| 2024-05-22 | 1.007 | 0.919 | 1.092 | 1.559 |
United Kingdom 10-year government bond yield (daily) above 4.5% — what happened next
| crossed on | value, % | +3 months | +6 months | +12 months |
|---|---|---|---|---|
| 2022-09-27 | 4.527 | 3.820 | 3.471 | 4.566 |
| 2023-06-19 | 4.527 | 4.271 | 3.620 | 4.078 |
| 2024-11-05 | 4.553 | 4.448 | 4.452 | 4.467 |
| 2025-05-08 | 4.538 | 4.589 | 4.487 | 4.870 |
| 2025-11-14 | 4.573 | 4.417 | 5.126 |
Italy 10-year government bond yield (OECD, monthly average) above 4.0% — what happened next
| crossed on | value, % | +3 months | +6 months | +12 months |
|---|---|---|---|---|
| 2006-12-01 | 4.038 | 4.176 | 4.772 | 4.538 |
| 2010-06-01 | 4.102 | 3.862 | 4.603 | 4.816 |
| 2013-06-01 | 4.379 | 4.541 | 4.114 | 2.920 |
| 2022-09-01 | 4.138 | 4.257 | 4.240 | 4.513 |
| 2026-09-01 | 4.362 |
United States 10-year government bond yield (OECD, monthly average) above 4.5% — what happened next
| crossed on | value, % | +3 months | +6 months | +12 months |
|---|---|---|---|---|
| 2005-11-01 | 4.540 | 4.570 | 5.110 | 4.600 |
| 2023-10-01 | 4.800 | 4.060 | 4.540 | 4.100 |
| 2024-04-01 | 4.540 | 4.250 | 4.100 | 4.280 |
| 2025-01-01 | 4.630 | 4.280 | 4.390 | 4.210 |
| 2026-07-01 | 4.600 |
Rules: base = last value known at T 09:00 ET; h-month result = last value whose known_at ≤ (T + h months) 09:00 ET
Sources: OECD DF_FINMARK IRLT (10-year government bond yields, monthly) ↗ · FRED DGS10 ↗ · FRED DGS2 ↗ · FRED DGS3 ↗ · FRED DGS5 ↗ · FRED DGS7 ↗ · FRED DGS20 ↗ · FRED DGS30 ↗ · FRED DFII10 ↗ · FRED TREAST ↗ · FRED THREEFYTP10 ↗ · FRED DTWEXBGS ↗ · FRED SOFR ↗ · FRED DTB3 ↗ · FRED DEXBZUS ↗ · FRED DEXINUS ↗ · FRED DEXUSEU ↗ · FRED DEXUSUK ↗ · FRED DEXJPUS ↗ · Deutsche Bundesbank BBSIS ZST 2y ↗
Government bond auctions outside the US
UK gilt auctions
Cash 24/7, tokenized government debt and stablecoins
Tokenized government debt: five kinds of record, never added · slice as of 2026-11-16
| record | kind | issuer | date | size |
|---|---|---|---|---|
| HKSAR-DIGITAL-GREEN-2024 ↗ | new digital government bond | HKSAR Government | 2024-02-07 | 6.00 bn HKD |
| BUIDL-ETHEREUM ↗ | fund share | BlackRock USD Institutional Digital Liquidity Fund (BUIDL), Securitize transfer agent | 2026-11-16 | 170.28 mn USD |
| TOKEN-ON-EXISTING-BOND | token on an existing bond | no issuer record with a primary source wired yet | in development | |
| TOKENIZED-REPO | collateral (repo) operation | repo platforms (e.g. intraday DLT repo) publish monthly volumes in press releases; parser plan | in development | |
| STABLECOIN-RESERVE-BILLS | bills in stablecoin reserves | Treasury bills in stablecoin reserves come from issuers' attestations (snapshot stablecoins, §35) | see stablecoins |
Rules: The five kinds are different records: never summed with each other, with stablecoins or with money-market funds; a wrapper is not a second count of the same bills.
Citi «Tokenization 2030»: in development a complete open counter of tokenized assets is not wired; the measured part is not divided by the 5.5 trillion target
Sources: HKSAR Government — HKSAR-DIGITAL-GREEN-2024 ↗ · Ethereum eth_call totalSupply 0x7712c34205737192402172409a8F7ccef8aA2AEc ↗
Cash 24/7: net yield after fees and haircut · slice as of 2026-11-16
| kind of cash | yield, % | fees, % | haircut | weekend | net, % |
|---|---|---|---|---|---|
| 3-month Treasury bill held directly (TreasuryDirect) | 4.14 | 0.00 | 0.00 | no | 4.14 |
| Overnight Treasury repo (SOFR) | 3.90 | in development | in development | no | in development |
| Tokenized Treasury money-market fund (BUIDL) | in development | in development | in development | yes | in development |
| Government money-market fund | in development | in development | 0.00 | no | in development |
net = yield − fees − haircut × funding cost
Rules: A row without a primary source for its yield or haircut is 'plan', not a number. A calculation, not a brokerage function: no orders, no advice on releasing collateral.
Sources: FRED DTB3 ↗ · www.treasurydirect.gov ↗ · FRED SOFR ↗
Stablecoins: circulation and reserves from the issuers · slice as of 2026-11-15
| issuer | coin | in circulation | reserves (total assets) | attestation |
|---|---|---|---|---|
| Tether | USDT | 175.40 bn USD | 180.86 bn | date not in the issuer’s feed |
| Tether | XAUT | 764,235 troy_ounce | 827,418 | date not in the issuer’s feed |
| Tether | MXNT | 19.56 mn MXN | 19.56 mn | date not in the issuer’s feed |
| Circle | EURC | 472.01 mn EUR | monthly report, date on the linked page | |
| Circle | USDC | 69.07 bn USD | monthly report, date on the linked page |
flows: accumulating in development
Rules: issuer's own publication only · stock (circulation, reserves) and flow (net issuance) are not added
Sources: tether.to ↗ · Tether transparency ↗ · www.circle.com ↗ · Circle stablecoins API ↗
Spot the regime shift — don’t guess tomorrow’s price
Choose an asset. See the bull case, the bear case, the arbiter’s read, the model votes and three conditional scenarios. Facts and votes come from the council’s answer of Nov 14, 2026; probabilities come from the live council.
From “I have a hunch” to checked reasons to act: one card turns a guess into fields you can test and names the one unknown that would change your decision and the most useful way to check it.
What Microsoft, Alphabet, Amazon, Meta and Oracle spend on data centres each quarter, what their suppliers sell, and only the contracts a supplier announced naming the buyer.
Data-centre debt: who lends, how it is marked, when it is due
Loading the engine slice…
AI campuses from space: how much warmer than the land around them
Loading the engine slice…
More: how the payback checks run — inputs, outputs, access
More: Nov 14 research note on capex scope and 2027 funding
Follow the financing. The IEA reports more than $400bn of 2025 capex for five large technology companies and a further 75% rise in 2026. Moody’s estimates $785bn of 2026 capex for six US hyperscalers. WTW’s February 2026 outlook projects more than $1.3tn of capex for the six largest US cloud companies across 2026–27, with 2026 more than 40% above 2025 — a two-year forecast, not the meter’s one-year total. Goldman Sachs models $765bn of annual AI infrastructure investment in 2026 and $7.6tn cumulatively through 2031; its model includes chips, data centers, power and supporting infrastructure. ING counts about $237bn of hyperscaler bond issuance across currencies already in 2026 and forecasts about $280bn in 2027, including about $190bn in US dollars. J.P. Morgan Asset Management puts capex at about 93% of the five hyperscalers’ operating cash flow. Money’s own meter above counts only final SEC prospectuses (424B2/424B5) for US-dollar notes and cash purchases of property and equipment, so its figures sit next to these estimates but are not the same measure. These forecasts have different company sets and spending definitions. The IEA’s 2026 study Powering Artificial Intelligence puts data-center operators’ 2026 capex above $700bn, mostly new data centers, and points to long grid-connection queues and slow permitting (the regions and the register in Physical Data Radar) — which is why the payback check dates a booked contract and a powered facility separately. IEA · Powering AI ↗ · IEA ↗ · Moody’s ↗ · WTW ↗ · Goldman Sachs ↗ · ING ↗
IMF · GFSR Oct 2026 ↗ · BlackRock · capital competition ↗ · BCG ↗ · IEEE ComSoc ↗ · Tech Insider / Fortune ↗
Does AI spending pay back? Three checks per company
the three payback checks, company by company
AI-revenue payback: three checks per company · T = 2026-11-15
US
| Company | Standard | 1 · Who pays | 2 · What remains | 3 · Will it repeat |
|---|---|---|---|---|
| MSFT | US GAAP | not disclosed | passed | passed |
| NVDA | US GAAP | not disclosed | passed | not indicative |
| ORCL | US GAAP | not disclosed | passed | passed |
| AMZN | US GAAP | not disclosed | passed | not disclosed |
| GOOGL | US GAAP | not disclosed | passed | passed |
| META | US GAAP | not disclosed | not passed | not disclosed |
Asia
| Company | Standard | 1 · Who pays | 2 · What remains | 3 · Will it repeat |
|---|---|---|---|---|
| TSM TSMC | IFRS | not disclosed | not disclosed | not disclosed |
| BABA Alibaba | US GAAP | not disclosed | not passed | not disclosed |
| 005930.KS Samsung Electronics | — | — | — | |
| 000660.KS SK hynix | — | — | — | |
| 0700.HK Tencent | — | — | — | |
| 9984.T SoftBank Group | — | — | — |
no sums across companies or columns: one dollar would be counted at the seller and the buyer, and different kinds of revenue would be added; no payback date is produced. A person checks 5 results before they count: waiting for a person in development
Engine slice v43 ai_payback · as of 2026-11-15 22:20:57 UTC · sources: SEC 10-K 0001564590-20-034944 ↗ · SEC 10-K 0001564590-21-039151 ↗ · SEC 10-K 0001564590-22-026876 ↗ · SEC 10-K 0000950170-23-035122 ↗ · SEC 10-K 0000950170-24-087843 ↗ · SEC 10-K 0001193125-26-323660 ↗ · SEC companyfacts MSFT ↗ · SEC 10-K 0001045810-26-000021 ↗ · +29
More: bubble test — when a run turns explosive
Bubble test — when a run turns explosive
Gold, copper, the Nasdaq, bitcoin and US house prices (also against rents): a recursive explosive-root test (Phillips, Shi & Yu, 2015 ↗) dated observation by observation. It dates the start and end of a run; it sets no price target.
Gold: explosive phase since Apr 2024 (slice of Nov 17, 2026; October 2026 average $4,105 after the February record of $5,020).
The bubble test now covers the S&P 500 since 1871 on Robert Shiller’s monthly series: the price/dividend ratio has no open phase (the last ended in May 2009; the test also dates 1997–2001), the real price has been in an explosive phase since June 2024, CAPE 39.4 — above 98% of months since 1881. Across the 12 Lab markets and 10 board markets (weekly since 2002), two are in an explosive phase today: the PHLX Semiconductor index since Jun 27, 2025 and Taiwan’s TAIEX since Nov 28, 2025 (Shiller data; PSY 2015).
Where money is hot — and where the odds split
Pick a market and a sector on the board, then compare council probabilities with prediction-market prices.
Heat map: market × model
Colour is the side (green — up, red — down), strength is the confidence. Pick the horizon.
Loading… Trends ↗
12 markets, 4 horizons: what the AIs would do
Sealed before the New York open. Pick a market: the Council’s call, every model’s target, entry and “wrong if”, the 30-session chart with those levels, and the crowd’s price.
Loading the day’s sealed answers… Open on Trends ↗
Where is capital flowing — which market, which sector?
Pick any cell — such as A7, Taiwan × chips — or narrow the board by market and sector: you get its coordinate, the dated fact behind it when there is one, the local exchange and a ready question for the council. Views: board, by market, by sector, map, flows.
| Market ↓ · sector → | A · Chips & hardware | B · Software & internet | C · Energy & power | D · Financials | E · Industrials & robotics | F · Materials & metals |
|---|---|---|---|---|---|---|
| 1 · United States | ||||||
| 2 · China (mainland) | ||||||
| Stock Connect carries about 71% of international investors’ holdings of mainland A-shares — around RMB 2.6tn at end-2025. Hong Kong is the gate between China and outside money. HKEX ↗ | ||||||
| 4 · Japan | ||||||
| 5 · India | ||||||
| Korean retail investors bought $7.63bn more US stocks than they sold in Q3 2026: local savings heading abroad — a cross-border flow the board follows (see the Flows view). Korea Times ↗ | ||||||
| 7 · Taiwan | Semiconductors were 52.22% of Taiwan Stock Exchange market capitalisation at end-2025: 94 companies, NT$49.28tn. Here a study of the chip supply chain says more than a view on “Asian equities”. TWSE Fact Book 2026 ↗ | |||||
| 8 · Europe | ||||||
| Since Feb 1, 2026 the main market is open to direct investment by all foreign investors, individuals included, with no QFI status required (CMA, announced Jan 6). A change in access is itself an event the board tracks. CMA ↗ · Regulation Tomorrow ↗ | ||||||
| 10 · Brazil |
Every cell with its numbers: the week’s move, what drove it, momentum and rank
Loading the board slice…
Loading the rotation…
- 6 → 1 · South Korea → United States. Korean retail investors bought $7.63bn more US stocks than they sold in Q3 2026 (a flow over one quarter). Korea Times ↗
- 3 → 2 · Hong Kong → China (mainland). International investors held around RMB 2.6tn of mainland shares through Stock Connect at end-2025, about 71% of their total — a holding at a date, not a flow, so it is never added to one. HKEX ↗
Marked cells and rows carry a dated, sourced fact; every other cell can still be picked, filtered and taken to the council. Colour by strength comes from the engine’s daily score, the flows from each market’s fund files, the rotation from weekly relative strength against MSCI ACWI; no view draws a made-up score.
Method: what goes in, how the score works, what comes out
Investors read two planes at once: where (the market) and what (the sector). The board puts both on one grid, like coordinates in Battleship: a hot spot is a cell such as A7 — Taiwan × chips. Every view draws the same dated snapshot, and rotation joins them with the engine’s scores.
What “country” means on the board
Loading the board…
Where AI odds and market odds split
Compare council probabilities with prediction-market prices and futures. Kalshi and Polymarket reached a record $50.6bn in July 2026 (Tangem ↗), and Kalshi a $15.27bn record week in September (DeFi Rate ↗). A wide gap flags an idea to investigate, with both sides laid out; after the event, the Accuracy Ledger shows who was closer — the crowd with money at stake or the council.
Live on 12 markets: on 49 crowd contracts the AIs and the crowd take the same side 92% of the time, but the crowd sits 22.7 points from 50 and the AIs 8.9: money is bolder. Each day the widest gaps lead (the Spread, the Crowdometer), and the Polymarket–Kalshi pair shows one market at two prices. The free divergence of the day: the dollar index — the crowd 15.6 points pricier than the Council (live page, Nov 16, 2026). Cut-off: session of Nov 13, 2026, data Nov 16, 2026, 11:38 ET, export ↗.
Loading today’s crowd prices… Crowdometer on Trends ↗
More: beyond the US, like-for-like contracts, one asset at two prices
Loading the latest survey…
Loading the comparability check…
Loading today’s pairs…
Closed-end funds: price against net asset value
Loading the engine slice…
Map +100% · where the AIs saw a doubling
Sealed before the open: each AI could name one market that may double within a week. Open on Trends ↗
Doubling in a week is rare. These are the AIs' own guesses; we score them when the week closes, and misses are shown too.
East against West
The same question, the US models against the Chinese ones: the mean chance of a rise per group, market by market.
Loading… East against West on Trends ↗
Ratings of the day
Today’s top lists load from Trends. Ratings of the day ↗
“10-year above 5%.” What happened next?
Money finds past episodes, shows what followed and tests which similarities still matter.
History Engine: how an era is like today and how it differs
Five eras, each given as it looked on the date T. The outcome and the lesson stay closed until you press Reveal, because the name is hindsight.
| Episode | Backdrop on date T | 10-year after 12 mo. | Gold after 12 mo. | Copper after 12 mo. | Similar to today | Not similar |
|---|---|---|---|---|---|---|
| Jun 1, 2007 · 10-year near 5% | Fed funds at 5.25% for a year; several subprime mortgage lenders have collapsed since February | High rates, debt rising | Then the debt sat with households and mortgage lenders; now it sits with AI companies | |||
| Oct 18, 2023 · 10-year near 5% | Fed funds at 5.25–5.50% after 11 hikes; AI capex rising | AI capex, concentration of leaders | Then the Fed had paused after 11 hikes; now the market expects a hike | |||
| Sep–Nov 2026 · 5.1–5.2% | Hot inflation, capex $0.7–0.8tn, debt for AI | → distribution across analogs + the Council 10+ assessment | The council explains which episode is closer and why | |||
Pick a date. See the market as it looked that morning
Hindsight makes every crash look obvious. Replay rebuilds a past morning from what was public by 09:00 ET on date T: the 10-year yield as published (it is never revised), payrolls as first reported (ALFRED), and copper and gold cut at T by their publication lag. Press “Reveal” to load what happened next.
Both episodes start with the 10-year near 5%. Press “Reveal in replay” to see what followed each one; the figures come from the engine’s after-T snapshot. The replay has 63 prepared dates since 1971 for the 10-year, copper and gold; oil, the Dow and currencies are next. Government bonds of every country: the same question on the 20 bond markets of the Atlas — “Japan’s 30-year above X: what happened next?” — with the 2022 gilt stress and the Bank of Japan’s exits from yield-curve control among the episodes. Free — a demo; $2 — every date.
How Money asks history an honest question — and the backtest safeguards
“Using only what was known on date T, can we improve the risk or return estimate for asset X over horizon H?”
“Feed the model the whole internet” is not a research design. Qlib and LEAN already exist to test models and strategies; Money competes on better data, a sharper question and a stricter test. History sets ranges, scenarios and risk. Forecasts are locked before the outcome and scored on genuinely new data — a model can earn its keep by managing risk better even when it can’t call tomorrow’s price.
| Market | Hypotheses Money tests |
|---|---|
| Gold | Real yields, inflation expectations, the dollar, ETF and central-bank flows |
| Copper | Industrial demand and Chinese construction, inventories, mine disruptions and new capacity, grid build-out |
| Equity index | Earnings and margins, valuations, rates and credit spreads, buybacks, market breadth, index composition and concentration |
- Overfitting → walk-forward testing and an untouched sample.
- Look-ahead → only data published by date T, using ALFRED vintages for macro series.
- Survivorship → index members as they stood on each date, including delisted companies.
