Direction (a move smaller than the day's noise is a draw), Brier score for the probability (a coin scores 0.25), error in %, a virtual $100,000 portfolio after costs. Baselines: a coin, “no change”, buy and hold, the S&P 500, Poly World-10, Poly USA-5 vs Poly China-5.

The Council's portfolio is the mean of the AIs' weights, fixed in the league protocol before the first session; the other Council lines are yardsticks. It is the Council’s main count — the figure at the top of the page, after the league’s costs; the portfolios by actions are a separate count, each with its own S&P 500 from the same day.

For the 10-year yield and VIX the AIs bet on the indicator itself, up or down; in the portfolio a bet on a rising yield is held as the inverse of IEF and on a falling one as IEF; on a rising VIX as VIXY, on a falling one as SVXY.

“Buy on dips” is a level set in advance: if the price touches it during the session, the AI buys there; if not, nothing is bought, and the forecast still counts.

The score counts the answer sealed before the reference price. An answer sealed later gets its own reference: the first price after its seal.

The 12 indicators of one day move together, so the league counts trading days too: a model is ranked after 30 forecasts on at least 10 trading days.

The AIs against the crowd: for the indicators with a daily question on Polymarket (“Up or Down”) or Kalshi (“above the last close at 4 pm ET”), we do not give the AIs the crowd's price and ask them not to look it up (the question is published word for word); an answer whose sources cite Polymarket or Kalshi is left out of this count. Each AI answer is compared with the crowd's price at that answer's own minute; Polymarket is scored on its own resolution, Kalshi on its strikes (between them — the official close), each with its own count.

Data: CC BY 4.0