The whole record.
Including what didn't work.
Every signal we track, every program we have registered, and the dated log of what was claimed and what came back. The raw file is at the bottom if you would rather check it yourself than take our word for it.
Graduation bar: ROI at or above 6 percent on a sample of at least 40 settled bets. Whiskers are 90 percent confidence intervals on return.
Ranked by return at the prices we actually got, losers included. Bars are the point estimate; whiskers are the 90% confidence interval. 4 clear the capital bar of ROI ≥ 6% on n ≥ 40, and 1 have an interval that clears breakeven entirely. A strong return on a small sample is a hypothesis, not a result — printing the interval is the difference between the two.
Committed in writing with a target n. The verdict is whatever the data says when it arrives.
2 supported · 13 not supported · 4 short of n
Production rules, re-graded nightly. When the forward record stops paying, the rule dies.
Registration and verdict are separate entries with separate dates. That gap is the point — anyone can write “we pre-register” in a method section. This is the receipt.
A perfectly calibrated forecast sits on the dashed diagonal. Every one of our bands sits below it, which means the model states its confidence too strongly. We publish it because the fix is measurable: isotonic recalibration moves the Brier score from 0.24314 to 0.23034 — 5.26% better across 4,913 settled positions. The forecast did not get smarter; it stopped overstating, which is the difference between a number you can size a position on and one you can't.
Everything we have failed to solve, in public.
1,070 games, walk-forward. Our error was worse than the market's in both directions and simulated spreads returned −6.2%. Well calibrated with no edge — those are different things. NBA therefore sizes nothing.
Next: lineup-derived ratings from play-by-play, after roster-level availability adjustment failed structurally.
Rules currently in production that the forward record has stopped supporting. They stay visible, with their sample and their return, until the nightly re-grade removes them.
Next: whether the decay is the class or the entry window — the rules were re-timed and the samples are still accruing.
0 of 387 hypotheses swept across the CBB, NBA, CFB and NFL corpora survived multiple-testing correction. The MLB signal book is a separate program with signals that do pay; these sweeps are the expansion effort.
Next: market-structure programs rather than outcome prediction — the 40 open registrations lean toward pricing behaviour, not who wins.
Check it yourself.
Every program we have registered — open and graded — as one file, regenerated nightly by the same script that builds this page.
Prefer the write-up? The working paper goes through the method and the results in full.