- Costs → fees, bid-ask spreads and liquidity in every result — plus each exchange’s own rules from the Market Profile: transaction taxes such as India’s securities transaction tax, lot sizes, daily price limits and settlement (trading days: exchange_calendars ↗), so a trade that could not have been executed is rejected. Free — the exchange-rules reference; Pro — applied to your backtest. Worked today on six listings — Shanghai, Taipei, KOSPI, NSE, Tokyo and Tadawul — with one rule (hold the stock above its 200-day average, otherwise cash): over ten years the transaction tax alone took 10.6 points from TSMC’s result (+496% frictionless, +484% by the exchange’s rules) and 2.1 points from Reliance’s; a sell order on Tokyo’s Greenery Day 2018 was not filled and waited for the next session.EXCHANGE RULES IN THE BACKTEST · SIX MONEY
Loading the exchange-rules backtest…
- Regime change → results reported by regime, not as one average — local regimes too: Japan 1989–90, Asia 1997–98, UK gilts 2022, Brazil 2015, with cash at the local central-bank rate and the result in the currency you spend. Regime comparison: an asset, a spending currency and a horizon → regimes set from inflation, real yield, currency and volatility as known on date T (a hidden Markov model, statsmodels) → the range of outcomes for that asset and for cash in past matching regimes, how many matches and which dates, and the condition that breaks the analogy; fewer than 10 matches → “regime outside the analogs,” with no probability. It feeds Portfolio Guard and the Lab. Free — the method and one dated example; $2 — the shared regime map while fresh, open after 30 days; Pro — your strategy by regime. Worked today: on September 2026 data the four drivers read high inflation, a high real yield, a weakening dollar and calm markets; over three months that regime has 12 independent past matches — gold’s median +0.6% (10–90%: −5.9% to +15.6%) against cash +1.0% — while at six and twelve months there are fewer than 10, so the answer is “outside the analogs”, with no range. Scores that waited for calibration are now published, each with its interval and its distance from a benchmark (walk-forward 1973–2022, Brier score against the base rate): a US recession within 12 months — 27% (90% interval 18–38%, Oct 2026), +0.054 ±0.026 better than the benchmark (skill +0.27); a negative S&P 500 real total return over 12 months — 31% (25–37%), −0.002 ±0.002 against the benchmark: the regime model does not beat the base rate yet, and the label says so (FRED, NBER dates, Shiller).REGIME COMPARISON · FOUR DRIVERS ON DATE T
Loading the regime comparison…
- Multiple testing → next in the engine: every variant tried is logged, the Sharpe ratio is deflated for the number of trials, and a single test across all variants at once (White’s Reality Check) runs across all variants — a thousand generated strategies will always contain a lucky one.
- Model memory → a frontier LLM may already know how 2020 ended; telling it “ignore the future” isn’t enough. Money grades LLMs only on forecasts made after their training cutoff, an issue the Federal Reserve and research on time-restricted models both study. Gao, Jiang and Yan built a test for it: in their headline application, a one-standard-deviation rise in a news item’s lookahead propensity strengthens the LLM signal by about 32% of its standalone effect — a sign the model partly remembers rather than forecasts; it holds for that test, not for every model (v2, June 2026, arXiv ↗). Our own measurement: the same fact pack up to date T runs twice — with the real names and dates and anonymized; the gap between the answers is each model version’s “memory gap,” and a backtest with a significant gap stays out of the Accuracy Ledger. Free — the method and the verdict on every backtest; the gap matrix by model stays internal (it feeds the Router). On Nov 17, 2026: 10 model versions from 4 families measured on one pack of 108 items; 5 answer significantly better with real names on dates they may remember, and their backtests over those dates stay out of the Ledger.
On Nov 17, 2026: 10 model versions measured on the same pack of 108 items, 12 markets — 5 answer significantly better with real names on dates they may remember.
Why the bar is this high. An April 2026 NBER working paper followed portfolios picked by ChatGPT, Claude, Gemini and Grok: raw returns beat the S&P 500, but the abnormal return after characteristic adjustment was not statistically significant — the picks leaned on momentum, large companies and media attention (NBER w35153 ↗). AMCIS 2026 found Gemini Deep Research no better than a coin flip on Nasdaq-100 stock direction. In Gu, Kelly and Xiu, the best model explained only ~0.4% of an individual stock’s monthly return. In October–December 2025 live-money contests such as Alpha Arena, most models lost money, overtraded and gave inconsistent calls on identical data. Open time-series models — Kronos, trained on 12+ billion price bars, plus TimesFM, Chronos-2 and Moirai — beat traditional methods on several tasks (dated research, 2025–2026). Money’s bar is the statsforecast baseline (naive, ETS); its challenger — the mean of AutoARIMA, AutoTheta and AutoETS — is sealed in the Ledger beside it and scored the same way: on Nov 17, 2026 it beat Naive in 5 of 9 series-horizons (MASE). See What’s inside.
Challenger (mean of AutoARIMA, AutoTheta, AutoETS) against the statsforecast baseline, MASE on the same origins.
Free series to start from: FRED copper since 1992 ↗ · World Bank Pink Sheet (gold) ↗ · FRED Dow ↗
QuantConnect · research guide ↗ · Federal Reserve · FEDS 2025-044 ↗ · Qlib ↗ · ThinkAdvisor ↗ · TSFM in finance ↗ · Chronos-2 ↗
Your money. Your goal. Your currency
Start from what you hold and what it is for — Money measures every move in the currency you will spend.
Make your call first. Seal it before seeing the Council, then compare the reader experiment in the Live score.
Decision Mirror: did you follow your own rules?
Write down what you hold, why, for how long and what would change your mind. On the review date the Mirror shows the price then and now and asks one question: did you keep your rule? Entries stay in this browser; nothing is sent.
Your journal, checklist and past decisions become your own data: hit rate, gains and losses, how you change your mind.
Seal your call before the AIs
One question a day, the same one the Council answers. Pick up or down and how sure you are. Your call is sealed with a SHA-256 hash in your browser, then you see the Council — and on the result day, who was closer.
Loading today's question…
Country by country: does money in a safe deposit keep its buying power?
Three doors: money until the next income, a small business, money and my work.
Three more doors beside «what it changes for me»: debt and mortgage, trips and purchases abroad, not investing yet — facts for your country with dates, how it works, steps to check for yourself.
Your money, wherever it sits — measured in the currency you will spend
Enter what you hold where it actually sits — a flat, a deposit, a local fund, government bonds bought straight from the treasury, gold — and the goal it serves. Goal Navigator is built to answer one question: how much will really be there, in the currency you will spend, on the date you need it. For government bonds: name the date and the currency you will spend — Money shows suitable bonds, the payment schedule, the amount on that date and the early-exit price. Most of the world’s savings never pass through Wall Street (Capital Flow Atlas).
One goal, two routes: what really stays in the currency you spend
A family in Brazil saves for a home down payment in 3 years. Same money, two routes: a local inflation-linked government bond, or dollars in a US Treasury and back. Every line — fees, taxes, the currency, inflation — is shown in reais of today.
| Line | BRL |
|---|
| Bond | 10,000 → in today's money in 5 years | Real a year |
|---|
| Issue passport | Yield | Duration | Buy directly |
|---|
What is really inside your funds — and does it fit your limits? →
How it works: the fields, how much is enough, the first route
Morningstar 2026 finds that individual investors value a dashboard that tracks progress toward their goals. The yardsticks people set themselves: a 7.9% average return target while 22% of portfolios sit in cash (Fidelity, 13,000 investors in 13 markets, Fidelity ↗), and $1.26mn for a comfortable retirement after a 20% rise in the target (HSBC ↗).
- Where you live, the goal and the currency you will spend it in — and the asset field by field: country of listing, domicile, revenue geography, quote currency and the currency of the assets inside. A Europe-listed fund of US stocks bought in euros is read field by field: listed in the EU, domiciled in Ireland, earning mostly in the US, quoted in euros, holding dollars — so a dual listing or one company inside two funds is never counted twice
- How much is enough — yearly spending ÷ a safe starting withdrawal rate: Morningstar’s 2026 rate is 3.9% for a 30-year US retirement (about 25.6× spending; Bengen’s 1994 rule of 4% gives 25×), set beside what cash and short government bonds pay in your spending currency today (Morningstar ↗). The rate is a dated parameter: for another country Money uses a local study where one exists, otherwise the line says “planned” rather than borrowing the US figure. Why it is looked up and never assumed: across 17 developed markets and 109 years of data, a 4% real withdrawal rate would have held in only 4 of them — even with the stock-bond mix picked in hindsight; a fixed 50/50 mix broke down at some point in all 17 (Pfau, Journal of Financial Planning, Dec 2010, FPA ↗)
- The real-return stack — nominal return → fees → currency spread → tax → inflation, with the line that eats the most highlighted
- What your tools cost: list your paid subscriptions — AI chat, data, screeners, charts — and the money they serve; Money adds them to the stack as a yearly percentage of that money, set beside fund fees, and flags two tools doing the same job, showing which research step each tool covers and where a gap is left — for example, “finding the source is covered; checking a claim against the original is still a gap.” Free — the worked example; One — your subscriptions and your money
Worked example: three subscriptions vs a robo-advisor fee
ChatGPT Plus $20 + Perplexity Pro $20 + TIKR Plus $24.95 a month is $779 a year — 3.1% of a $25,000 portfolio, about 12.5 times the $62.50 a year that a robo-advisor such as Wealthfront charges at 0.25%. US list prices, monthly billing, before local taxes; in the Navigator each price is converted into your spending currency at the ECB reference rate of the day. Prices checked Nov 16, 2026: OpenAI ↗ · Perplexity ↗ · TIKR ↗ · Wealthfront ↗.
WHAT YOUR TOOLS COST · DATED PRICES IN THE CURRENCY YOU SPEND · TWO TOOLS ON ONE JOBLoading the price book…
- Income runway: enter dated paychecks, coupons, rent and withdrawals alongside the goal; the Navigator converts each cash flow into the spending currency and checks when available cash first falls short of planned spending, including a delayed payment or a vacancy. Free — one worked route; One — your cash flows and the first shortfall date.
- What stays with me — government bonds: a date and a spending currency → for each country, the amount on that date after inflation, withholding tax and the exchange rate, and a maturity ladder built to the goal. Every issue has a passport: ISIN, coupon, payment calendar, yield to maturity, duration, exit terms, the primary document and how to buy it directly — Tesouro Direto, RBI Retail Direct, TreasuryDirect, BTP Valore, NS&I. Free — worked examples for a resident of the US, the EU, India and Brazil, and the facts of every issue; One — your data and your bonds. The country map is in the Atlas
- What stays with me — tax wrappers: the same savings kept in a UK Stocks & Shares ISA, a Japanese NISA or a US Roth / Traditional IRA against an ordinary account. Example: $10,000 a year for 10 years at 7% — the ISA keeps about 4.5% more than an ordinary account for a basic-rate taxpayer and 7.5% more for a higher-rate one, NISA about 7.4%, a Roth IRA about 4.0%. Every limit and rate links to the official page (gov.uk, fsa.go.jp, nta.go.jp, irs.gov) with the date it was checked; dividend tax rates by country come from the OECD Tax Database, Table II.4. Free: the rules table and the example. With One: enter your own country, contribution, years, return and tax rate — the calculation runs in your browser and nothing is sent. Limits and rules change by tax year; the table shows the standard case.WHAT STAYS WITH ME · TAX WRAPPERS · ONE
ISA, NISA, IRA: what stays with you
Loading the rules…
- Exit terms — early-redemption price, lock-ups and settlement: what you can actually take out on the date
- Local sources — the treasury, central bank and statistics office of the asset’s country, each with date and line
- Fast entry, next — a statement screenshot, PDF or CSV becomes the same card; the form stays as the fallback
Loading the runway…
Loading the import…
Loading the routes…
Loading the board map…
Three small calculators for your own money
A daily routine you can tick off, what really stays from a return after every cost, and what your research subscriptions cost as a share of the money they serve.
Your $100,000 against the AIs
Spread virtual $100,000 over the same markets the AIs trade: 5 to 25% each, the rest stays cash. Seal it at the last close. From the next session it is scored on the same prices as the 10+ AIs and their Council.
| # | Who | Value | Since your seal |
|---|
Humans vs machines: one table for everyone
Pick any market and your chance it closes higher at the next session. The seal is a SHA-256 hash made in your browser, the same kind the Council's seals use. Settled calls put you in the same table as the AIs, the Council and the prediction-market crowd.
Shared table
| # | Who | Settled | Side right | Brier |
|---|
Council or crowd: who wins the week?
Back the 10+ AI Council or the prediction-market crowd. The week goes to whoever has the lower Brier score on the questions settled during it. Stake $1,000 of play money, even odds.
Do the AIs make readers sharper?
Every reader gets a group by a published rule. Group A sees the Council before its first call of the day; group B only after. Both are scored on the same outcomes as the AIs. If seeing the AIs helps, group A will be right more often.
| Group | Calls | Right | Brier | Toward the target |
|---|
Learn to read a forecast by playing with one
Learn to read a forecast by playing with one
Five short games. Everything runs in your browser; the money is virtual. The closed questions are real: the Council's chance, the crowd's price and the outcome, from the open data.
Understand forecasts: five games
Prompter: change a belief with a fact
Hash duel: did anyone change a sealed forecast?
Battle of the titans: your rule against closed questions
Confidence poker
Bet against the Council
Get each issue the moment it is sealed
Every weekday before the US open: 10+ AIs answer, the answers are sealed, the issue comes out. Pick how it reaches you.
Webhook for developers
https://polyhelper-trends-votes.vadymchernets.workers.dev/hookPut the live score in your article
Copy one line of code. The widget shows the AIs against the market and updates with every issue — free, no sign-up, in all site languages.
One system. Four ways in
Independent advisors and small research teams — the people who check filings and broker notes every week. Become a partner →
See the market clearly. Stay disciplined
Get the Pulse, see what news means for your goal and ask the council in plain English. Already a weekly AI habit for 41% of them (Finimize Q3 2026 ↗); Money brings 10+ models instead of one.
A sourced answer for clients, fast
Turn a sourced council brief and its disagreement map into a client-ready email: what the models see and what it means for the plan.
A second opinion, backed by data
Monitor your own topics 24/7, compare council odds with prediction markets and, under contract, feed the forecast archive of 10+ models into your own models via API.
See which AI gets the call right
A public ranking shows which AI best predicts rates, metals and earnings. The table gives researchers a citable, dated record.
What each type of investor is really after
- Individual investor: financial freedom, savings that hold up, predictable income and less anxiety. Retirement is the leading reason to invest (Fidelity, 13 markets; Amundi, ~18,000 people in 26 countries — many doubt their savings will be enough). The mood is still upbeat: ~70% expect markets to rise over 12 months and 58% favor the US (Finimize Q3 2026, 2,808 people).
- Active trader: a repeatable edge that survives fees and mistakes.
- Professional manager: deliver the client’s mandate, a good result for the risk taken, and keep the assets.
- Long-term institution: meet future liabilities, such as pension payments — and keep the portfolio resilient. In Schroders’ 2026 survey of 1,000+ institutions, ~69% named Middle East conflict as their top concern, half feared a downturn and almost half were adding non-US assets.
The shared goal: capital that works — and an owner who understands what it is doing. Money is built for the second half of that sentence.
- Income they can count on — 90% prioritize reliable income (Global Atlantic) → Goal Navigator.
- Understand risk before a trade — growth without devastating drawdowns (Schroders, Morningstar) → Portfolio Guard.
- “Bloomberg in your pocket” at a monthly price, and a 24/7 “board of directors” that speaks up only when something material changes → Thesis Monitor, Smart Alerts.
- Spot “the next Nvidia” early → Sector Researcher.
- Keep positions secret so they can’t be front-run → Portfolio Vault.
- A place to learn — 74% would invest more if they had one (WEF); only 18–24% understand private markets → ask the council in plain English.
- 51% of Americans say “financial FOMO” has driven a purchase or investment (Empower, CFP Board 2026) → every council answer carries a mandatory counter-case.
- 60% of Gen Z get financial news from social media; almost half say AI has influenced a financial decision (Betterment 2026) → the Report Checker holds a claim against the primary filing.
- 41% of individual investors use AI for market research every week (Finimize) → the council shows how independent its models really are and keeps score in a public ledger.
That is why every council answer comes with the counter-case and the disagreement map — so a confident answer never arrives alone.
Where Money stands apart
AI research is already a large business: AlphaSense reports more than $600mn in annual revenue and a $7.5bn valuation (June 2026); Perplexity, $450–500mn. Both give one answer, not a measured panel of models. Open arenas such as 1rok and Alpha Arena let models trade and compete, but give no disagreement map, no measured independence and no investor-facing product.
Not sure where to start? Pick the job you are doing — the research route shows the path from source to watch ↓ · free
| What | Bloomberg | FactSet / LSEG | Koyfin | AlphaSense | Perplexity / ChatGPT | Kalshi / Polymarket | Money |
|---|---|---|---|---|---|---|---|
| AI views across regions | ASKB — its own agents | — | — | 1 provider | 1 company | — | 10+ models, 3 blocs |
| Disagreement map | — | — | — | — | — | price = one number | who dissents and why |
| Public forecast track record | — | — | — | — | — | the market itself | a score for every model |
| Chinese primary sources in the original | secondary translations | — | — | Western transcripts | in part | — | the original beside a checked translation; seats that search in Chinese (DeepSeek, Qwen, Kimi…) — whether they add anything is measured in the Accuracy Ledger |
| Historical analogs and outcome ranges | manual | manual | charts | — | — | — | auto + explanation |
| Open components, listed with status | closed | closed API | closed | closed | closed | public API | What’s inside ↓ |
| Keep client positions private | — | — | — | — | prompts are logged | — | Portfolio Vault |
| Price | $31,980/yr | $12–40k/yr | $39–299/mo | $10k+/yr | $20/mo | fee | from $0 / $2 / $24 |
Prices: Bloomberg — per seat, 2026 (NeuGroup ↗) · Koyfin (pricing page ↗) · Perplexity Pro and ChatGPT Plus (Perplexity ↗ · OpenAI ↗) · FactSet / LSEG and AlphaSense publish no list price — the range is an estimate until checked against a quote · Money — the plans below.
Research route: pick a job, get the path from source to watch · free
Start with the decision you need to make
Choose a job. Money shows the evidence to gather, the check to run and what to watch next, on any market and any kind of investment. Free — every route and its steps; One — your thesis, holdings and alerts (planned); Pro — your reports and strategy (planned).
- Find the original filing, fund document or issuer disclosure for the asset in its own market. For US filings, start at SEC EDGAR ↗; for other markets, the local filing systems are listed under Other markets in the Report Checker; for a fund, bond or private holding — the issuer’s documents and dated cash flows.
- Use the Sector Researcher (Pro · planned) to trace the claim through customers, costs and competitors; put the decisive claim to the council.
- Record the condition that would change your view. The Thesis Monitor then watches for that condition.
