How many seconds behind BTC's own price do Polymarket's 15-minute Up/Down trades price?
FRAME
- Subject: Polymarket "Bitcoin Up or Down" 15-minute markets; trades from Data API v2 (second timestamps); BTC from Binance Vision spot BTCUSDT 1-second klines.
- Why it matters: if trades price BTC's move tens of seconds late, a faster trader can buy the cheap side; if they keep up within seconds, the edge seen on 1-minute price points is a feed artifact.
- Scope: one market per day, random (seed 0), markets ending 2025-09-13 to 2026-08-31; 353 markets, 861,420 distinct in-window fills, 671,059 used (trade and model between 0.05 and 0.95 at every lag). Descriptive, no claim. Method locked in
docs/bets/2026-11-18-pm-15m-lag.md, commit447dc78, written before any trade or 1-second data was loaded. - Terms: model = Phi(ln(P/P_start) / (sigma sqrt(T - s))), P = 1-second close at s - 1, sigma = std of 1-second log returns over the 900 s before start. Lag L = model at trade time - L. Fill = one (transaction, price, size); fills after the window ends are dropped.
MEASURES
Mean squared gap by lag (best 3 s):
| lag_s | mean_sq_gap |
|---|---|
| 0 | 0.00614 |
| 1 | 0.00581 |
| 2 | 0.00551 |
| 3 | 0.00534 |
| 5 | 0.00583 |
| 10 | 0.00811 |
| 15 | 0.01017 |
| 20 | 0.01212 |
| 30 | 0.01557 |
| 45 | 0.02028 |
| 60 | 0.02409 |
| 90 | 0.03095 |
| 120 | 0.03695 |
Best lag per market (markets with at least 20 used fills, n=346): median 3 s, 25th-75th percentile 2-5 s.
Buying at a trade's price when the model, read L seconds before the trade, disagrees by more than 0.10 (before fees, size ignored; CI by resampling markets):
| act_after_s | trades | markets | mean_per_share | ci_low | ci_high |
|---|---|---|---|---|---|
| 0 | 98501 | 335 | +0.0917 | +0.0157 | +0.1636 |
| 2 | 85392 | 321 | +0.0809 | -0.0020 | +0.1603 |
| 5 | 91239 | 331 | +0.0528 | -0.0224 | +0.1278 |
UPDATE
- Status: MEASURED.
- What it means: trades price BTC best with a lag of 3 s (per market median 3 s); the 1-minute history points trail BTC by about a minute, so they are not trade prices. Disagreement trades show +0.092 per share acting at once, CI +0.016 to +0.164, shrinking with each second of delay; before fees and without knowing if size was there.
- Next bets:
pm-15m-fast-taker(Does a taker who reads BTC within a second and buys Polymarket 15-minute Up/Down when the model disagrees by more than 0.10 make money after fees?) - Rejected:
pm-15m-minute-edge(the minute price history trails BTC by about a minute while trades trail it by about 3 s (pm-15m-lag), so the +0.18 per share seen on minute points comes from late-recorded points, not from prices anyone could trade (decided 2026-11-18))