Signal explorer: eight signals nobody shows, tested
Pre-registered on 2026-10-25: definitions, variants, expected directions, baseline, windows, statistics and the guard were fixed before any result was read. Every variant is shown, whatever it says.
Each signal is an entry filter, side picker or exit on top of the plain consensus (5+ tracked holders, 60 %+ on one side, price 0.05–0.95, held to resolution), replayed point in time on the tracked wallets' fills with a 1 ¢ half-spread and Polymarket's taker fee. Returns are per $1 staked; every figure carries its n and 95 % interval; nothing is called an edge unless it passes the guard (Benjamini–Hochberg over every variant's in-sample p, then positive out of sample).
The eight signals
The "best variant" column picks the highest pooled mean after the fact — which is exactly why the guard runs over all 40 variants and not the best one. Click a signal for its definition, every variant per window, the chart and the caveats.
| Signal | Best variant (pooled, after the fact) | Trades A / B | Mean a trade, 95 % CI (pooled) | IS / OOS (pooled) | vs its baseline | Guard | Verdict |
|---|---|---|---|---|---|---|---|
| 1. Conviction Does a holder betting far more than usual mark a better consensus? | 1d: none on the consensus side, >= 1 on the other side (control) | 62 / 53 | +9.9 % [−10.4, +30.1] n = 115, hit 61 %, price 0.52 | IS +16.8 % (71) OOS +14.5 % (37) | +14.5 [−4.8, +34.2] points vs the dropped trades, p = 0.0874 | BH-IS no, OOS positive | nothing |
| 2. Speed of formation Is a consensus that formed fast more informed than one that built up over days? | 2d: slower than 24 h | 2,074 / 1,321 | −2.9 % [−6.3, +0.0] n = 3,395, hit 64 %, price 0.63 | IS −2.3 % (2,029) OOS −3.3 % (1,145) | +2.8 [−1.4, +7.4] points vs the dropped trades, p = 0.0914 | no | nothing (some variants negative) |
| 3. Buying below smart money's entry Does buying the consensus side cheaper than the holders paid beat buying it after the price moved their way? | 3d: price >= 5 points above the holders' entry (control) | 768 / 666 | +0.2 % [−2.8, +3.3] n = 1,434, hit 80 %, price 0.78 | IS +4.6 % (851) OOS −4.7 % (470) | +5.8 [+1.9, +10.0] points vs the dropped trades, p = 0.0045 | no | nothing (some variants negative) |
| 4. Elite vs crowd When the best-record wallets disagree with the broad consensus, who is right? | 4c: elite (win rate) disagrees: buy the elite side | 106 / 88 | −0.1 % [−20.7, +22.4] n = 194, hit 49 %, price 0.47 | IS −2.6 % (133) OOS −15.3 % (48) | +17.5 [−11.0, +45.6] points vs base-hold, p = 0.1084 | no | nothing |
| 5. Live score + consensus In play, is smart money on the side that is currently behind a buy? | 5b: in play, >= 5 points below pre-match (behind) | 1,620 / 2,238 | +0.7 % [−3.7, +5.1] n = 3,858, hit 48 %, price 0.45 | IS −0.6 % (2,103) OOS +2.3 % (1,667) | +15.4 [+13.3, +17.8] points vs s5e-any, p = 0.0005 | BH-IS no, OOS positive | nothing (some variants negative) |
| 6. Cross-market consistency When the winner market and the spread market of one match agree, is the consensus more often right? | 6e: spread market: the winner market disagrees | 14 / 35 | +2.2 % [−28.1, +32.9] n = 49, hit 57 %, price 0.51 | IS +10.2 % (19) OOS −4.8 % (29) | +9.5 [−23.0, +42.6] points vs the dropped trades, p = 0.2892 | no | nothing |
| 7. Smart money leaving Should a position be sold when the holders behind it leave, and should money that fell before the signal block the entry? | 7d-fell: the side's money fell >= 25 % in the last 6 h (the dropped set) | 20 / 23 | +16.5 % [−25.0, +62.1] n = 43, hit 60 %, price 0.57 | IS +20.4 % (18) OOS +24.0 % (22) | +20.9 [−20.5, +73.8] points vs the dropped trades, p = 0.1863 | no | nothing (some variants negative) |
| 8. Fresh wallets with big precise bets on quiet markets Does a new wallet putting $5k into a market that smart money barely touches know something? | 8d: no fresh wallet in the market (control) | 2,747 / 3,744 | −4.1 % [−6.3, −2.0] n = 6,491, hit 65 %, price 0.64 | IS −1.9 % (3,278) OOS −6.0 % (2,732) | +5.3 [−6.0, +15.4] points vs the dropped trades, p = 0.1873 | no | nothing (some variants negative) |
Windows
Two windows of the production export, as in every backtest report: A, the curated wallets in March–April 2026 (clean fill times), and B, all tracked wallets in October 2026, provisional until the repaired export lands (fill times run early in parts of October; the affected ranges are dropped). Pooled = A + B. A third window, the one-year leaderboard backfill, is hindsight-biased by construction and will be added as its own labelled column when assembled, never pooled.
| Window | Entries | IS / OOS split | Baseline (plain consensus, held) | IS / OOS | Variants under the guard | Data |
|---|---|---|---|---|---|---|
| Window A: curated wallets, Mar 10 - Apr 28 2026 | 2026-03-10 → 2026-04-28 | 2026-04-13 | −3.8 % [−7.4, −0.2] n = 3,063, hit 64 %, price 0.62 | IS −2.8 % (2,552) OOS −12.4 % (303) | 34 | clean |
| Window B: all tracked wallets, Oct 2 - 22 2026 (provisional) | 2026-10-02 → 2026-10-22 | 2026-10-13 | −4.9 % [−7.3, −2.3] n = 3,756, hit 65 %, price 0.64 | IS −2.3 % (1,030) OOS −5.1 % (2,439) | 30 | provisional, drift ranges dropped |
How to read it
At signal variants keep or drop the baseline's own trades at the hour the consensus formed, using only data before that hour; their comparison is kept − dropped with a paired cluster bootstrap over events. Strategy variants run natively in the replay with the baseline's gates and are compared with their baseline paired by market. IS / OOS are the pre-registered halves of each window (a trade belongs to a half when entered in it and its market ended before the half closed). The guard: BH at q = 0.05 over the in-sample p-values of every variant with 30+ in-sample trades, then a positive out-of-sample mean on 30+ trades. "Promising" means positive in both windows and both halves with 100+ trades and a comparison interval above zero — and is still not proof.
Full method and numbers: research/backtest/results/signals/REPORT.md and research/backtest/features/PREREGISTRATION.md in the repository; the figures as JSON at /api/research/signals. The broader track record is on /research.