1. Conviction
Does a holder betting far more than usual mark a better consensus?
Definition
A conviction holder has a position cost in the market of at least $1,000 and at least 3x the holder's own 75th percentile of position costs over the previous 90 days (positions opened 1-90 days earlier, each measured 24 h after its first fill; at least 5 such positions, else no verdict). Counted at the hour the plain consensus formed, on the consensus side and on the other side.
Baseline for every comparison: the plain consensus on the same markets (5+ tracked holders, 60 %+ on one side, price 0.05–0.95, first signal, held to resolution), 1 ¢ half-spread plus Polymarket's taker fee; $1 a trade.
Mean return per trade, with 95 % intervals
One marker per window: circle = A (curated wallets, Mar–Apr 2026, clean), square = B (all tracked wallets, Oct 2026, provisional, drift ranges dropped), diamond = pooled. Variants with fewer than 30 trades in a window are left out of the chart; the table below has every number.
Every variant
| Variant | Expected | Window A | Window B | Pooled A + B | IS / OOS (pooled) | vs its baseline (pooled) | Guard | Verdict |
|---|---|---|---|---|---|---|---|---|
| 1a >= 1 conviction holder on the consensus side entry filter, decided at the signal | better | −4.5 % [−8.5, −0.1] n = 1,178, hit 68 %, price 0.66 | −0.5 % [−6.6, +5.1] n = 640, hit 67 %, price 0.66 | −3.1 % [−6.6, +0.6] n = 1,818, hit 68 %, price 0.66 | IS −3.6 % (1,201) OOS +0.5 % (496) | +1.7 [−2.8, +6.0] points vs the dropped trades, p = 0.2173 | BH-IS no, OOS positive | nothing |
| 1b >= 2 conviction holders on the consensus side entry filter, decided at the signal | better | −5.4 % [−12.3, +1.7] n = 403, hit 69 %, price 0.68 | +3.3 % [−6.3, +13.1] n = 180, hit 76 %, price 0.72 | −2.7 % [−7.9, +3.1] n = 583, hit 71 %, price 0.70 | IS −4.3 % (431) OOS +6.0 % (121) | +1.8 [−3.6, +7.8] points vs the dropped trades, p = 0.2622 | BH-IS no, OOS positive | nothing |
| 1c >= 3 conviction holders on the consensus side entry filter, decided at the signal | better | −12.1 % [−23.4, −0.5] n = 125, hit 69 %, price 0.71 | +3.8 % [−10.8, +16.3] n = 72, hit 79 %, price 0.75 | −6.3 % [−15.2, +2.2] n = 197, hit 73 %, price 0.73 | IS −7.5 % (141) OOS +2.6 % (44) | −2.0 [−11.2, +7.0] points vs the dropped trades, p = 0.6638 | BH-IS no, OOS positive | nothing |
| 1d none on the consensus side, >= 1 on the other side (control) entry filter, decided at the signal | worse | +14.8 % [−13.9, +43.9] n = 62, hit 61 %, price 0.51 | +4.2 % [−21.8, +30.2] n = 53, hit 60 %, price 0.54 | +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 |
| 1e no conviction holder on either side (control) entry filter, decided at the signal | control | −3.9 % [−9.2, +1.5] n = 1,823, hit 61 %, price 0.60 | −5.9 % [−8.7, −3.2] n = 3,063, hit 64 %, price 0.64 | −5.1 % [−7.7, −2.5] n = 4,886, hit 63 %, price 0.63 | IS −2.7 % (2,310) OOS −7.6 % (2,209) | −2.8 [−6.9, +1.9] points vs the dropped trades, p = 0.8946 | no | nothing |
What this data can and cannot show
Can: Whether the tracked wallets' unusually large bets, judged against their own history, predict the consensus outcome better than the plain head count.
Cannot: Bets by wallets the site does not track, bets placed outside the order book (conversions), and a wallet's history before the site started storing its fills (leaderboard wallets of October 2026: only 20 days of history, so most have no verdict).
Caveats that apply to every signal
- Only tracked wallets' fills exist, from when the site started tracking them; the replay is closer to the chain than to what the site displayed (no Polymarket position overrides, merges missing May–Oct 2026).
- Window B's fill times run early in parts of October 2026; the affected ranges are dropped and the rest is not proven clean. Its numbers are provisional until the repaired export is in.
- Prices are the last tracked fill, not the order book (median error 0.5 ¢, p90 up to 16 ¢ on fast markets); entries happen on the hour (5-minute steps in play).
- The windows are 7 and 3 weeks long; effects of a few points cannot pass a guard over 40 variants there. A failed guard is not evidence of no effect — it is the absence of evidence.
- Nothing here is trading advice. Past replays, even clean ones, do not pay future trades.