Performance

Methodology note — Aug 17, 2026. On Aug 7, Polymarket changed how 5m/15m Up/Down markets resolve: from the raw price at the window boundaries to the Chainlink TWAP-60s feed (hourly markets stayed on the Binance 1h candle; ties now resolve Up). Our grader was still applying the old rule, so roughly 7% of windows between Aug 7 and Aug 17 — the tight ones — had been graded with the wrong outcome. On Aug 17 we re-graded every night since Aug 7 against each market's actual on-chain resolution (UMA), and the collector now reads the real resolution instead of recomputing it. Every figure on this page reflects the corrected outcomes; nothing was removed or cherry-picked.

Model note — Aug 18, 2026. Since 06:18 UTC today, the live model prices the variable that actually resolves these markets: the final 60-second TWAP, not the terminal spot. Outside the last minute the horizon variance becomes σ²·(τ−40 s); inside it, the probability conditions on the part of the average already locked in. The price-to-beat is read from the Chainlink TWAP-60s stream itself. Earlier nights keep the model that actually ran that night — outcomes were corrected (note above), model predictions are never restated.

Lines: four long-running rules, plus the two controls (dashed). Both controls sit significantly below zero — CTRL_RND prices the spread, CTRL_FAV30 the favorite premium — which is the bar a live rule has to clear. Table below: every strategy.

95% CI ≈ normal (mean ± 1.96·σ/√n) over all settled trades. Controls (CTRL_*) land significantly below zero — the spread and the favorite premium, priced; a rule only counts once its interval clears the zero the controls cannot reach.

Common rules: simulated buys at the book's real ask, $1 stake, at most 1 trade per strategy per window, never in the last 20 seconds. Strategies added along the way only have history from their introduction date (see the "nights" column above).