- Open the issuer’s original report and note its publication date. TIKR ↗ also indexes earnings-call transcripts across global companies.
- Send a specific claim and document to the Report Checker (Pro · planned). It matches the number, period and wording to the source.
- Keep the dated verdict and ask the Thesis Monitor to flag a material change (One).
- Write the entry, exit, market, costs and stop rule before seeing the result — plain words are enough: the rule compiler of the Strategy Backtest turns them into a tree of conditions and downloadable code (planned). For a quick look on the chart, TradingView ↗ runs scripts and alerts.
- Use the History Engine to replay only information available on each date, then the Strategy Backtest to compare with a dated benchmark after costs (Pro · planned).
- Keep the result and stop condition in the decision journal; the Decision Mirror later checks you kept to them (One).
- Start with your holdings, currencies and goal — a broker or bank statement (CSV or PDF) is enough. IBKR clients can also ask about their account in plain language through Ask IBKR ↗ (IBKR, Oct 15, 2025).
- The Portfolio Guard looks through funds and concentration; the Goal Navigator checks the path to your goal (One).
- Use Smart Alerts to hear about a change that touches your thesis or risk limit (One): browser notifications, and the daily issue push to bring you back.
Proven tools, the jobs they do, the edge that compounds — and where AI already works in finance
- TradingView ↗ — charts, alerts, scripts; #1 in the world in the Investing category (Similarweb, October 2026), with ~100mn users; $12.95–199.95/mo. billed annually, real-time exchange data often extra. Investing.com is #2, then Rakuten Securities, Fidelity and Moneycontrol.
- Yahoo Finance — the biggest free ecosystem, billions of monthly visits; CNBC, Reuters, WSJ, FT, Seeking Alpha and Finviz remain daily stops; Moneycontrol leads in India, Eastmoney in China. Since May 2026 its AlphaSpace has been pitched as a Bloomberg-style research canvas for everyone else (Fortune, May 19, 2026 ↗).
- Terminals — Bloomberg (325–350k subscribers), LSEG Workspace, FactSet ($2.44bn trailing-12-month revenue), AlphaSense with its SuperAnalyst agent.
- Koyfin ↗ — screeners, data, portfolios, reports; $39–299/mo. Alongside: TIKR, YCharts, Finviz Elite, Benzinga Pro, InvestingPro, Seeking Alpha Premium, Morningstar; Trade Ideas and TrendSpider for scanners.
- Crowd signals — Reddit and X; the ECB built R-RISI, an official retail-sentiment index from subreddits (Nov 12, 2026); Kalshi and Polymarket prices as a gauge of expectations.
- Bloomberg ASKB ↗ — agents inside the terminal, with sources.
- OpenBB / TradingAgents ↗ / 1rok / Alpha Arena — open data, agent debates, model arenas; not a unified system. Money takes patterns from them, not code — see What’s inside.
- AlphaSense raised $350 million at a $7.5 billion valuation (3 June 2026) with annual recurring revenue above $600 million; Rogo closed a $160 million Series D at about $2 billion (April 2026); Hebbia raised $130 million at about a $700 million valuation (2024). Company and press figures (AlphaSense ↗).
- AlphaAgents (BlackRock researchers, arXiv, 15 August 2025): role-based AI agents — fundamental, sentiment and valuation — debate stock picks and are tested against benchmarks at different risk tolerances. A pattern of process, not a product (Zhao et al. ↗).
- TradingAgents (open source): every extra bull-and-bear round adds 2 model calls and every extra risk round adds 3 (three risk analysts), read from the project's own code; the dollar cost depends on the model and the length of the answers (TradingAgents source ↗).
Models are available to everyone; what Money accumulates over time is the asset:
- Measured council independence — ≈ 1.71 positions on the key verdict from 13 answering models (council), with a mandatory opponent and a visible disagreement protocol; what each model adds by market is claimed only once the Accuracy Ledger measures it.
- A traceable evidence chain — checked facts and calculations, from every number to the document version and publication time; source labels and model calibration build up with it.
- Point-in-time history — a past council sees only what was available then.
- A record of decisions — why a thesis changed and which model was more accurate.
- A specialist database with usage rights and deep knowledge of one sector before the next.
- A track record measured honestly — published, dated and impossible to edit.
- Daily workflow fit — alerts, client letters and exports where the team already works.
The bar: Bloomberg (ASKB), LSEG (AI Search, Deep Research) and FactSet already join AI to proprietary data, client documents and workflows — that is the bar. Money goes deep before it goes wide: the Thesis Monitor, one sector at a time, built to earn daily use.
| The job | Where professionals and savers go (money trackers as of Nov 15, 2026) | What Money adds |
|---|---|---|
| News and what it means | Reuters, Bloomberg, Financial Times, The Wall Street Journal | The Pulse: what 10+ models from 3 regions see in it |
| Professional data and analytics | Bloomberg Terminal, LSEG Workspace, FactSet | BYOK: bring your own licensed feed |
| Charts, watchlists, alerts | TradingView | Alerts only when a thesis moves |
| Company and portfolio analysis | Koyfin | Portfolio Guard, Goal Navigator |
| Search across filings, research and expert calls | AlphaSense | Report Checker with a verdict on every claim |
| Free primary filings of US companies | SEC EDGAR | The evidence chain, by publication time |
| Macro statistics and their vintages | FRED and ALFRED | Replay: what was known on date T |
| Net worth across accounts, property and private holdings in any display currency; an MCP connector lets ChatGPT and Claude read it | Kubera | The real-return stack and exit value for each goal — Goal Navigator |
| A strictly read-only connector for Claude, ChatGPT, Mistral, Grok and Perplexity: the tracker becomes a data source for assistants | Finary | The layer that checks what the assistant concludes and keeps score of it |
| Portfolio analysis by country, currency and fees; a €12mn round led by State Street Investment Management and Portage (May 19, 2026) | getquin | Fees as one line of the stack, next to currency spread, tax and inflation |
| Return split into price, payouts, currency and fees; a look inside funds; holdings tied to a goal | Sharesight · INDmoney | Portfolio Guard look-through, without counting the same money toward two goals |
| AI translation of foreign companies’ earnings calls, 1.5mn cumulative users by January 2026 | Toss Securities | Local filings quoted in the original, the translation checked by a second model |
| Automations (Aug 24, 2026): a user-defined study runs when a filing lands | Quartr | The Thesis Monitor reports only what touches a break condition |
| Thesis Checker among its workflow agents (May 2026) | AlphaSense | A public, dated score for every call — the words “thesis monitor” alone no longer set anyone apart |
| Over 200mn active consents by end-July 2026 on the official dashboard — 208.79mn in one cut (Let’s Money), 239.79mn for receiving institutions on Jul 31 (Belvo) | Open Finance Brasil | A national data channel for the first route, Brazil |
Kubera MCP ↗ · Finary MCP ↗ · Portage · getquin ↗ · Sharesight ↗ · INDmoney ↗ · Asia Economy · Toss ↗ · Quartr ↗ · AlphaSense ↗ · Open Finance Brasil ↗ · Open Finance Brasil dashboard ↗ · Belvo ↗
- Research first. Banks, hedge funds and private-equity firms use AI for research, drafts and ideas; people still make the investment call. Morgan Stanley’s AskResearchGPT searches and summarizes the firm’s research library; Citadel runs an internal chatbot for stock pickers.
- Toward execution. BNY and Goldman are moving agents from assistance toward execution in trading, treasury and wealth management; JPMorgan uses agents in trading systems on a ~$20bn technology budget. In BNY fund accounting, dynamic fund benchmarking during the NAV calculation cut false positives from 90% to under 20%.
- At the brokerage edge. Robinhood Agentic Accounts (May) and Robinhood Agents (September); autonomous trading at Interactive Brokers (Oct 21); agent access at Coinbase via MCP (Oct 22); an MCP server at Schwab.
- Inside the terminals. Bloomberg ASKB runs specialist agents; LSEG offers AI Search and Deep Research; FactSet builds company-research and portfolio-commentary tools. Morningstar’s MCP server brings its licensed data into Claude, ChatGPT, Copilot and Perplexity — the data provider moves inside the assistant. Federal Reserve Financial Services describes AI for payment-fraud detection.
- 97% of 100+ fund managers with ~$1tn between them are adopting generative AI; 77% use it for investment memos, 73% for company or asset analysis, 71% for ideas — and only 7% consider it core to portfolio construction (AIMA, Nov 12, 2026).
- 5% of asset managers give AI autonomous or semi-autonomous authority over recommendations or trades (Mercer, 2026).
- 44 / 42 / 41 / 35% of asset owners see AI gains in workflows, information access, data quality, and research and forecasts (Morningstar).
- 15k jobs cut at the six largest US banks in Q1 while they earned $47bn, as roles move toward oversight.
- Reading vs. arithmetic. Language models read filings, compare companies and draft research; code and conventional machine learning compute the numbers, risk models and fraud signals. Money works the same way: LLMs read and debate, code calculates.
InvestmentNews ↗ · AlphaMaven ↗ · AI News ↗ · AIMA ↗ · Mercer ↗ · Traders Magazine · AskResearchGPT ↗ · Bloomberg ASKB ↗ · Semrush ↗ · Sacra ↗ · Find My Moat ↗ · BNY · AI in fund operations ↗ · Morningstar MCP server ↗
Not “Where is the market going?” but “Has my case changed?”
Know when your investment case changes. Investors pay to save time on work they can check: Money delivers a brief in which every figure leads to its source and the moment it became public. Four focused products sit at the core; the rest of the system feeds them.
Thesis Monitor — live on the engine's daily log
Example thesis “Grid equipment wins the data-center build-out” — under pressure · review (Nov 17, 2026).
Latest: NVIDIA 8-K Item 5.02 filed Nov 12, 2026 — the Board accepted Jen-Hsun Huang’s resignation as President and CEO (SEC EDGAR ↗). US 10-year past its fiscal limit by 29 bp (fiscal-limit rule holds).
Notify me on my thesis
Loading the topic feed…
KNOW WHEN THE REASON TO INVEST CHANGES
Grid equipment wins the data-center build-out
Loading the flag log…
More: who runs the company, what the price assumes, tail-risk triggers — your own spread, curve or rate threshold, one macro condition across markets
Who runs the company: a filing that changes the CEO, the CFO or the board — 8-K Item 5.02 in the US, the ad hoc disclosure in Japan (TDnet), Taiwan (MOPS) and Brazil (CVM) — lands in the Thesis Monitor's log the day the engine reads it, filing linked and quoted in the original, as with Nvidia’s Nov 12 change in November’s context.
What the price assumes: a reverse DCF from the local filing, the exchange price and the local government-bond yield returns the growth today’s price needs — first market: Brazil (CVM filings, B3 closes, Tesouro Nacional 10-year; WEG, Petrobras, Vale below). Japan, Taiwan and Germany are built too — Toyota, Sony, TSMC and SAP from their latest annual report, the close in Tokyo, Taipei and on Xetra, and the JGB 10-year (2.92%), Taiwan’s latest 10-year auction (2.00%, Oct 27, 2026) and the Bund 10-year (3.41%): at the Nov 16 close the price needs revenue growth of about 15.5% a year for 10 years at SAP, 12.4% at TSMC, 8.0% at Sony and 4.4% at Toyota; India joins as its local filing and price are connected; for every listed link of a value chain at once, see “Already priced?” in the Sector Researcher. Free — the worked example; One — your thesis.
What the price assumes — Japan, Taiwan, Germany
Loading the engine slice…
What the price assumes — local filings, local price, local 10-year
WEG (B3: WEGE3), transformer and motor maker: at R$50.40 (Nov 16, 2026) the price implies revenue growth of ~51.5% a year for 10 years at a 19.4% operating margin, discounted at 20.2% (Brazil’s 10-year 15.3% + equity premium). Petrobras and Vale: growth earns less than the cost of capital, so the price does not read as a growth rate.
Tail-risk triggers: set a credit-spread, yield-curve or rate threshold for your thesis — on the market and in the currency the thesis is tied to → each day the dated primary series (FRED and ALFRED in the US, the central bank or regulator elsewhere) is checked against your threshold and its own past range → one alert only when your condition is crossed, with the source, the time, the affected holdings and a counter-case; a repeat of the same value sends nothing new. Free — worked rules (in the Thesis Monitor above); One — your conditions.
One macro condition across markets: up to three dated thresholds joined by AND / OR — for example inflation above your level and an energy-price shock and a wider credit spread — each read from the statistics office, central bank, debt agency or exchange of the market your thesis is tied to, as published at the time → one alert only when the whole condition holds; a part not yet published shows “waiting for data,” not “no,” and a reprint of the same release sends nothing new → one source-linked log of which part fired, which did not and which is still pending. Free — a worked rule (inflation, Brent and the credit spread together, in the Thesis Monitor above); One — your rule and alerts.
Each thesis gets a verdict — held, broken or waiting for data — with the chance that a break rule holds within 90 days. Behind each rule a point forecast (the median of the 90-day corridor) is sealed with a hash chain and an RFC 3161 time stamp before the date, then scored by CRPS against climatology.
Thesis verdict and the sealed point forecast
Loading the engine slice…
What an event does to your portfolio
Loading the engine slice…
Collect
New filings, releases, transcripts and events for the companies and sectors you follow.
Detect the change
An analyst model states what is new and which part of your thesis it touches.
Try to refute it
The council’s skeptic seat looks for counter-evidence and alternative explanations.
Recalculate
Code, not prose, checks every number against the primary source.
Report
Sources, the degree of uncertainty and what it means for the portfolio — in one screen.
Score later
Every call is stored with its date and graded once the outcome is known.
Agreement is not evidence. Several models can repeat the same wrong source, so Money ranks primary data above any model vote.
How any AI market call gets checked here: each factual claim is traced to a primary source and its publication time (SEC EDGAR filings, FRED vintages that keep figures before revisions); the council’s skeptic seat tests it for flaws and alternative explanations; a forecast is replayed using only what was known on its date. Money publishes three scores — factual accuracy, scenario calibration and out-of-sample results. Each line is tagged in the engine’s evidence packet — fact (with its primary source), estimate (a calculation) or opinion (a model or an analyst); the tag now shows on every line of the day’s evidence (Nov 17, 2026: 14 facts, 17 estimates, 13 opinions, 0 unconfirmed). Spot a wrong fact? Challenge it with your source.
Every line of the day’s evidence with its tag: fact, estimate, opinion or unconfirmed.
Check an earnings claim against the filing
Below: a check of three companies from the engine’s current snapshot; in the product you pick the company and date T. Money matches figures in its earnings release to the same period in the filed 10-Q or 10-K, then shows which claims the filing supports. The current check covers Microsoft, Nvidia and Oracle; the results below come from the engine.
Any US-listed company and any date T: the request goes into a queue the engine answers on its next run, and the answer stays in the record.
Every Report Checker answer lists its documents: document · source · publication date · time it was read
- Each claim links to the release, filing and publication time
- The compared period and XBRL figure appear beside the claim
- Confirmed, contradicts the report or not checked when no comparable figure exists
- Any article, post or note: The daily public check is live: one public claim a day, traced to the filing, the release or FRED, with its verdict below. Paste a news story, a post or a broker note → each figure and claim is pulled out and traced to its primary source — the filing, the statistics office, the central bank or the exchange, on any market → confirmed, contradicts the source, sources disagree, or no primary source behind it (a story that only repeats other stories is flagged as such), each verdict with the quote and its link; a product’s own “AI-powered” claims too (More ↓). Free and $2 — one public claim checked every day, plus a trial: one claim of yours a day for two days, answered within 1–48 hours; One and up — your texts, answered in minutes; Pro — documents in bulkCLAIM CHECK · ONE PUBLIC CLAIM A DAY · ENGINE SLICE
Claim of the day: traced to the primary source
- AI revenue check: outside demand, unit economics and repeatability of a company’s AI revenue, three sourced statuses — the verdict in the AI-capex meter, the method under More ↓
Report Checker: any company, any date T
More: AI revenue check, “AI-powered” claims, wording screen, other markets, ownership flags, private funds, guidance, redemption queues, spreadsheet models and lawsuits
- AI revenue check: choose a company and reporting date → compare customer and related-party disclosures, AI segment margin after compute and server depreciation, and renewals or remaining contracts with the previous filing → three sourced statuses: outside demand, unit economics, repeatability. The Asia column now reads home filing systems: Samsung Electronics and SK hynix from OpenDART (eight quarters, consolidated), Tencent from its HKEXnews quarterly announcement. Latest: SK hynix operating margin 43.0% → 74.5% over the last four quarters against the four before, revenue +220%; Samsung 9.5% → 46.1%; Tencent 32.6% → 32.9% (three months to June 30, 2026 against a year earlier). Who pays and contract repeatability are not in these systems’ machine-readable data and are shown as not disclosed. A change becomes a Thesis Monitor event; the verdict sits in the AI-capex meter. Free — one Lab company and the method; $2 — the fresh company table, open to all after 30 days; One — your company
- “AI-powered” claims: a product’s own marketing is checked the same way as any article — a fund, adviser, robo-adviser or app on any market that says it runs on AI is read against its prospectus, Form ADV or key information document → supported, unsupported, or not described there; a performance figure in the ad is labeled by what it is — live client accounts (a GIPS composite), a model portfolio or a hypothetical backtest — and whether costs and fees were taken out: for SEC-registered advisers these are the distinctions of the SEC Marketing Rule (SEC Marketing Rule ↗), on other markets the local document and rule apply. False claims about using AI are a fact on record: the SEC fined two advisers $400,000 in total for them (SEC, Mar 18, 2024 ↗). Access as for “Any article, post or note” aboveREPORT CHECKER · “AI-POWERED” CLAIMS · SEC 497K
“AI-powered” claims: the name against the prospectus
- Wording screen: a topic — AI spending, China exposure, tariffs, going-concern doubt — and a list: an index, a cell on the board or your watchlist → the companies whose latest annual or quarterly report changed its wording on that topic against the previous one: sentences added, dropped or softened, both quotes side by side with their dates, section and page; a topic the report does not address reads “not disclosed”; on SEC filings and on the local systems below. A change becomes a Thesis Monitor event. Free — one topic on the demo companies; $2 — the topic of the week across the S&P 500 and the board’s markets, open to all after 30 days; Pro — your topic and listREPORT CHECKER · WORDING SCREEN · 10-Q / 10-K
Wording screen: topic of the week
- Other markets: the same check on local filings — ESEF annual reports in the EU, RNS announcements ↗ in the UK, EDINET and TDnet in Japan, OpenDART in Korea, MOPS in Taiwan, CNINFO in China, NSE/BSE in India, CVM in Brazil — with IFRS and US GAAP figures never mixed, the quote in the original beside its translation and the change in management’s wording quarter to quarter; today the check runs on 14 companies in three systems — TDnet in Japan (Toyota, Sony, SoftBank Group, MUFG, Tokyo Electron, Lasertec: the earnings-summary page against the statements in the same XBRL package and the printed % change against its two figures), ESEF in the EU (ASML, Airbus, Nokia, LVMH, TotalEnergies: balance sheet, cash roll-forward, last year’s figures against last year’s report) and CVM in Brazil (WEG, Petrobras, Vale: line sums, quarter = year-to-date difference, prior-year figures): 152 checks, 148 agree, 2 differ and are shown as questions (Airbus cash roll −€5m, LVMH 2024 +€108m); the first demo is TSMC, cell A7 on the board (ESMA · ESEF ↗ · EDINET ↗ · OpenDART ↗)
- Ownership-structure flags: local filings (NSE/BSE, HKEX and CNINFO, CVM) → the share of founder stock pledged, VIE structures, related-party dealings → a flag block with the document on the company card; free — a demo on companies from the board, Pro — any company, Desk — in batches Today: JSW Steel — 12.21% of the promoter holding pledged (shareholding of Sep 30, 2026); PDD — the VIE contributed 13.1% of revenue in 2025 (20-F of Apr 29, 2026); Petrobras — 61 related-party transactions in its latest Formulário. China from its own filings too: of 9 A-share companies checked in CNINFO on Nov 17, 2026, the six largest (Moutai, Wuliangye, CATL, Ping An, Midea, BYD) have no pledge announcement in 365 days, while three smaller ones have controlling holders with 19.8%, 29.9% and 53.1% of their stake pledged; Tencent, Alibaba, Meituan, JD.com and Xiaomi all describe contractual-arrangement (VIE) structures in their latest HKEXnews annual reports — 6.8% of JD.com’s 2025 revenue and 24.1% of Xiaomi’s come through them.OWNERSHIP FLAGS · PLEDGES, VIE, RELATED PARTIES
Ownership-structure flags, each with its document
Loading the engine slice…
OWNERSHIP FLAGS · CHINA FROM LOCAL FILINGS · CNINFO, HKEXNEWSChina: pledges (CNINFO) and VIE structures (HKEXnews)
Loading the engine slice…
- Private markets and infrastructure: capital-call and distribution notices and quarterly NAV statements of private equity, private credit and infrastructure funds (PDF or scans) → commitments, fees and carried interest extracted, DPI and TVPI checked against the bank statement and the fund agreement (LPA) → hidden dilution and fees the agreement does not allow; free — a worked demo fund, Pro — your fund documents, Desk — batch checks for the team The worked demo fund: paid in $7.7m, DPI 0.61, TVPI 1.44, net IRR 10.8%; the check finds a fee above the LPA ($15,000), carried interest taken before the European waterfall allows it ($80,000) and a distribution that differs from the bank statement ($90,000).PRIVATE FUND CHECK · WORKED DEMO FUND
Private fund check: a worked demo fund
Loading the engine slice…
- Redemption queue: semi-liquid funds that promise periodic exits — non-traded BDCs and REITs and interval funds in the US, ELTIFs and open-ended property funds elsewhere → each quarter’s requests against the cap the fund documents allow, the share actually paid and the history of gates, from the fund’s own filings and tender results. Latest: in its third-quarter 2026 tender, Ares Strategic Income Fund received requests for 13.1% of its shares and kept repurchases at its 5% limit (Ares filing, SEC ↗); the next filing replaces it. Free — the latest gate events; $2 — the full queue with its history, open to all after 30 days; One — the funds you hold: the first date your money is actually available, how much that leaves you short of your goal, and a position under a gate never counted as liquid in Portfolio Guard; Pro — fund documents you upload, or many funds at once Of 23 such funds that reported a tender result in the last 400 days, 12 paid pro rata in their latest tender — Blackstone Private Credit Fund paid 49.9% of what was asked, Blue Owl Technology Income Corp. 12.8% (fund tender filings, SEC). Non-traded REITs report no tender offers, so their monthly repurchases are read from the 10-Q and 10-K: in September 2026 BREIT repurchased 15.6m shares at $14.79 — 0.4% of NAV against its 2% monthly and 5% quarterly limits — and met every request; SREIT repurchased 773,596 shares at $19.43, since Apr 29, 2026 only on death, disability and accounts under $5,000, after months of paying about 3–4% of each request under a 0.33–0.5% monthly cap (SREIT 10-Q, SEC ↗).REDEMPTION QUEUE · REQUESTS AGAINST THE CAP
Redemption queue: requests against the cap
Loading the engine slice…
NON-TRADED REITS · MONTHLY REPURCHASES AGAINST THE LIMITNon-traded REITs: monthly repurchases against the limit
Loading the engine slice…
- Guidance trail: every forward figure a company gives — revenue, margin, capex — set against what it said a quarter earlier, from its own releases and call transcripts, here or on a local filing → raised, held or cut, with both quotes and dates; the council’s seats argue the reading and the call goes to the Accuracy Ledger; free — the demo companies, Pro — any companyREPORT CHECKER · GUIDANCE TRAIL · 8-K EX. 99.1 AND CALL TRANSCRIPTS
Guidance trail: what the company guided, and what came of it
Calls are read too, from the companies’ own transcripts: Microsoft guided October–December revenue at $96–97.1 billion on Oct 28, 2026, and beat its July range for July–September ($92.3 billion against $89.85–90.95 billion); Alphabet raised its 2026 capex range from $180–190 to $195–205 billion in July and held it in October; Micron’s GAAP revenue came in above its own range five quarters running (microsoft.com/investor, abc.xyz/investor, SEC EDGAR).
- Financial-model audit: a spreadsheet model (.xlsx) → the formula tree parsed: numbers typed into forecast rows instead of formulas, circular references, a balance sheet that does not balance, history that differs from the company’s filed figures — SEC XBRL in the US, the local filing elsewhere → an audit sheet cell by cell; free — a sample audit, Pro — your models, Desk — the team A sample audit of a simplified Microsoft model finds all four planted errors: a number typed into a forecast row, a circular reference, a balance sheet off by $5,000m and a history cell that differs from the 10-K (net income FY2025: $110,832m in the model, $101,832m filed).FINANCIAL-MODEL AUDIT · SAMPLE
Financial-model audit: a sample, cell by cell
Loading the engine slice…
- Lawsuits and contingent liabilities: a company and date T → Item 3 “Legal Proceedings” and the contingencies note of the 10-K and 10-Q, set against the federal court dockets made public by T (CourtListener RECAP) and the regulators’ own enforcement releases (SEC, CFTC, CFPB); elsewhere, the provisions and contingent-liabilities note of the annual report (IAS 37) and the exchange’s own disclosures → each case with its first filing date, the amount the company accrued or called “reasonably possible,” and what changed since the last filing — including a move in the company’s own wording from “remote” to “reasonably possible” to “probable”; a new case, or an accrual that grows past a share of the company’s cash and free cash flow you set, becomes a Thesis Monitor event. Free — one demo company; One — your own company, without uploading documents; Pro — any company and date, with your documents; Desk — the team’s list Demo: Meta and Alphabet — named matters and every amount sentence of the latest 10-Q verbatim (Meta, Oct 22, 2026: “the court issued an order imposing a civil penalty in the amount of $25 billion”), set against the federal dockets that name them (CourtListener RECAP, last 120 days). Regulators’ own releases are read too: in the 90 days to Nov 17, 2026 — 85 enforcement releases (SEC 75, CFTC 10), 29 with a stated amount, the largest a CFTC court order for over $288 million (Oct 19, 2026); none names Meta or Alphabet. Outside the US — IAS 37 in an ESEF annual report: Eni’s other long-term provisions fell from €15,774m to €14,580m in 2025 (−7.6%), with its contingencies note quoted (SEC and CFTC feeds, CFPB enforcement actions, filings.xbrl.org).LAWSUITS · OWN WORDS VS COURT DOCKETS
Lawsuits and contingent liabilities
Loading the engine slice…
REGULATORS’ RELEASES · SEC, CFTC, CFPB · IAS 37 OUTSIDE THE USRegulators' enforcement releases and IAS 37 provisions
Loading the engine slice…
AI revenue check — Asia column, from local filings
Loading the engine slice…
Report Checker on local filings: TDnet, ESEF, CVM
Loading the engine slice…
One value chain, followed in depth
Pick a chain — power → data centers → equipment — and Money tracks it end to end, asking second-order questions: if data centers keep growing, who earns on the grid, and who pays more for electricity? Grid load, capex plans, equipment backlogs and regulation sit in one view.
A chain with a check on every arrow →
- Four layers of the chain: compute (chips, memory, networking) → physical infrastructure (power, cooling, grids) → platforms and models → applications
- The “picks and shovels”: power, copper, chips and memory, cooling, credit
- PJM and ERCOT load, IEA, SEC filings; AI capex evidence in the Lab
- A weekly sector memo with a counter-case
- Across borders: each link carries its exchange of listing, so an event lights the cells it touches on the board — chips in Taiwan, Korea, Japan and Europe — and Follow turns the chain into a thesis; chains beyond silicon too: a strait or a tariff → freight and raw materials → producers in Europe and Asia — live today: transit calls through the Strait of Hormuz over the last 30 days are 96% below a year earlier and the Red Sea route 27% below (IMF PortWatch); seven US Section 232 proclamations in 180 days (Federal Register); 16 board cells lit; deep-sea freight +36% and Brent +44% over 12 months (FRED). TSMC lights two cells, Taiwan and New York.
- Pricing power: eight quarters of price and volume, cost of goods and gross margin from filings on any exchange → how the margin responds when costs rise → “confirmed / neutral / margin squeezed,” or “not enough data,” on each link of the chain — no composite score First read on 24 listed links: confirmed at 7 (AMD, Amazon, Broadcom, GE Vernova, Micron, Palantir, Vertiv), neutral at 3, margin squeezed at 3 (Eaton, Centrus, Microsoft), not enough data at 11 — mostly filers whose quarterly cost line is not in SEC XBRL (utilities, defence primes, TSMC, SK hynix, Cameco). Where filings do not split price and volume, revenue stands for both.
- Theme → checkable chain and Already priced?: a theme turned into a sourced chain, and each listed link read for growth against its valuation (More ↓)
Loading pricing power…
Loading the border chains…
More: theme → checkable chain, already priced?, and a worked second- and third-order chain
- Theme → checkable chain: a theme and a horizon → buyers, suppliers, financing and constraints, each link tagged fact, estimate or hypothesis with its source, money traced from spend to orders, revenue and cash flow and counted once, a company with two listings treated as one (TSMC in Taipei and New York, Cameco in Toronto and New York) → a sourced chain with its break condition, ready to Follow in the Thesis Monitor; the theme’s basket is checked against the Crowding Index basket of the same theme. Today four themes: AI compute, AI power, defense, uranium. Free — the example chains; Pro — your theme (same file format); Desk — the team’s lists
- Already priced?: each listed link of the chain → revenue growth over the last four quarters from its own filings, against the median growth of comparable companies on its home market, beside its valuation multiple against its own ten-year range → “growing, not yet priced” (growth above that median, multiple below its own median), “growth already priced” (multiple in its top quartile with growth above the median), “valued above its own history” (top-quartile multiple without that growth) or “not enough data”; anything in between shows both numbers without a verdict. Each link on its own exchange, read on one scale; for one company, the reverse DCF in the Thesis Monitor. Today 24 listed links on 4 themes, multiple = price/sales against its own quarter-end values of the last ten years. Companies that file under IFRS or at home are read too: TSMC and Cameco from the annual IFRS figures of their 20-F and 40-F, SK hynix from OpenDART, each priced on its home exchange over ten years. Today: TSMC — valued above its own history (P/S 15.1× against its ten-year median 7.6×, growth 31.6% in line with the peer median 32.0%); Cameco — growth already priced (11.8× against 5.2×, growth 11.0%); SK hynix — growth already priced (3.0× against 2.0×, revenue of the last four quarters +220%). Free — the example chain; Pro — your chain
Loading the chains…
Loading growth and valuation…
Already priced? — outside the US
Loading the engine slice…
Worked example · first order · fact: US data-center demand goes from 31 to 66 GW by 2027 (Goldman Sachs, Lab). Second order · fact: grid connections and manufacturing capacity are the constraint (IEA) → order backlogs for transformers, cables and switchgear, checked in the makers’ SEC EDGAR filings. Third order · hypothesis: who ends up paying more for power — households and industry near the sites — and how voters answer: on Nov 3, 41 local data-center measures went to a vote in 10 states (Ballotpedia) → a risk to connection dates and, back along the chain, to orders. A link with no supporting document or historical precedent is tagged “hypothesis”; every link opens to its source (Thesis Monitor).
Loading this week’s memo…
The weekly client letter, drafted from evidence
The letter that eats an advisor’s Friday: Money drafts it from the theses you follow — what changed this week and why, where the 10+ models disagree, and what it means for the client’s plan. Every figure links to its source. Nothing goes out until you approve it.
- Changes by thesis, not a news digest
- The disagreement map in plain English
- Your brand, your voice, a PDF and an email version
- Portfolio attribution: a portfolio and a period → return split into market, factors, currency and security selection → a “why the portfolio moved” paragraph with each part’s contribution and sources; the free example letter shows it on a model portfolio Today, on the model portfolio in euros, Nov 2025 – Oct 2026: +8.3% against +11.0% for the benchmark — market +7.5 pp, allocation −0.05 pp, selection −2.6 pp, interaction +0.1 pp, the dollar against the euro +3.3 pp; of the −2.6 pp gap, the value factor explains −3.9 pp.
Loading the attribution…
Loading the example letter…
Attribution of your own portfolio and your PME
Loading the engine slice…
Add any cell of the country × sector board to your letter: the cell, its figures with dates and links to the primary sources (fund holdings file, price source) go in as one paragraph. Copy it as text or as HTML with links, or download .md. The free example is one model cell; every cell is in One (your own letter) and Desk (your brand). Your draft stays in this browser.
Add a board cell to your client letter
Loading the board…
21 modules. One research system
Investors pay for speed, clear research and a track record. These modules add the views of 10+ AIs and connect each decision to a personal goal.
The market, in one screen
Each hour the engine re-ranks what is moving markets now and lists what could matter next; the council’s calls come in the morning issue. The dated facts that moved most against their own history rise to the top, with how many seats agree shown beside them; outlier views are labeled with their seats.
- Asia morning, Europe open, New York open and close
- Rates, inflation, AI capex, metals, oil, EM
- Calendar, RSS, web push — live in the record (served by the live record ↗): a calendar file (.ics), RSS and web push, one line per issue — the base is Daily alerts; the embeddable “AIs vs the market” widget is in the live record; the Pulse’s own RSS feed, one line per new edition; email and Telegram · planned
- The morning issue in 2 minutes · live: three disputes of the day with the widest spread and a quote, one fact against the majority, the next event and yesterday’s result — sealed before the open (live record)
- Issue of Nov 16, 2026 · live in the record: Copper for the week splits the room: ChatGPT, Claude, Kimi, DeepSeek, Grok, Meta AI and StepFun trim it, while Qwen, MiMo, Z.ai, Amazon Nova and Gemini buy the dip. On the 10-year yield for the week, Kimi, MiMo, DeepSeek, Meta AI and Amazon Nova bet it keeps rising. Qwen and Claude bet it falls, and Claude says it will drift down from here. On fear (VIX) for the week, Kimi, DeepSeek, Grok, Meta AI and Amazon Nova bet it rises. Claude and Gemini bet it falls, and Claude expects the event-packed week to drain the nerves by Friday. The most surprising call: MiMo (Xiaomi) is the only one to sell gold today, and Amazon Nova is the only one to sell both oil types today. Kimi sells Nasdaq and chips today while Claude, MiMo and StepFun hold. issue ↗ · export ↗
- Card of the day · live in the record: the Council’s calls for Nov 13, 2026 — 8 right of 11 calls, 1 market with no call: ❌ S&P 500 ✅ Nasdaq-100 ❌ Dow Jones ✅ Chips ❌ US 10-year ✅ Dollar ✅ VIX ⬜ Gold ✅ WTI ✅ Brent ✅ Copper ✅ Bitcoin. A 1200×630 image to share; the day’s score after 16:30 ET. image ↗ · export ↗
- What rises to the top: a primary release with its date, previous value and the council’s sealed expectation → the size of the change against that series’ own history, the assets and thesis conditions it touches; reprints of one release merge into one event, and a revised release keeps its earlier version in the log; model agreement is a separate field, not proof of importance. Free — the lead fact and its source; $2 — the full hourly ranking, open to all after 30 days. Ranked now: first prints of US releases (σ against their own history), 8-K filings of the watched companies, Thesis Monitor events, link breaks and council drift; the rank adds 0.5 per asset touched and 0.75 per thesis condition.WHAT RISES TO THE TOP · THIS HOUR
Loading this hour’s Pulse…
- Who was right this week: the council’s one-session calls of the last seven days, scored by code against the close — Brier score beside the hit rate, every seat with enough calls ranked, a coin scores 0.25; one link to the scored week in Word and deed — no second feed
Loading the scored week…
More: council drift, a link between assets breaks, theme watch
- Council drift: a recurring question with the same horizon → the share of seats in favor and the median, week by week → a line, and an event on a sharp shift; free — the latest point, $2 — the history, open to all after 30 days; the same drift among public voices is Consensus drift in the Divergence Index
Loading the council’s weekly drift…
- A link between assets breaks: pairs that usually move together or apart on any market — gold and real yields, copper and Chinese equities, the dollar and emerging-market bonds, bitcoin and the Nasdaq-100 → the rolling correlation against its own history, with the date it broke → the break is restated as a claim the data can settle and sealed in the Accuracy Ledger (the council’s own step D, ↑, follows); the break that holds becomes a Portfolio Guard scenario, and one you want to trade goes to From an anomaly to a testable rule in the Backtest. Free — the day’s largest break, delayed; $2 — the full list and its history, open to all after 30 days
Loading today’s pairs…
Loading the lead facts…
Ask 10+ AIs
Any question: a company, a sector, a scenario. One answer, a map of disagreement, sources.
- Deep mode — the full council (latest answer)
- Fast — the 3 seats with the highest current weight
Every forecast, scored
Every forecast carries a date and probability, is sealed before the event and after the model’s training cutoff, and is scored against simple benchmarks once it resolves. Live since Oct 16, 2026 on 12 markets and four horizons: sealed before the open, anchored outside, scored against a coin, climatology, the crowd and the S&P 500 — the live record. The same book next takes questions on rates, metals, earnings and elections, and the calls of public voices.
- Benchmarks: “no change,” buy-and-hold, momentum and a random picker — as many calls on the same dates, drawn at random — the control every named voice in Word and deed has to beat; banks’ year-ahead targets and AI rating services join them (More)
- Word and deed: what public voices said and what their own filings show they did, on one scored record (sample card: Berkshire Hathaway ↓); free every day — the lead number on the main theme, “N mentions — M independent primary sources”
- Brier score and calibration by question type — live today per AI, per market regime and per horizon; places by the 30/10 rule with 95% intervals and tie groups (More)
- No edge claimed before the sample can show one: telling 55% right calls from a coin takes about 145 sessions ≈ 6.9 months; a verdict against the crowd needs 350 comparable questions
- Scored in the asset’s currency and in dollars, market by market — a strong record in India can’t hide a miss in China
- For property, private credit or venture, the observable outcome and its date are fixed in advance; no data means “unresolved,” never a win
- Government bonds of the world: 10-year yields in 20 countries, auctions, ratings — scored country by country, free Sovereign rating actions enter the Ledger as dated facts linked to the agency's own release: 23 Fitch actions from Oct 16 to Nov 16, 2026 across 22 countries, six of them changing a rating or outlook — Cyprus up to ‘A’, the Dominican Republic up to ‘BB’, Cameroon down to ‘B-’, France’s outlook to negative at ‘A+’, Georgia’s to positive, Paraguay on Rating Watch Negative. Each of the 20 countries on the bond map gets the question “will Fitch change its rating or outlook within 90 days?”, resolved by the release.
- Can’t be erased or edited after the fact: the SHA-256 chain is anchored in Bitcoin (OpenTimestamps), RFC 3161 time-stamping authorities and Sigstore Rekor; a correction is a new dated line, never a rewrite; a late answer counts from its own first price; a crowd contract under $1,000 traded is labelled thin and left out of the AI-vs-crowd count
More: word and deed with a sample card, benchmarks in full, how pundits and analysts scored, government bonds and long-range forecasts, reverse source lookup, the research quality scoreboard, outside AI scores
- Word and deed — what public voices said vs what they did: choose a public voice, a market or an investment theme → a sourced portrait: what the voice said, what its own disclosures show it held or did, when each fact became public, what changed from its earlier stance and how its measurable calls ended.
In: public calls by bank strategists, central bankers, fund managers and commentators on any market — speeches, letters to investors, interviews, posts — beside what the same firms and people disclosed doing with money (fund and insider filings) — today 13F of seven funds (Berkshire Hathaway, Pershing Square, Bridgewater, Soros, Duquesne, ARK, Renaissance), their own Form 4 trades and the Periodic Transaction Reports of House members, prediction-market prices and the council itself, in one table.
Does: each call is extracted blind — asset, direction, target, horizon, condition and a short verbatim quote with its source and date — by models that see neither the date nor what happened next; a call read two ways is checked before it is scored, and “the market will crash one day” is kept as “too vague to score.” Each day’s edition is sealed into the Ledger’s own hash chain with an RFC 3161 time stamp, so a call its author later deletes stays on file with its link and hash, labeled “deleted by the author on [date].” Code, not a model, scores each call when its horizon ends: a stock call counts only if it beat its own market’s index; Brier score and calibration sit beside the hit rate wherever a probability was stated, and every voice faces the random picker on hit rate and excess return.
Out: a card per voice — calls, disclosed actions and results at 1, 3 and 12 months, by sector and horizon; how many calls were found and how many could be scored; calls recorded live kept apart from older calls collected later; “score building” until the minimum sample set before the test is reached.
WORD AND DEED · DISCLOSED DEEDS WITH THREE DATES · DAILYLoading today’s deeds…
WORD AND DEED · CONGRESS · SENATESenate trades (eFD)
Words now come from open speeches and notes — Federal Reserve Board and ECB speeches, ARK’s research notes — read in parallel by AI of different families; a call is kept only when two families find it with the same quote, character for character, and a call found by one family goes to a skeptic of another family and then to a manual check. Older documents are read with their own date removed.
WORD AND DEED · WORDS FROM OPEN SPEECHES AND NOTES · DAILYWords of public voices: dated calls from speeches and notes
Loading today’s words…
For a long-horizon investor the card has two layers: lasting principles, each a short quote from the investor’s own letters with its year and link, and the decisions of the current cycle, each tied to a 13F or Form 4 filing with its three dates; a quarter without a trade is not read as a change of principle.
Sample card: Warren Buffett and Berkshire Hathaway — principles, current management, decisions · rebuilt daily from filings · cut-off Nov 16, 2026 for the text above · free
Use it: read the investor’s stated rule, then check each decision against the filing and the date it became public. New primary filings add dated rows to the decisions layer; a principle changes only with a new letter.
Lasting principles — Warren Buffett, in the letters’ own words
- Circle of competence: “knowing its boundaries, however, is vital” (1996 letter, letter ↗)
- Buybacks follow value: the 2013 letter ties repurchases to per-share intrinsic value (letter ↗); the written rule today — only below intrinsic value “conservatively determined,” never taking cash and T-bills below $30bn (10-Q ↗)
- Cash comes second: Berkshire “will never prefer ownership of cash-equivalent assets over the ownership of good businesses” (2024 letter, published Feb 22, 2025, letter ↗)
- Mistakes are named: “The cardinal sin is delaying the correction of mistakes”; the words “mistake” or “error” appear 16 times in the 2019–23 letters (2024 letter)
- Size has a cost: “We can’t come and go on a dime” (2024 letter)
- Earnings are reinvested: one cash dividend in sixty years — 10¢ per A share on Jan 3, 1967 (2024 letter)
THE CARD, REBUILT FROM PRIMARY FILINGS · DAILYLoading the card…
Current management — Greg Abel, CEO since Jan 1, 2026, in his first letter (2025 letter, Feb 28, 2026, letter ↗)
- Cash: to the reading that a large cash pile “signals a retreat from investing” — “It does not.” Cash and Treasury bills stood above $370bn at the end of 2025 (same letter)
- A miss named: Kraft Heinz “has been disappointing”
- Beyond Wall Street: the Japanese holdings are “comparable to our major U.S. holdings in importance.” A 13F lists only US-listed securities, so the card takes those stakes from the letter and from Japan’s large-shareholding reports (EDINET ↗)
Decisions of the current cycle, 2025–26 — each from a filing or a company release
- Succession: Greg Abel CEO from Jan 1, 2026 (board vote May 4, 2025, release ↗); on Sep 18, 2026 Warren Buffett became Chairman Emeritus and stays a director, Howard G. Buffett was elected chairman (release ↗). Decisions from 2026 are Berkshire’s under its current management
- Alphabet: a position first disclosed for Q3 2025; on Jun 1, 2026 Berkshire agreed to buy $10bn in a private placement — $5bn of Class A at $351.81 and $5bn of Class C at $348.20 — alongside Alphabet’s equity raise for AI compute, announced at $80bn (Alphabet 8-K ↗) and upsized to $84.75bn (Alphabet 8-K, upsized ↗). The 13F for Jun 30, 2026 shows 78.8M Class A and 27.2M Class C shares worth $37.8bn (13F, filed Aug 14, 2026 ↗). The June price is marked to market daily in the Ledger (planned)
- Whole companies: OxyChem — agreed Oct 1, 2025, closed Jan 2, 2026, $9.7bn announced (release ↗), about $9.4bn after post-closing adjustments (10-Q); Taylor Morrison — agreed May 31, closed Jul 24, 2026 at $72.50 a share, ≈$6.8bn equity and ≈$8.5bn enterprise value, to be unified with Clayton Properties Group (release ↗)
- Buybacks: none in 2025 (2025 annual report ↗); $14.4bn in the first nine months of 2026, $9.6bn of it in Q3 (Nov 7, 2026 release ↗)
- Cash, each figure with its own definition: above $370bn at the end of 2025 (2025 letter); $352.3bn on Sep 30, 2026 for the insurance and other businesses, net of unsettled T-bill purchases (10-Q ↗). Figures with different definitions are never compared
Money’s assessment: compare the stated preference with both deployed capital and retained liquidity, then watch later filings for whether the pattern persists. One acquisition folded into an existing unit is one observation, not a new model. The filings report Berkshire’s decisions without naming who made each one; the card names the filer, Berkshire.
More: study · the record after a lagging year
Since 1990 Berkshire’s market value per share trailed the S&P 500 with dividends by 10 points or more in nine years — 1990, 1996, 1999, 2003, 2009, 2015, 2019, 2020, 2023. The following year it beat the index in seven and trailed in two (2004 and 2020). Sample: nine years out of 36. 1965–2025: 19.7% a year against 10.5% (Berkshire’s own table, 2025 letter).
Next on this card: the 13F for Sep 30, 2026 (due mid-November; not yet on EDGAR at cut-off) adds its rows the day it is filed; every 10-Q, 10-K, letter, Form 4 and company release adds a dated row; news without a primary document does not. Cut-off Nov 16, 2026 · request a correction · note 14. Free.
Every day: what changed and how unusual it is for this voice; every week: who was right. A “said vs did” line appears when a voice’s public view and its disclosed positions point opposite ways, with the possible reasons beside it: price moves, mandate limits, positions not visible. A consensus counts independent primary sources, not mentions: ten articles retelling one bank note are one voice. People and funds named here are quoted briefly, with source and date; their numbers come from primary filings (SEC EDGAR 13F and Form 4) and their own shareholder letters; every reading of them is signed “Money’s assessment”; each card shows its cut-off date and a link to request a correction (see note 14).
Dated forecasts from outside: who said what, and when
Forecasts by banks and investment houses that the archive holds with a dated source — each with its horizon and whether it can be scored yet. “Not yet measurable” is a result in its own right: the date is fixed, the check comes when the horizon ends.
Loading…
Method, open and versioned: one open method file says what counts as a call — a checkable statement about the future move of an asset, a rate or an event; views on the past, general remarks, irony, questions, ads, someone else’s call retold and conditional scenarios without the author’s own position do not count — how each call is extracted, how code scores it and how it is weighted. Every change is a new version with its date, and each score names the version of the rules it was computed under. A correction is a new record linked to the old one, never an edit. Version 0.1 is a draft; the engine runs it once the prototype is merged (planned) — method v0.1 ↗. Free.
Free — every voice’s record and score, the method, and the day’s lead number on the main theme: “N mentions — M independent primary sources”; $2 — the fresh daily slice in full, with filters, open to all after 30 days; One — an alert when a voice you follow changes its call, and the voices who talk about or hold what’s in your portfolio
VOICES YOU FOLLOW · WHAT CHANGED SINCE YOUR LAST VISIT (KEPT IN THIS BROWSER)Loading the voices…
- Benchmarks in full: the year-ahead targets banks and asset managers publish each November for indexes, rates, currencies and metals on any market are recorded with their date and link and scored by the same rule as the council; the public scores of AI stock-rating services join them the same way (“Outside AI scores” below); the random picker draws as many calls on the same dates and assets at random. Today: five S&P 500 targets for 2027 (Jefferies 9,000, UBS 8,900, Barclays 8,800, Goldman Sachs 8,700, Morgan Stanley 8,300 for mid-2027), each checked on its source page. Free — every benchmark’s score
- How pundits scored, 2005–2012 · study: CXO Advisory graded 6,582 market forecasts by 68 pundits collected in 2005–2012; accuracy was 46.9% across all forecasts and 47.4% averaged by pundit (final report, grades no longer updated; CXO Advisory ↗). Word and deed adds what that archive could not have: calls sealed at the moment they were made, scored against the voice’s own market and horizon. Free
- How analysts’ price targets fared · study: Exent replayed 988,127 ratings by 1,191 Wall Street analysts over a seven-year lookback; the average analyst’s price target was reached within 12 months 53.0% of the time (Exent’s own study, statistics as of Nov 16, 2026; Exent ↗); the next published figure replaces it. The Ledger scores bank and analyst targets forward from the day they are published. Free
- Government bonds of the world: Every day the reference forecast of the 10-year yield is sealed for the 17 markets with an open monthly series; council calls join it as the council seals them. The council seals calls on the 10-year yield in 20 countries, on auctions and on rating changes, and the Ledger scores them country by country. Published long-range forecasts will be sealed the same way, starting with Citi’s “Tokenization 2030” (Jun 1, 2026; base case $5.5tn of tokenized assets by 2030, Citi GPS ↗), with the share reached so far updated monthly. Free — the score
- Reverse source lookup: a URL, a filing number or a FRED series → every conclusion, forecast and thesis built on it; when the source is revised, a “source corrected” flag and a note to the thesis owners (free — public conclusions, One — your theses) Today 700 sources sit behind 310 published snapshots; the 10-year Treasury yield (FRED DGS10) feeds 220 of them.REVERSE SOURCE LOOKUP
Reverse source lookup
Loading the engine slice…
- Research quality scoreboard: over 30 and 90 days — the share of claims traced to a primary source, errors found and fixed (as in “What the fact check found”), the median time to a memo — three dated numbers with the counting method, free Counting from Nov 17, 2026: 160 claims checked, 24.8% traced to a primary-source figure, 5 of our own false alarms withdrawn before publication.RESEARCH QUALITY SCOREBOARD · 30 / 90 DAYS
Research quality scoreboard
Loading the engine slice…
- Outside AI scores — a backtest gets a live record: stock-rating services publish scores and backtests. Danelfin, for one, reports +456% for its Best Stocks strategy from Mar 29, 2017 to Sep 3, 2026 against +228% for the S&P 500 — a backtest, with an estimated 0.15% cost per trade (Danelfin, through Sep 3, 2026 ↗); the next published figure replaces it. The Ledger starts the live clock: a score published on date T, on any market the service covers, is recorded with its date and link, hashed, and scored forward against the same benchmarks as the council. Other signal services and the theme lists brokers publish, such as IBKR Investment Themes, are scored the same way, on the same dates, markets and costs; a service appears here by name from its first recorded score. A claimed result keeps its label — live accounts, model portfolio or backtest, as in “AI-powered” claims — and a backtest stays labeled “backtest” until the live record exists. Free — each service’s live score ForecastBench, the open benchmark of AI forecasters, is read daily and each release is hashed with its date: As of 16 July 2026 several systems on the ForecastBench tournament leaderboard (Cassi AI at the top, plus xAI and Google DeepMind submissions) are statistically indistinguishable from superforecasters; the organizers' own caveat is that this is consistent with parity, not outperformance (Forecasting Research Institute ↗). On market questions a forecaster can score well by copying the market’s price; that is why every Council answer carries the flag «saw the market price».OUTSIDE AI SCORES · LIVE RECORD
Outside AI scores: a backtest gets a live record
Loading the engine slice…
Find the historical analog
Find past setups like today’s: 10-year yields above 5%, a capex boom and record copper. Show the range of outcomes instead of a point forecast. The numbers come from code: statsforecast baselines (naive, ETS) set the bar, with a challenger beside them — the mean of AutoARIMA, AutoTheta and AutoETS, sealed in the Ledger and scored the same way; the council tests what carries over and what has changed. Try the replay ↓
- How alike are these cases really: the same buyer, the same price, the same rules, the same speed?
- What has changed in the world since then that would make this comparison misleading?
All 21 modules — 14 more
For funds and fintechs
The forecast ledger of 10+ models captures what frontier AIs believed at each point in time across thousands of questions. Access the live feed via API and license the archive.
model, question (text, resolve date, rule, source), p or value, sealed_at, hash, prev_hash and the outside timestamp — the same fields as every entry in the open Accuracy Ledger. Open today · live: council-book.jsonl ↗, council-forecasts.csv ↗, arena-protocol.json ↗, OpenAPI ↗, llms.txt ↗; the data release on Zenodo (doi:10.5281/zenodo.23552700 ↗) and 30 dated predictions registered on OSF (doi:10.17605/OSF.IO/9Q86N ↗), with BibTeX and CITATION.cff on the method page ↗. Outside models can apply to the sealed ranking with the same JSON. Read-only MCP server: a request from Claude, ChatGPT or Cursor → the Accuracy Ledger, the day’s edition, verdicts and your own theses → an answer with its date and primary sources; nothing is written and no trade is placed through it. It runs on your computer: uvx --from https://polyhelper.ai/money/data/snap/mcp/markets_mcp-0.1.0-py3-none-any.whl markets-mcp, or one line in Claude Code — claude mcp add markets -- uvx --from <that link> markets-mcp. Six tools: the list of open slices, any slice with its sources and sha256 check, the Ledger, the day’s edition, verdicts and your theses, read from a file you export from the Thesis Monitor; nothing leaves your machine. It answers Claude, Cursor and any client that speaks MCP over stdio. Free — the open Ledger; $2 — the day’s edition; One and Pro — your theses; the raw feed — by contract.Read-only MCP server
Loading the engine slice…
12 markets. 4 horizons. Before the bell
Before the New York open, the Council’s AIs seal forecasts for 12 markets (S&P 500, Nasdaq-100, Dow, chips, US 10-year, dollar, VIX, gold, WTI, Brent, copper, bitcoin — the live record) over a day, week, month and half-year. Each call includes a range, probability and condition that would invalidate it. Results are tallied that evening; recent accuracy sets council weight.
Honest refusal: when a series can’t be forecast
On Nov 17, 2026: of 48 market-horizons, 12 forecastable, 28 a range only, 0 not forecastable, 8 outside the analogs.
Where independent views split
Independence is measured (≈ 1.71 positions, council). The index maps where those positions split — copper, supply chains, tariffs, oil, rates and the dollar — from each seat’s own words in the sealed council book, and once those calls resolve, the Accuracy Ledger will show, market by market, which side was closer. The free number of the day, live on 12 markets: herd index 81% on Nov 16, 2026 — how alike the AIs’ probabilities are; independence by horizon is in the Council ↑.
Where the council splits, by theme
Loading the engine slice…
More: consensus drift
Consensus drift: In: dated calls on one question — a central-bank decision, an index target, gold or oil, on any market — from the public voices in Word and deed and from the council · Does: follows the median call and the spread between independent voices week by week; when the spread suddenly collapses and everyone stands on one side, the moment is flagged and checked against history — the Philadelphia Fed has run its Survey of Professional Forecasters since 1968 and publishes how widely the forecasters disagree (Philadelphia Fed ↗); its individual forecasts are published under ID numbers, not names, so they serve to test the spread and the scoring method, not to rank anyone by name (SPF individual forecasts ↗) · Out: the drift of the median, today’s spread against its own history, and what followed such moments, scored in the Ledger. The council’s own drift is Council drift in Pulse. Free — the day’s number; $2 — the history, open to all after 30 days. Tested on the SPF since 1968: after the spread collapsed, the median four quarters ahead missed by about as much as at other times (CPI inflation: 1.24 vs 1.57 points, the difference inside its 95% interval) — a narrow consensus is not a sign of being right. Today the narrow one is the 10-year yield: its spread is tighter than in nine of ten earlier surveys.
Consensus drift: the median and the spread of independent voices
Loading the engine slice…
From research to execution
In fall 2026, Robinhood, Interactive Brokers, Coinbase and Schwab opened account access to AI agents (dates in Where Money stands apart); JPMorgan went the other way on Oct 13 — no third-party agent may log in as a client or move money, read-only balance access stays (Cybernews ↗). So the agent account is one the broker opens to agents, never a client’s borrowed login. Money is built to work the same way: the council prepares a trade, rebalance or hedge, and a separate agent account executes it with position limits that live in code, outside any model — the size is the smaller of your mandate’s cap and the council’s suggestion, cut further when the council’s reasoning is less firm, and no model’s argument can raise it — a kill switch, approval for large trades and no order whose triggering fact lacks a second independent source — paper-account testing first (What’s inside). Your own brakes sit on your side: a daily loss limit and an order-rate limit you set block an order locally, before it reaches the broker, and a separate paper sandbox runs the agent on live prices with no account access until it passes the stress, out-of-sample and look-ahead checks and the execution gate of the Strategy Backtest. The paper account runs on this device on daily prices, without a broker.
Your broker: draft → you confirm → limits → journal
Loading the engine slice…
Connect your free Alpaca paper keys: they go only from your browser to Alpaca. Money drafts the dollar orders that move the account to your targets. You confirm each one by typing its code. Your per-order and daily limits, an allow-list and a stop switch hold everything else. Every step goes into a hash-chained journal on your device, which you can download. Live execution follows the same steps once the broker agreement is signed.
Can this order be filled on that exchange today?
Before the agent sends anything, the trade passes the exchange's own rules. An order that could not be filled is rejected — in live trading, in the backtest and in replay alike.
Cases from the engine's rulebook
Other markets · planned
Outside the US the same contour runs through brokers’ own APIs — Zerodha’s Kite Connect in India (Kite Connect ↗) — and holdings can be read, never moved, through Open Finance Brasil and India’s Account Aggregator network, and current-account balances through UK Open Banking, straight into Goal Navigator (planned; Open Finance Brasil ↗ · Sahamati ↗ · Open Banking UK ↗).
Red team
Enter an investment thesis, its horizon and what would change your mind. Money tests the strongest case for it and against it against dated primary sources, then returns a review brief: supported claims, unresolved questions, a measurable break condition and the decision you record. For documents, it runs as the Report Checker.
Check your idea → · Two sentences first, then the Council →
Does the deal survive a higher rate and a lost tenant?
Enter the loan and the income behind it. Debt-service cover and interest cover are recomputed under +100 and +200 basis points and with the largest tenant gone — and set against your covenant.
| Rate shock | Tenant leaves | DSCR | ICR | Covenant |
|---|
Worked example: the memo's claims against the documents
Where the edge comes from, private deals and property loans
Where the edge comes from: a thesis → three answers on record before any money moves: why this opportunity still exists, who is on the other side of the trade, and what you know that the price does not yet reflect; if the idea has a backtest, its result after market, size, value and momentum factors from the strategy’s own region → “edge named,” “this is market or factor exposure” or “not enough data.” Free — an example; One — your idea. Worked today on three example ideas: Japanese value (EWJV) earned +10.6% a year over cash from April 2019 to September 2026, yet after Japan’s market, size, value and momentum factors its alpha is −0.5% (95%: −3.4% to +2.3%) — market and factor exposure; semiconductors (SMH) keep +12.4% of alpha after US factors, but the third answer names nothing the price lacks, so no edge is named; European defence has two answers missing and 23 months of history — not enough data. The answers are checked by models of other families than their author.
Loading the edge check…
Private deals: send the deal memo, loan terms and cash flows; get back where the income comes from, what happens at another rate or occupancy, the covenants, collateral and payment waterfall, and which documents contradict each other. Property and infrastructure loans the same way: the rent roll and covenants → debt-service and interest cover recomputed in code, the loan schedule and rate scenarios in QuantLib, a central-bank rate, tenant-loss and late-grid-connection stress, private valuation multiples set beside listed peers → a memo for the investment committee (planned).
How concentrated is my portfolio?
With AI behind most of the S&P 500’s growth (Pulse), diversification can be an illusion. Money exposes hidden concentration, then stress-tests the portfolio against “−25% AI capex,” “10-year at 6%” and “oil at $120,” plus the regimes of 1973–74, 1979–82, 2000–02, 2008 and 2022. It checks for spurious correlations and leading indicators. It also looks through funds — which companies, countries and currencies are already inside, and how much two funds overlap, before you add a “new” asset — and answers “does this fit my risk and horizon, and which lines does that rest on.” It sorts every holding into core, defense, growth and cash, shows which regime each sleeve needs — falling rates, hot inflation, an AI boom, a recession — and flags a portfolio that only works in one of them; the review comes once a quarter or when a break condition moves, not on every CPI print.
In: holdings on any market and off it — a flat, a deposit, gold, a private stake, entered once in Goal Navigator — weights, fund look-through, horizon, loss limit and the currency you will spend · Does: compares your mix with sleeves you choose — broad core, defensive cash flow, growth, cash and short bonds — under the same dated shocks and regimes, without counting one company twice; fund holdings come from the funds’ own filings — N-PORT for US funds, 13F for managers — with the report date and a delay tag · Out: the largest shared exposure, each sleeve’s share of the drawdown and the first scenario that breaks your loss limit — no ready-made barbell weights; how much rests on one country and one currency, and a national index fund that is really a few companies (weights from the index provider’s factsheet); for money spent outside the dollar the shock is the local rate plus a devaluation, not the US 10-year. Worked example: under the same 13 dated regimes the example mix falls 33.5% at worst (2008) and breaks a 20% loss limit in 3 of 10 measured regimes; broad core alone breaks it in 5, core with 40% cash and short bonds in 2. In a pound, rupee or real the shock is local: the UK 2022 gilt sell-off took the mix −4.1% in pounds; India 2013 and Brazil 2015 show how a devaluation lifts the foreign part (+9.8% in rupees, +28.6% in reais) while a US-10-year view of the same months says −2.4% and −1.7%.
Loading the sleeve comparison…
Loading the statement formats…
Loading…
Loading the mandates…
What is really inside your funds — and does it fit your limits?
Six funds can be one bet. The look-through adds up every holding through every fund, then checks each of your limits: one country, one currency, one company, one sector, home bias.
Your limits
| Rule | Limit | Value | Status |
|---|
The example portfolio and national funds
Buy X for $N? Size from your own risk budget
Loading the engine slice…
Does the sizing rule hold? Backtest
US stock market since 1926: the rule holds the market only up to a monthly loss budget of 8% (95% CVaR, the rest in cash).
Ten more checks: home bias, drift from mandate and the rebalance, one yardstick for the whole portfolio, models exiting together, trusting the council, a fast crash, what a product really pays, position size, hedges falling with stocks, a bubble phase ending
Home bias: next to the one-country share comes a second number — that country’s share of world market capitalization (WFE, Atlas) — so a tilt toward home is measured, not felt. In the example, the United States is 57.7% of the portfolio through the funds against 48.8% of world market capitalisation (2025) — +8.9 pp, 1.18× its world weight; Japan and India are tilted up too (+6.3 and +3.0 pp), China down (−10.2 pp).
Loading the home-bias table…
Drift from mandate: you or your client set the rules once — weights, limits, exclusions, currency; Portfolio Guard shows which rule is broken, by how much and since when, and what brings the portfolio back inside. The way back is a rebalance plan, as robo-advisers do it: new cash and payouts go first, then the fewest trades that restore the weights, each priced after fees, spread and — when you add your tax lots — tax; the band that triggers it is set once, and the order goes to the agent account only after your approval. Worked example: with QQQ 9.0 pp over its target and IEF 8.0 pp under it, $30,600 of new money and dividends restores every weight inside a 5-point band with four purchases and no sale; with only $2,600, one sale is needed — $3,480 of QQQ, taken from the lot bought in December 2021 at a loss, so the trade realises a $309 loss instead of a gain — and all trades cost $0.78 in spread. Free — an example; One — your portfolio.
Loading the rebalance plan…
One portfolio, one yardstick: the total-portfolio approach used by large pension funds — no fixed share per asset class; every holding is set against the simple listed mix it replaces. A private credit fund, a rental flat, a startup stake or gold → the stock-and-bond blend with the same risk in the same currency → whether it earned more after fees, illiquidity and the years your money was locked up, and how the whole portfolio would look with the simple mix in its place. Beside it, a second, separate calculation: the whole portfolio on one map of shared drivers — rates, inflation, credit spreads, liquidity — across stocks, property and private credit, so one bet held twice (data-centre loans in a private credit fund and data-centre spending in your chip stocks) is added up, not counted as two diversifiers. A private holding valued once a quarter looks calmer than it is, so its risk is read from unsmoothed values, never from the flat line of its reports. Worked example: a $500-a-month gold plan since January 2015 against the 78/22 stock-and-bond blend with gold’s own monthly risk — PME 1.35, $48,099 more than the blend; a US flat bought for $250,000 in June 2012, on price alone (add your net rent) — PME 0.83, $124,771 less than its 52/48 blend, with its reported monthly risk of 0.5% read as 2.5% once the index’s smoothing is removed and 14.2 years locked. The PME is the same in dollars, euros or reais; the money gap moves with the exchange rate. Free — an example; One — your portfolio.
Loading the yardstick…
Models exit together: the holdings the Crowding Index flags on date T as the shared AI answer → a scenario where they are sold at once → your portfolio’s share in the “AI consensus” and the drawdown in that scenario. The same scenario counts what selling costs: positions large against their daily turnover slip nonlinearly, ETF spreads widen, margin calls come due, and private holdings are flagged “liquidity frozen” rather than counted as cash by your date — replayed on past liquidity shocks (May 6, 2010; Aug 24, 2015; Aug 5, 2024). Worked today: on Nov 16, 2026 the council’s one-month answers held no AI consensus — at most 7 of 14 seats on one side — so the example has 0% in it. If the answers converged on the side the example holds (80% of it is in assets the council tracks), the Aug 6–9, 2007 quant unwind — the crowded momentum factor fell 3.9%, 7.3 times its prior volatility — takes the example −12.6%; a private credit fund is counted as “liquidity frozen”, not as cash. The ECB's Financial Stability Review (November 2026) cites Situational Awareness, a US hedge fund betting on AI, as concentration meeting leverage: in July 2026 its assets under management fell from about $45bn to $10bn (ECB ↗).
Loading the joint-exit stress…
If you trusted the council · opens after three months of resolved forecasts (Feb 19, 2027 at the earliest): your portfolio and the council’s views, weighted by its record in the Accuracy Ledger → a Black–Litterman model (skfolio) → how far the weights would move; a what-if calculation, not a recommendation. The calculation already runs daily on the example; the moved weights stay hidden until the class of questions has its record.
Locked until the council’s one-month forecasts have three months of resolved record (Feb 19, 2027 at the earliest).
Fast crash: your holdings, order types (stop, stop-limit, limit), leverage, the cash you need and by what date, and any put hedge with its premium → a replay through minutes when prices broke from value, under your exchange’s own halt and price-limit rules (Market Profile) and your broker’s margin rule → which positions would be sold automatically, at what price in the worst minute, how much a margin call takes, and what a limit order or no leverage would have changed; what can really be sold before your cash deadline given each market’s hours, the date and size of the first cash shortfall in the currency you spend, and what the hedge costs after its premium. A fund held twice counts once; undisclosed borrowing terms give “not enough data,” not zero margin. Anchors: on May 6, 2010, over 20,000 trades in more than 300 securities printed 60% or more away from prices moments earlier (SEC–CFTC ↗); on Aug 24, 2015, there were 1,278 trading pauses, 1,058 of them in 327 exchange-traded products (SEC ↗). Next episode beyond New York: Tokyo, Aug 5, 2024, a yen carry-trade unwind — the Nikkei 225 fell 12.4% in one session (Nikkei ↗); the replay runs on that day too. Worked example on published daily open, low and close: on May 6, 2010 the −7% stops on the US funds fill at the day’s low (−10.1% and −11.5%) and miss the rebound to −3.3% — the example ends the day −6.7% with stops and 30% borrowing, −2.1% with limits instead of stops. On Aug 24, 2015 the iShares Core S&P 500 fund traded as low as −26.0% while the S&P 500’s low was −5.3%: a stop on the fund filled 20.7 points below what it holds. In Tokyo on Aug 5, 2024 the computed Nikkei 225 close, −12.40%, matches Nikkei Inc.’s published −12.4%; a −5%/−8% stop-limit sells at −8% while the fund closed −11.4%. No margin call at a 25% maintenance margin, the $8,000 cash need is met in all three, and the put pays nothing in any of them while costing its $600 premium.
Loading the fast-crash replay…
What a product really pays: a ticker or term sheet — leveraged or inverse ETF, buffer or option-income fund, structured note, autocallable, a tokenized Treasury or money-market fund; the EU key information document (KID) where there is one → payoff math (QuantLib), fees and a replay through the History Machine regimes → what you get in up, flat and down years, the cap, the cost of the buffer, how daily leverage decays in a choppy month, the worst historical path and whose credit stands behind a note; for a tokenized fund, the yield after fees against the bills it holds, the redemption terms and days, and who holds the assets; whether the backing held at the custodian matches the tokens issued, with the same bills never counted twice through a wrapper or mixed with stablecoins. Worked today, from the issuers’ own files: ProShares SSO (2× the S&P 500 each day) — in October 2026 the index fell 0.9%, the fund’s NAV 2.4%, the daily-reset model 2.3%; in the choppiest flat month (April 2025) the index ended −0.7% while the daily reset lost 3.2%, not 1.4%. Innovator BJAN — cap 16.46% before fees, 9% buffer, 0.79% fee: over 22 calendar years of SPY’s price return the cap would have cut growth 7 times (a median 7.8 points given up) and the buffer saved a median 6.2 points in the 5 down years. Free — one worked product; $2 — the card of any public product while fresh, open to all after 30 days; One — the products in your portfolio; Pro — your term sheet as a PDF.
Loading the product cards…
Position size: a holding or a new idea, your loss limit for one position and for the whole portfolio → the largest size that keeps the loss inside both limits in the asset’s worst History Machine regime and in a fast crash, computed in code from your limits — no model’s text can change it — beside the size you hold now; a fractional Kelly figure appears only as a ceiling, and only for a strategy with a walk-forward record in the backtest. Free — an example; One — your positions. Worked example with limits of 5% for one holding and 20% for the portfolio: the 40% in the S&P 500 fund can be at most 20.1% — bounded by 2022, when it fell 24.9% to its trough while the rest of the example lost 11.2%; the Nasdaq-100 fund 10% → 6.4%, Japan 8% → 3.3%, bitcoin 5% → 3.0%, while gold has room up to 48.0%. The S&P 500 fund’s worst opening gap in ten years was −10.4% (Mar 16, 2020): an −8% stop would have filled there. Portfolio CVaR 95% (monthly, Nov 2016–Aug 2026): 6.4%, 59% of it from the S&P 500 fund.
Loading position sizes…
Hedges fall with stocks: your holdings, with the gold, bitcoin and long bonds you keep as protection → a scenario where rising real yields push stocks and the usual hedges down at the same time, and only cash and short bills hold their value → what the portfolio loses with its hedges failing, which hedge breaks first, and how much cash is left by your deadline, in the currency you spend. On data since 1960: in US stocks’ worst 10% of months gold fell with them 34% of the time, long Treasuries 41%, bitcoin (since 2015) 86%. In 2022, as the 10-year real yield rose 2.62 pp, the example portfolio lost 17.4%, its hedges together −20.5%; bitcoin broke first (January), long Treasuries by April; cash and bills held $10,111 against an $8,000 need by June 2027. Free — an example; One — your portfolio.
Loading the hedge check…
A bubble phase ends: the Lab’s bubble test (↑) turns “the phase ended on [date]” into a stress on your holdings → what the portfolio loses if each asset in an explosive phase falls back to where the phase began. Today (bubble test of Nov 16, 2026) gold is in an explosive phase since April 2024; if it ends as badly as gold’s worst past end (−13.2% within 12 months), the example portfolio loses 1.3% on its 10% gold; bitcoin’s last phase ended on May 14, 2021 and stays history. Free — an example; One — your portfolio.
Loading the bubble-phase stress…
Route each job by measured record
Every seat the independent finance benchmark measures (Vals AI Finance Agent v2.1) starts with a weight from it; a model it has not measured starts at the lower quartile of the measured ones and claims no edge until the Accuracy Ledger scores it. As questions resolve, the Accuracy Ledger’s own scores take over — by question type and by market. Routine jobs go to the seats with the highest current weight.
On 5 May 2026 Anthropic released ten ready-to-run agent templates for financial work (pitchbooks, KYC screening, month-end close) and says Claude Opus 4.7 leads Vals AI's Finance Agent benchmark at 64.37%. A company claim (Anthropic ↗).
Signals from the physical economy
PJM and ERCOT grid load, AIS tanker tracks, port use, transformer and cable backlogs, Chinese state raw-material procurement records, Taiwan’s monthly company revenue reports, management’s tone and pauses on earnings calls, and synthetic stress scenarios feed the debate. “Node load is rising” can challenge a report of falling GPU use. Every signal carries source and usage rights.
Permits and grid queues: a dated register of local data-center measures, moratoria and connection queues — each with the decision, its effective date and source, matched to announced sites. It starts from facts already on the page (41 local measures in 10 US states on Nov 3, see Thesis Monitor); queues are not a US-only limit — the IEA names Texas, Virginia, Finland and the UK (IEA ↗). A delay in months appears only where the project’s own data supports it. The queues now come from the grid operators’ own files: PJM’s queue — 70.7 GW in 610 active generation projects, 18.3 GW of them in Virginia and 10.8 GW in Ohio; projects that reached service in the last 36 months waited a median of 79 months from request to service, by their own dates. ERCOT counts about 474 GW of large loads seeking connection as of June 30, 2026 — up from about 410 GW on March 26 — and says about 90% of them are data centers.
Grid-connection queues: PJM and ERCOT
Loading the engine slice…
Physical data: what changed on the ground, and which thesis it touches
Physical signals: monthly company revenue · as of 2026-11-15
$0 single facts; Desk stream; contract licensed data
- value, period, publication time and source for every point
- a revised figure is stored as a new version
- no display rights → the page shows only a fact link
| Company | Series | Period | Value | Unit | Published |
|---|---|---|---|---|---|
| Hon Hai | TW_REV_2317 | 2026-10 | 1 305 319 612 | thousand TWD | 2026-11-14 |
| TSMC | TW_REV_2330 | 2026-10 | 586 890 319 | thousand TWD | 2026-11-14 |
| MediaTek | TW_REV_2454 | 2026-10 | 55 025 656 | thousand TWD | 2026-11-14 |
European grid load (ENTSO-E), Korea exports, first 20 days, China copper imports in development — not connected yet: a free access key, a parser for the press PDF, a parser for script-rendered tables
- TSMC monthly revenue below the same month last year two months running — not enough history yet
- TSMC monthly revenue falls month on month two months running — not enough history yet
Sources: TWSE OpenAPI t187ap05_L (monthly revenue of listed companies, MOPS) ↗
Register of data-centre permits and queues · as of 2026-11-15
$0 single facts with source; $2 fresh full registry, open after 30 days; Desk own sites; contract licensed data
- a decision enters the slice of T only from its effective date
- every status links to the official document, checked by the engine
- Ballotpedia counts are summaries, not statuses; measure kinds are not merged into 'bans'
- delay in months only from the project's own dates
| Where | Regulator | Status | Date | Conditions | source |
|---|---|---|---|---|---|
| Ireland (national) | Commission for Regulation of Utilities (CRU) | permitted | 2025-12-12 | new data-centre connections must bring on-site or nearby dispatchable generation/storage matching their demand; Greater Dublin region constrained | source ↗ |
| Singapore — Jurong Island | Singapore Economic Development Board (EDB) and Infocomm Media Development Authority (IMDA) | provisionally allocated | 2026-08-21 | provisional allocation; facilities on Jurong Island; >50% of capacity from green energy | source ↗ |
- PJM and ERCOT large-load interconnection queues in development — queue files need parsing and matching to projects
- Ohio municipal restrictions (Venable: at least 14 municipalities) in development — each ordinance must first be linked to the municipality's own document
- Netherlands data-centre restrictions in development — only after the primary source is opened
- Excluded: 21 bans in Ohio / 25 MW — not confirmed by a primary source; not used
Sources: Commission for Regulation of Utilities (CRU) — IE-CRU-2025236 ↗ · Singapore Economic Development Board (EDB) and Infocomm Media Development Authority (IMDA) — SG-DC-CFA2-2026 ↗ · Ballotpedia — US-BALLOTPEDIA-2026-11-03 ↗ · Ballotpedia — US-BALLOTPEDIA-2026-11-06 ↗
More: site → suppliers, other markets
Site → suppliers: a change in a site’s status (approved, moratorium, queued) → a site–buyer–supplier graph built only from documented links → a Thesis Monitor event: “an N MW site is affected; transformer, cooling and grid suppliers exposed” — and a link in the Sector Researcher chain. Today: 3 sites and 9 links, each checked in its own document — Singapore’s Jurong Island allocation (Digital Realty, Equinix, Keppel Data Centres and ST Telemedia, 50 MW each; EDB); the customer site in Richland Parish, Louisiana (four new Entergy Louisiana units, Entergy 10-Q; gas transport by Energy Transfer, its 8-K); Abilene DC 1 in Texas (owners Blue Owl and Crusoe; the tenant, a hyperscaler, is not named in the filing).
Site → suppliers: only documented links
Loading the engine slice…
Korea’s exports for the first 20 days of each month (Korea Customs Service), European grid load and cross-border flows (ENTSO-E Transparency Platform), China’s copper imports (General Administration of Customs) — each a dated fact matched to a thesis condition (Korea Customs ↗ · ENTSO-E ↗). European load is read without a key from ENTSO-E data republished by Energy-Charts (Fraunhofer ISE, CC BY 4.0): in the 7 days to November 16 it ran 3.1% above a year earlier in Germany and 10.2% in France.
European grid load against a year earlier
Loading the engine slice…
Outside the US the register starts from regulators’ own decisions: on Dec 12, 2025 Ireland’s CRU required new data centers to bring generation or storage on site or nearby matching their maximum import demand and, on a six-year glide path, to cover at least 80% of annual demand from additional renewable projects generating in Ireland (CRU ↗); Singapore allocates capacity by call for application — 200 MW provisionally allocated to four operators on Aug 21, 2026 (EDB · IMDA ↗).
Every alternative-data signal shows: source · delay · check by another party · when it goes stale
Alerts only when the thesis moves
Powered by the Thesis Monitor. Get an alert when model disagreement spikes or your thesis breaks. Each alert gives the thesis, risks, opposing case and next step. Each morning brings a 90-second voice brief, delivered through Telegram, Slack, email or webhook.
Loading the alerts…
Test a rule before you use it
Choose a market, state a rule that can be checked — or describe your idea in plain words, or start from a council thesis — and set the date when you could have acted. Money turns it into a strategy; the generated code is shown and can be downloaded, so the test can be read and rerun. It runs walk-forward on point-in-time data, including fees, and is compared with a dated buy-and-hold or cash baseline (skfolio; a QuantStats-format report — CAGR, Sharpe, Sortino, drawdowns, monthly returns — to every run, downloadable as one HTML file). Export signals to trading terminals via API or webhook. Every result is reported twice — raw, and after market, size, value and momentum factors, with beta and sector weights beside it — so a strategy that simply bought high-beta chip stocks in a rising market is labeled as such (why).
Loading the backtest reports…
Skill or luck — and what a fund adds over its region’s factors
Skill or luck: trial register and walk-forward · as of 2026-11-15 · example
| What | Value |
|---|---|
| Idea | country-momentum-aqr-v1 |
| Trials in the register | 60 · effective 32.2 |
| Best of the grid (deflated Sharpe) | Sharpe 0.35 · DSR 0.09 · no |
| Out of sample | 2021-09–2026-08 · Sharpe 0.19 · max DD -35.6% |
| Stability (first / second half) | L9-s0-k3-inverse_vol · 0.71 / -0.25 |
Funds after their own region’s factors · as of 2026-11-15
Pro (strategy or fund) · $0 method
| Fund | Alpha a year after factors | CI 95% | p | Verdict |
|---|---|---|---|---|
| iShares MSCI Japan (EWJ) vs Japan factors | -0.01% | -1.56% … 1.54% | 0.99 | alpha not distinguishable from zero; market beta 0.96 [0.88; 1.03] |
| Vanguard FTSE Europe (VGK) vs Europe factors | -0.21% | -1.72% … 1.31% | 0.79 | alpha not distinguishable from zero; market beta 1.01 [0.97; 1.05] |
| iShares MSCI United Kingdom (EWU) vs UK factors | -1.46% | -2.89% … -0.04% | 0.04 | alpha distinguishable from zero; market beta 0.95 [0.91; 0.99] |
| iShares MSCI EAFE (EFA) vs EAFE factors (Europe + Pacific + Israel, no US/Canada) | -0.27% | -1.83% … 1.29% | 0.73 | alpha not distinguishable from zero; market beta 0.99 [0.95; 1.04] |
| VanEck Semiconductor (SMH) vs US factors — high-beta chips | 15.25% | 5.47% … 25.03% | 0.00 | alpha distinguishable from zero; market beta 1.26 [1.00; 1.52] |
two reports: as is (market only) and after the five factors; 'skill' only when the alpha interval after factors excludes zero
Sources: AQR Data Library — QMJ factors ↗ · Yahoo Finance chart (adjusted closes; internal calculation) ↗
More: a machine-learning bar, persistence of outperformance, when a strategy stops working, execution gate, plain words to a rule, an anomaly to a testable rule, strategies on an AI signal, following a filing, other markets
A machine-learning bar: In: the strategy’s universe, dates and costs · Does: a standard factor model — Qlib’s Alpha158 price-and-volume factors with a LightGBM tree — is retrained each year on past data only, trades the same universe with the same fees and slippage, on the universe you give it — your own file with constituents as they were on each date, delisted names included; the free run uses today’s large caps, so the model and buy-and-hold carry the same survivor tilt · Out: “beats the standard factor model after costs” or “does not,” beside the buy-and-hold baseline — a council thesis that cannot clear an off-the-shelf model is labeled as such. Qlib imports any exchange’s daily prices from plain files (Qlib ↗), so the same bar runs for Tokyo, Frankfurt, Mumbai or Shanghai stocks; a fund, bond or private investment is tested on its own dated prices or cash flows instead, without an equity factor model. Run today on four exchanges, 2019–2026, after costs: the off-the-shelf model earned 11.6% a year in New York, 18.9% in Tokyo, 8.6% in Frankfurt and 11.8% in Mumbai — below the equal-weight buy-and-hold of the same stocks in all four (22.6%, 21.9%, 14.4%, 16.7%). Of the three worked rules, two clear the New York bar on the same dates and one does not. Free — the method; Pro — your strategy.
Loading the machine-learning bar…
Persistence of outperformance: beside the result — the share of rolling 12-month windows above the benchmark and the information ratio, because an average edge can hide years of lagging; for the council’s seats it is open in the Accuracy Ledger. On the worked rules: the yield-switch rule beat the market in 56% of 109 rolling 12-month windows, the trend rule in 17%, the spread rule in 3% — while all three beat cash in 58–81%.
Loading persistence…
When a strategy stops working: the stop rule is set before launch, from the strategy’s own walk-forward record — the range its drawdown and rolling results stayed in → live or paper-account results checked against that range every day → “inside the range” or “outside since [date],” and the rule fires: cut the size or stop. Before the first run the strategy also records why it should still work, the instrument and venue where it can actually be traded and the condition that would cancel it; the result is set against a dated buy-and-hold or cash baseline after spread, fees and capacity (see the execution gate), window by window and in local regimes. The worked rules carry their stop rules now: fixed before launch from the record before it — a floor for the rolling three-month result against the market (its 5th percentile, breached at two month-ends in a row), a drawdown limit at 1.5 × the deepest one, and a CUSUM alarm on monthly results; one breach cuts the size by half, two stop it. Each rule and each firing is written with its hash to a decision journal before the result is read. Set as of Jan 3, 2023 as a worked example, the spread rule went outside its range on Apr 11, 2023; the live launch is Nov 17, 2026, checked every day against the paper book. Free — the method; Pro — your strategy.
Loading the stop corridor…
Execution gate: a proposed signal, its venue and date → the exchange calendar, trading halt and price-limit notices, available liquidity and the strategy’s own stop rule are checked before a paper order is counted → “tradable now,” “wait until [session]” or “blocked,” with the notice and the cost assumption. The backtest, the paper run and the broker order read the same target weights, so the paper result tests the strategy that was backtested rather than a rewritten version — the March 2026 FinRL-X paper names the backtest-to-paper and paper-to-live gaps as where results inflate (FinRL-X, arXiv ↗). The paper run is set against the backtest on the same dates, signal by signal — signal → paper order → fill or skip, with the reason — and the gap is split into costs, slippage, skipped orders and the calendar instead of one number.
- A backtest with zero slippage or zero fees is flagged as unrealistic.
- A signal that arrives after the close goes to the next session, never at the closing price.
- A halt or a non-trading day is a skip with its reason.
- An illiquid instrument gets no invented fill price.
- A private investment is followed through its cash flows, not an exchange order.
The same strategy is rerun with wider spreads, thinner liquidity and larger amounts to show where its edge starts to shrink — its capacity. On the worked rules the yield-switch rule’s edge over the market, +18 points over ten years at 5 bp a trade, is gone at 10 bp; it trades 23 times its size a year. A real order goes to the broker only with a separate recorded human approval.
Sample: a rule signals “buy” after Friday’s close → the gate finds Monday is an exchange holiday → the paper order waits for Tuesday’s open and fills at Tuesday’s price, not Friday’s close → the gap to the backtest is logged under “calendar,” with the exchange notice.
Loading today's signals…
Loading the paper account…
Free — the method and a worked signal → paper order → fill; Pro — your strategy; execution through a broker remains a separate contract.
From plain words to a rule: the idea you type is parsed into an explicit tree of conditions — trigger, asset, weight, rebalancing date on the exchange calendar — that you confirm before the run; the same text always yields the same downloadable code, a rule that peeks at data not yet known on the date is refused, and the downloaded script reruns offline to the same curves. Sample: “if the 10-year minus 2-year Treasury spread has stayed negative for three months, move 25% of stocks into Treasury bills on the first session of the month” → trigger: 10Y–2Y spread below zero for three months; asset: stocks → Treasury bills; weight: 25%; rebalancing: first session of the month on the exchange calendar → code that reads only the spread published by each date. Today the grammar reads four kinds of rule — a spread or yield that stayed above or below a level, a fast average against a slow one, a yield move over a month, one asset's return against another's; text outside them is answered with the list, and a rule that reads tomorrow's price, trades at the close that made its own signal or acts on an announcement without waiting a session is refused with the reason. On the sample, Nov 2016 – Nov 2026, after 5 bp per trade, on split-adjusted prices without dividends: the rule switched 3 times and grew 218% against 244% for the S&P 500 ETF and 28% for Treasury bills; without its best day, 195%. Free — the method and a worked rule; Pro — your rule.
Loading the worked rules…
From an anomaly to a testable rule: a dated anomaly — a broken link between assets from Pulse, a Lab reading or your own question — with its market, horizon and costs → the council proposes a few explicit explanations and a rule for each; you confirm the rule before the run, and every variant, failures included, goes into the History Engine’s trial register → a test sheet: the rule and its code, source dates, the result outside the training period after fees against the market and cash, the number of trials, and a condition to stop watching it. The best of many tries is never shown without the count of tries. Each day the engine takes the anomalies of the link-break watch — a break, or a shift in a pair's correlation within the last year — writes two or three explanations as rules, and runs each: trained before the last two years, checked on them after costs against the market and Treasury bills. Every variant goes into the History Engine's trial journal on first try; the best of a family is shown only as "best of N" with its Sharpe ratio deflated for N, and each family carries the condition that ends the watch. Free — the method and one worked rule; Pro — your anomaly and tests.
Loading today’s anomalies…
Strategies built on an AI signal: when a strategy reads news or filings through a language model, the test runs only on dates after that model’s training cutoff; earlier dates are shown apart, labeled “the model may remember the outcome,” and the look-ahead score of Gao, Jiang and Yan (described in History) joins each window. Training cutoffs come from each provider's own model page — Anthropic and OpenAI print them today; a model without a published cutoff gets no clean window. Example: a strategy reading filings through Claude Opus 5.5 (cutoff June 2026) is scored only from Jul 1, 2026 — under a year, so its score is still accumulating. Free — the method; Pro — your strategy. The score in Money is our own two-run memory-gap test from History: for claude-opus-5-5, the example model, 0.09 [95%: 0.05; 0.12] on Nov 17, 2026.
Follow a filing — with its real delay: In: a voice from Word and deed and what it disclosed — a fund’s quarterly holdings, insider buys, a lawmaker’s trades, a strategist’s calls · Does: the copy buys at the next session after the filing became public, not on the day of the trade, with costs · Out: the copy against the voice’s own result and the local index — the delay’s cost in one number. Every backtest result is also shown with its single best day removed, and with Sharpe ratio, annual growth (CAGR) and maximum drawdown recomputed without the top 1% and the bottom 1% of sessions; a curve that rests on a few days is labeled so. The voice’s own entry is the quarter-end close for a 13F (its trade dates are not public) and the reported price for a Form 4; the copy enters at the open of the first session after the filing’s acceptance time, both pay the same cost and exit together, against the S&P 500 ETF, price return. Free — the method and one worked voice (Berkshire Hathaway, every 13F and Form 4 buy of the last year; one House member as a second example); Pro — your selection.
Loading the copy and its delay…
Outside the US the factors come from the strategy’s own region — Kenneth French’s data library publishes them for developed, European, Japanese, Asia Pacific ex Japan and emerging markets (French data library ↗ · AMFI ↗ · IIM Ahmedabad factors ↗ · CVM open data ↗ · NEFIN ↗) — and the same test reads a fund you hold from its published unit prices, a UCITS, Indian or Brazilian fund: skill or luck, with its uncertainty. Run today on the funds’ own published unit values, after fees, against each country’s factors: in India (AMFI NAVs, IIM Ahmedabad four factors, 86 months to December 2025) the Nifty 50 index fund and two large active funds are indistinguishable from luck, and Parag Parikh Flexi Cap shows skill — +7.7% a year after factors [+2.3%; +13.2%]; in Brazil (CVM daily reports, NEFIN five factors, 50 months to August 2026) the Ibovespa fund and two large active funds are indistinguishable from luck, and one active fund lags its factors by 11.8% a year [−23.3%; −0.3%]. A bond or a private loan is tested on its own cash flows — coupon and maturity dates, defaults and recoveries, yield to maturity and spread to the local curve (gilts, Bunds, JGBs) — and a fund on its published unit value after fees and currency. Today’s run: a 10-year gilt bought at par at the end of 2021 at 1.01% has earned −2.7% a year since — a 23% drawdown by September 2022; the same Bund bought at −0.23% −3.8% a year, the JGB at 0.09% −2.2% a year. Bought in January 2016 and held to maturity, each earned its purchase yield (gilt 1.6%, Bund 0.3%, JGB 0.1%). A five-year sterling loan at gilts + 450 bp kept its 456 bp spread when paid in full; the same loan missing its coupons from June 2024 with 60% recovered a year later returned −8.6% a year — 813 bp below a gilt of the same maturity held over the same dates.
Loading fund unit values…
Loading bond cash flows…
Keep positions inside your perimeter
Funds keep positions inside their own systems. External models analyze markets and factors; a local open model (Llama, Qwen) reconciles that work with holdings inside the client’s environment. Reconciliation runs in a secure enclave (TEE: Intel TDX, AMD SEV) or the client’s data center.
- Leaves the perimeter: anonymized questions, factor betas, volatility, covariances
- Never leaves: tickers, amounts, positions, sizes, orders
- Local models: Llama or Qwen on vLLM / Ollama
When everyone asks the same AI
Millions of people can ask the same model, get the same answer and crowd into one trade. Asking AI for research is already a weekly habit for 41% of individual investors (Finimize, Q3 2026, 2,808 people); how alike the answers are and where the money then goes is what Money measures, beside the Divergence Radar. CFA Institute’s August 2026 report on the Algorithmic Market Hypothesis names the risk — prices increasingly reflect the dominant algorithmic reading of the news, and algorithmic monoculture becomes a systemic risk (CFA Institute ↗).
Crowding by market: 0DTE, surveys, investor types
Loading the engine slice…
More: leveraged money (last dated news, Apr 1, 2026: a 2x SK hynix fund, ~$1.6bn) and theme baskets; why AI crowding is measured from the answers (FSB, 2025)
Leveraged money: daily units outstanding of leveraged, inverse and single-stock funds from the issuers’ own files — ProShares (full NAV history with units, from 2007) and Direxion (one file a day, the series grows daily) — semiconductors, US tech, single stocks → money piling in or leaving over 1, 5 and 20 days, and the 5-day flow against its own last 250 days (a ≥ 2σ move is an event for the Theme watch). Last dated news (Apr 1, 2026): a Hong Kong-listed 2x fund on SK hynix drew about $1.6bn in 2026 to that date — topping global inflows, ahead of single-stock funds on Tesla and Microsoft (Bloomberg ↗); it stays until the issuer’s own file (CSOP, Hong Kong) can be read.
Loading units outstanding…
Themes: a theme is a basket, not a word — AI power, uranium, defense or tokenized bonds are defined by the holdings files of the funds built on them, so the same theme scores the same way on any exchange. Today: AI power and data centers (Global X ZAP, DTCR), AI compute (iShares SOXX), uranium (Global X URA), defense (Global X SHLD, iShares ITA), copper (Global X COPX); one company with two listings counts once; the page shows counts, order and countries of listing, not the funds’ weights.
Loading theme baskets…
Research, Oct 10, 2025 — why this is measured from the answers: the Financial Stability Board names herding through “similar AI models and data sources” as a market-correlation vulnerability, notes that direct indicators such as interactions between AI models “are not currently available” and that there is still little empirical evidence of the effect, and lists the number of firms using each widely used pre-trained model as a proxy (FSB ↗). For each theme the Crowding Index measures what can be seen directly: how alike the answers of 10+ models are on date T and whether money then moves the same way — the answers of the Money council, not the interaction of every trading AI in the market.
No. Recomputed with only the data known on each date, the index did not flag crowding before 8–15 October; its AI-answer part did not exist yet. Shown as it is. Did the Crowding Index see October 2026 coming? →
Models answering alike and money sitting in the same positions are two different things: the first is read from the answers, the second from positions and flows.
One move, explained with evidence
A stock drops 6% at 11 a.m. Money lines up what became public just before — filings, data releases, headlines, sector and factor moves, prediction-market odds — and the council ranks the explanations.
The 30-minute investor routine
Spend 10 minutes on overnight news, the Fed calendar and portfolio risk; 10 on a council memo with bull and bear cases and key levels; 10 deciding whether to act. Log the reason, then review your calls and the council’s each week. The payoff includes trades you avoid.
Decision Mirror: did you follow your own rules? → · Three small calculators for your own money →
Decision Mirror: did you follow your own rules
Research that earns its price
Proof is free. What the engine computes once for everyone is $2 while it is fresh. Work on your own data is One, Pro and Desk — each includes $2. A call moves to the open archive and the Accuracy Ledger once it resolves; editions, debates and explainers after 30 days.
- The whole Accuracy Ledger, always: hashes, outcomes, misses, the score; every Word and deed record and the day’s lead number (planned)
- A short daily Pulse, delayed
- The council poll: verdict and disagreement map
- Pre-Open Calls: hash and direction at once, everything after resolution
- 3 quick council questions a day
- Demos: one Lab asset in full, a History Engine episode, the Report Checker, the Country × Sector board
- Claim check trial: one claim from an article or post a day for two days, answered within 1–48 hours (planned)
- Government bonds of the world: the country map and the top 10, worked “what stays with me” examples, the facts of every issue
- Tax-wrapper rules table and a worked example
- The full Pulse at once, every hour, with the counter-case and the archive
- The whole council debate: every answer and how the argument went
- The council’s pre-open ranges, probabilities and invalidation conditions before resolution
- Every Lab asset and the AI-capex meter; all 63 History Engine dates
- Full Atlas tables, period comparison and series history; the full Divergence Radar list of the day; index history and the daily scale of crowding themes
- Handy Accuracy Ledger slices and charts; the fresh Word and deed slice with filters, open to all after 30 days (planned)
- Every “Why is it moving” explainer and the daily bulls-vs-bears memo
- The shared alert channel: “an index left its normal range”
- The fresh permit and grid-queue register (Physical Data Radar)
- Fresh report checks — the AI-revenue table and the wording topic of the week across the S&P 500 and the board’s markets — and the bubble-test status of every Lab asset and board market
- Government bonds of 20 countries: the full table with history, fiscal limits, holder history, the maturity wall, index events, auction surprises and flags
- Board scores, the market / sector / rest split, the trail and rotation — once the engine scores daily
- Calls open to everyone once resolved; editions, debates and explainers after 30 days
One helps you research and test your own edge from your own work and knowledge, instead of copying other people's signals.
- Everything in $2
- Personal consultant with your paid plan
- Saved theses and change alerts (Thesis Monitor); an alert when a Word and deed voice you follow changes its call
- Council 10+ without a limit: the 3 seats with the highest current weight in fast mode; the full council on request or when models conflict
- Your own asset in the Lab, your own ticker in “Why is it moving”
- Goal Navigator with your accounts read in — import a broker statement (CSV or PDF), read in your browser, Portfolio Guard with fund look-through
- What stays with me: government bonds for your date and spending currency, a maturity ladder, a fiscal-limit alert at your own threshold
- Thesis Auditor: check an idea
- Claim check on your own articles, posts and notes, answered in minutes (planned)
- Smart Alerts on your thesis, the 90-second brief, a personal decision journal with Decision Mirror
- Position size from your own risk budget (95% CVaR), the 3-month scenario value and “what if equal-weight”, computed in your browser
- The cause map on your own holdings
- Tax wrappers on your own amounts (ISA, NISA, IRA), calculated in your browser
- Board cells in your client letter
- Everything in One
- Personal consultant with your paid plan
- Report Checker on any company and date, your own documents; Sector Researcher
- Full research briefs and the full Council 10+ on every question, within the plan’s limits
- Thesis Auditor and private deals; Client Letter for one advisor
- Divergence Radar on your markets, index cuts, your own crowding theme
- Strategy and fund backtests, exports of your work, API quota
- Portfolio Guard across several portfolios
- Your own PME yardstick
- Everything in Pro
- Personal consultant with your paid plan
- Report Checker in batch for the team
- Shared thesis review and sector watchlists
- Team, roles, SSO and audit (planned); an oversight pack as of date T; Client Letter in your brand
- 24/7 monitors on your own topics, alerts to team channels
- Client mandates and drift from mandate across the team — a mandates file checked in your browser
- Physical data, licensed feeds
- BYOK: connect your own keys to paid data feeds
- Priority API and webhooks
- Board cells in your client letter
- Portfolio Vault inside the client’s perimeter
- On-premises deployment in your data center, with an SLA
- Live API feed and the point-in-time AI opinion archive
- Data export: Atlas and board series
- Broker Agent, after paper-account testing
- BYOK for your licensed data contracts
- License for funds
Investors pay for speed, clarity and control; Money sells the time saved on checking — every figure traced to its source — with research capabilities priced far higher elsewhere. Partner pricing: partners lock in their rate. Money benchmarks: the median comparably priced investing tool costs $22.50/mo. (Find My Moat, Nov 11, 2026 ↗; details below), Koyfin $39–299/mo ↗, TradingView up to $199.95/mo ↗, Bloomberg $31,980/yr per seat ($28,320 with 2+ seats) NeuGroup ↗.
What investors pay for — and what each plan delivers
- Terminals — Bloomberg $31,980 per seat a year, rising to $33,840 in January 2027; FactSet $12–40k (median contract $25,160, Vendr); S&P Capital IQ Pro $13–30k; LSEG Workspace up to ~$22k per user.
- AI research — AlphaSense: more than $600mn in annual revenue (Sacra estimates $700mn) and 7,000+ corporate clients; no list price — the median contract is $18,750 a year (Vendr, 41 purchases).
- Private markets — PitchBook and Preqin, $20–35k a year.
- Alternative data — ~$2.8bn of buy-side spending in 2025, up 17% (Neudata).
- Everyday tools — of 366 investing tools, 169 price comparably, at a median of $22.50/mo. (Morningstar Investor $249 a year, Seeking Alpha Premium $299); 82% keep a permanent free tier, and AI tools carry no consistent premium (Find My Moat, Nov 11, 2026 ↗). As commissions fell to zero, spending moved to data, reports, transcripts, document search, screening and backtesting. What your own tools cost: Goal Navigator.
| What they pay for | What they are really buying | How Money delivers it |
|---|---|---|
| Accurate, timely data | Decisions grounded in sound inputs | Evidence chain: every figure to its source and date |
| Professional research | Knowledge a search engine won’t surface | Council 10+ + Sector Researcher |
| Document search and processing | Hours of work saved | Report Checker |
| Portfolio analytics | Understanding of risk and where returns come from | Portfolio Guard |
| Alerts | Catching a material change in time | Thesis Monitor · Smart Alerts |
| Strategy testing | Weeding out weak ideas before money goes in | History Engine · Strategy Backtest |
| Integrations and reporting | Less manual work | API, exports, Client Letter |
| Goal-based guidance | Knowing whether the plan is on track | Goal Navigator |
FinTech Global ↗ · Find My Moat ↗ · Seeking Alpha / FT ↗ · Sacra ↗
Compare team plan costs
Compare Bloomberg seats ($31,980 for one, $28,320 each for 2+) with Money Desk. The calculator shows savings on plan costs, not investment returns.
What people pay for, and access by tool — from the live page ↓
Build it with us
Money works hand in hand with a select group of partners: independent advisors, research teams and boutique funds who live with recurring, time-heavy research — the weekly client letter, the sector review, the thesis check before an investment committee.
- The Thesis Monitor and Report Checker on your own coverage — your theses, your sectors and your companies’ filings; broker note audit comes next (planned).
- A direct line to the team — your workflow shapes what ships next.
- Partner pricing, locked for as long as you stay.
- Priority access to every new module — History Engine, Sector Researcher and the team Desk.
- Data partnerships welcome — licensed datasets, alternative data and distribution.
Where the numbers come from
The primary sources behind every figure on this page: official statistics, company filings and market data, each with its date and link. Every section also links its own sources.
How to cite · DOI → · Put the live score in your article → · SHA-256 · check the seal →
Full list of sources
- CNBC · market Nov 13, 2026
- FRED · S&P 500 · DJIA
- The Fed · decision of Sep 16, 2026
- Morningstar · Asset Owner Perspectives 2026
- Schroders · Global Investor Insights 2026
- Goldman Sachs · AI investment
- IEA · Powering AI
- World Bank · Commodity Outlook, Oct 2026
- Global X · Outlook 2027
- Similarweb · Investing, Oct 2026
- Bloomberg · ASKB
- BCG · Investor Survey 2026
- McKinsey · Investor Survey 2026
- AIMA · AI in the front office, Nov 12, 2026
- Gartner · AI spending 2027
- Crunchbase · venture Q3 2026
- InvestmentNews · ETF flows
- Kronos (AAAI 2026) · TSFM in finance
- SEC · EDGAR API · FRED API
- NYSE · data redistribution
- LiteLLM · routing
- QuantConnect · look-ahead bias
- Morningstar · asset owner survey, 2026
- Wells Fargo Investment Institute · commentary
- State Street · ETF flows to Oct 2026
- iShares · global ETP flows, Oct 2026
- World Gold Council · gold ETF flows
- IEA · energy investment · Energy and AI
- Federal Reserve · FEDS 2025-044 (LLM look-ahead)
- ALFRED · data vintages · SEC · search filings
- Bloomberg · ASKB agents · AlphaSense · Deep Research
- skfolio · QuantLib · OpenBB docs
- Charts: TradingView Lightweight Charts™ 5.3.0 (Apache-2.0) · replay series: FRED DGS10 · PCOPPUSDM · IR14270 · World Bank Pink Sheet
From question to evidence
Five auditable layers. Every step leaves a record.
The five-layer research cycle
Radar
Track events, reports, macro data, flows and alternative data; link each fact to assets, sectors and supply chains.
Evidence
Extract figures from filings (SEC XBRL) and data releases. Preserve the source document, publication date and exact quote or figure. Flag stale, disputed or unverified claims.
AI council
Models in the US, China and Europe work independently; an advocate, skeptic and arbiter test the case, and the case is run against history before the verdict (how).
Lab
Test historical regimes, factors and stress scenarios on point-in-time data, without future knowledge.
Workbench
Every decision keeps an audit trail. Track scenarios, triggers, alerts, watchlists, notes and decisions. Revisit each call as facts change. Oversight pack: a decision and date T → the documents, versions, calculations, manual edits and sign-offs → a PDF/JSON with hashes and links for a client, a committee or a regulator. Free — a sample pack, rebuilt daily from the Ledger; Pro — your own work; Desk — the team.
Loading the sample pack…
Sources: SEC EDGAR + FRED + licensed market data → fact graph + model council → research brief, scenarios, ledger and API.
What runs inside Money
Each component, what it powers in Money, its license and its status. “Works” appears only when the engine’s acceptance tests for that module have passed; the date of the run is shown next to it.
| Component | What it powers in Money | License | Status |
|---|---|---|---|
| TradingView Lightweight Charts™ 5.3.0 ↗ | The replay chart: date T, the future hidden until you reveal it; next, the Lab charts | Apache-2.0 · TradingView attribution | |
| FRED ↗ (Federal Reserve Bank of St. Louis) | Replay series: 10-year Treasury yield (Federal Reserve Board), copper since 1992 (IMF); the BLS gold import price index as a cross-check | FRED terms; each series keeps its owner | |
| ALFRED ↗ | What was known on date T: macro series as first published, before revisions | FRED terms | |
| World Bank Pink Sheet ↗ | Gold and copper prices in dollars since 1960 for the early episodes | CC BY 4.0 | |
| edgartools ↗ | Report Checker: a figure → the SEC filing (accession number, acceptance time) → the exact passage | MIT | |
| skfolio ↗ | Portfolio Guard: CVaR and each holding’s contribution to risk | BSD-3-Clause | |
| statsforecast ↗ | The baseline every council forecast must beat: naive, seasonal naive, ETS | Apache-2.0 | |
| Polymarket and Kalshi APIs | Brier score of the market itself; the divergence radar | API terms; data kept local | |
| Accuracy Ledger (built in-house) | Sealed forecasts: hash, independent timestamp, the model’s training cutoff, Brier only on resolved questions | Money’s own | |
| QuantLib · QuantStats-format report (built in-house) · Chronos-2 · Qlib Alpha158 features + LightGBM (features computed in-house from Qlib’s definitions) · paper account (built in-house; Alpaca paper optional) · statsmodels · DuckDB + networkx · ccxt · broker note audit (built in-house) | Bond math for the Report Checker and product payoffs for Portfolio Guard · backtest report · a challenger to the baseline · the Strategy Backtest’s machine-learning bar · paper account for the Broker Agent · skill or luck: factors, leverage, deflated Sharpe · Sector Researcher value chain: a chain and a date → reported spend, orders, revenue and cash flow, each dollar counted once · coin quotes across exchanges for the Divergence Radar’s “one asset, two prices” · Report Checker: upload a broker note → each figure and claim matched to a dated primary source → an annotated copy for analyst review | BSD · Apache-2.0 · Apache-2.0 · MIT + MIT · MIT · BSD · MIT + BSD · MIT · Money’s own | |
| AI-capex meter (built in-house on SEC companyfacts and 424B2/424B5 prospectuses) | The Lab, AI capex: capex against operating cash flow over the same four quarters, new US-dollar notes from final prospectuses, and a status from fixed thresholds | Money’s own · SEC public data | |
| Narrative check (built in-house on edgartools) | Report Checker: the company’s words in its earnings release (8-K, Ex. 99.1) against its 10-Q/10-K XBRL — confirmed, contradicts the report or not checked | Money’s own · MIT | |
| Council independence (built in-house) | Council 10+: how many independent texts and positions the seats really give — TF-IDF distances, participation ratio, verdict entropy | Money’s own | |
| Router prior (built in-house on Vals AI Finance Agent v2.1) | Router: starting seat weights from the benchmark; the Accuracy Ledger’s own Brier takes over as questions resolve | Money’s own · benchmark figures cited |
Patterns Money rebuilds from scratch — and what it does not run
- TradingAgents ↗ — the bull / skeptic / risk-officer roles and debate rounds behind each council seat.
- LiveTradeBench — a forecast sealed with a hash and a time before the outcome. PolyForm Noncommercial license, so the Accuracy Ledger is written from a clean sheet.
- freqtrade — look-ahead and warm-up checks in every backtest. GPL, so only the idea.
- vectorbt — the parameter heat map. Commons Clause, so only the idea.
- FinSight — claim statuses “confirmed / contradicted / no support”. GPL, so only the idea.
- FinRL-X ↗ — one target-weight contract from backtest to paper to broker, behind the Strategy Backtest’s execution gate. Apache-2.0, but a framework that sets the architecture, so only the pattern.
- OpenBB ↗ — every number carries its provider, time and source.
LEAN and NautilusTrader (heavy execution frameworks), Kronos (after Chronos-2), TimesFM 3 and Moirai 2 (noncommercial weights for Moirai), LangGraph, LiteLLM, vLLM and Ollama (when a client needs local models), FinGPT, FinRL, FinRobot, Hummingbot — research Money compares against, not components it runs.
LEAN ↗ · Qlib ↗ · Kronos ↗ · LiteLLM ↗ · vLLM ↗ · AI4Finance ↗
What was measured about AI in money management, each line with its date, source and kind of evidence.
See what moves markets
Markets worldwide. 10+ AI models, their independence measured on every poll. A view you can challenge, trace to sources and check against the public record.
The whole of Money on one page: every chapter, dataset, explanation and source. See everything →
Money gives analytics and probabilities, not personal investment advice; figures carry their source and date, illustrations are marked as such, and past outcomes do not guarantee future ones.
All 14 disclosures
- Money provides analytics and probabilities, not individual investment advice; for personalized recommendations in the US and EU the Goal Navigator works through a registered investment adviser (RIA) or a licensed partner — as Betterment and Wealthfront are set up.
- The top strip shows Nov 13, 2026 closing quotes from open sources. Product data updates live.
- The council screen captures the real council answer of Nov 14, 2026; the Pulse ranking refreshes every hour.
- GitHub star counts and private-company revenue are third-party estimates; ranges are shown where available.
- Historical outcomes and scenarios do not guarantee future results. Interface illustrations are not trading signals.
- The “10-year above 5%” episodes and historical regimes use rounded reference data. The product calculates them from dated series.
- The Scenario Lab uses real facts to show the format; probabilities come from the live council.
- Model, data and broker connections require a terms and licensing review at launch; every model and data connection has a fallback.
- Survey figures describe their own samples: Morningstar covers 500+ institutional asset owners in 11 countries, not all investors. State Street flows cover the ETFs in its sample; the World Gold Council figure covers ten months (Jan–Oct), not a full year; the IEA’s $3.4tn is industry spending, not fund flows.
- The 2027 capital map is a working map, not a forecast. Forecasters disagree, and rising investment in a sector does not guarantee returns on its stocks.
- The Thesis Monitor card is an illustration assembled from facts on this page, not a recommendation on grid-equipment stocks.
- Web search is not access to all data: licensed research and expert calls are not on the open web. Open-source code does not grant rights to financial data.
- No public evidence shows a general-purpose LLM consistently out-forecasting markets.
- Word and deed records what public figures said and disclosed, with a short verbatim quote, its source and date, a link and a neutral result — not a judgment of anyone’s honesty or intent. Numbers about named people and funds come from primary filings and their own letters; interpretations are labeled “Money’s assessment”; Money is not affiliated with or endorsed by anyone it names, and none of them takes part in it. Commentators who are not public figures appear under a type label (“video commentator 7”) until they choose to be named; a disclosed position may be partial or out of date.