When the order book says 70%, how often does it happen?
The short answer
When the book priced Up between 70% and 75% one minute before the close — 72.4% on average — Up won 73.3% of the time: 1,533 windows, 95% interval 71.0%–75.5%. The price sits inside that interval.
The same band in each half of the period: 74.5% of 826 windows (Aug 7 – Sep 13), then 72.0% of 707 (Sep 14 – Oct 21); the book's price was inside both intervals.
Over all 20 bands of 5 points, the book's average price fell inside the 95% interval of what happened in 14 of 20. Split in two, the period gives 10 of 20 and then 18 of 20: the second half was better calibrated than the first.
The reliability curve
| book price | windows | book average | Up happened | 95% interval | calibrated |
|---|---|---|---|---|---|
| 0–5% | 42,332 | 1.19% | 0.86% | 0.78%–0.96% | no |
| 5–10% | 5,189 | 7.04% | 6.59% | 5.95%–7.30% | yes |
| 10–15% | 2,898 | 12.23% | 11.42% | 10.31%–12.63% | yes |
| 15–20% | 2,131 | 17.23% | 14.88% | 13.43%–16.45% | no |
| 20–25% | 1,815 | 22.27% | 20.94% | 19.13%–22.87% | yes |
| 25–30% | 1,585 | 27.23% | 25.11% | 23.04%–27.30% | yes |
| 30–35% | 1,412 | 32.40% | 31.09% | 28.73%–33.55% | yes |
| 35–40% | 1,370 | 37.36% | 34.74% | 32.27%–37.31% | no |
| 40–45% | 1,207 | 42.28% | 40.27% | 37.53%–43.06% | yes |
| 45–50% | 1,324 | 47.43% | 44.71% | 42.05%–47.40% | no |
| 50–55% | 2,630 | 51.22% | 53.46% | 51.55%–55.36% | no |
| 55–60% | 1,203 | 57.35% | 57.27% | 54.46%–60.04% | yes |
| 60–65% | 1,301 | 62.40% | 62.95% | 60.29%–65.53% | yes |
| 65–70% | 1,433 | 67.41% | 68.67% | 66.22%–71.02% | yes |
| 70–75% | 1,533 | 72.38% | 73.32% | 71.05%–75.47% | yes |
| 75–80% | 1,819 | 77.40% | 79.05% | 77.12%–80.86% | yes |
| 80–85% | 2,087 | 82.48% | 82.56% | 80.87%–84.13% | yes |
| 85–90% | 2,818 | 87.57% | 88.29% | 87.05%–89.42% | yes |
| 90–95% | 4,812 | 92.70% | 92.79% | 92.02%–93.49% | yes |
| 95–100% | 43,047 | 98.83% | 99.27% | 99.18%–99.34% | no |
At the extremes
68.9% of all windows were priced below 5% or at 95% and above one minute before the close. Below 5% (42,332 windows) the book said 1.19% on average and Up happened 0.86% (0.78%–0.96%). At 95% and above (43,047 windows) it said 98.83% and Up happened 99.27% (99.18%–99.34%).
At both ends the outcome was more one-sided than the price: the side priced near zero came in 0.32 points less often than its price below 5%, and 0.44 points less often above 95%. Both halves show the same direction below 5% (0.93% then 0.81% for 1.26% and 1.12% priced).
By duration
| duration | windows | order book | Student-t model | normal model | book bands calibrated |
|---|---|---|---|---|---|
| 5-minute | 87,498 | 0.0599 | 0.0615 | 0.0629 | 17 of 20 |
| 15-minute | 29,157 | 0.0345 | 0.0344 | 0.0354 | 16 of 20 |
| hourly | 7,291 | 0.0279 | 0.0254 | 0.0262 | 7 of 10 |
On 5-minute windows the book had the best score (0.0599 against 0.0615 for the Student-t model). On 15-minute windows the two were level: 0.0345 and 0.0344. On hourly windows our Student-t model scored best, 0.0254 against 0.0279 for the book.
| nights | windows | order book | Student-t model | normal model |
|---|---|---|---|---|
| Aug 7 – Sep 13 | 61,944 | 0.0569 | 0.0584 | 0.0599 |
| Sep 14 – Oct 21 | 62,002 | 0.0473 | 0.0476 | 0.0487 |
Our two models against the same windows
Our models were calibrated in far fewer bands than the book: 3 of 20 for the Student-t model and 2 of 20 for the normal model. Between 10% and 90%, the weighted gap between predicted and observed was 1.4 pts for the book, 3.4 pts for the Student-t model and 7.0 pts for the normal model.
The direction is the same for both models: outcomes landed further from 50% than they predicted. Worst band of the Student-t model: 60–65%, 62.5% predicted, 67.9% observed (1,571 windows). Worst of the normal model: 65–70%, 67.6% predicted, 77.7% observed.
This is why our pages show the model next to the market's price and publish both scores every night, rather than presenting the model as the better forecast.
How it was measured
Population: every Polymarket crypto Up or Down window (Bitcoin, Ethereum, Solana, XRP; 5-minute, 15-minute and hourly) whose session date falls in Aug 7 – Oct 21, 2026 and that Polymarket resolved Up or Down — 123,966 windows over 76 nights. Outcomes are Polymarket's official resolutions.
Forecasts at T−60 s: the order book's mid price on the Up side, and our two model probabilities (normal and Student-t), all read from the first snapshot taken 60 seconds or less before the close. 20 windows without that snapshot are left out of every figure, so the three forecasts are scored on the same 123,946 windows.
Bands: 20 bands of 5 points; a band with fewer than 100 windows is merged with its neighbour closest to 50%, series by series. Interval: Wilson 95% on the observed frequency. Brier score: mean of (forecast − outcome)², pooled over windows.
Variability: the 76 nights are cut into two halves of 38 and 38 nights, and every ratio is also given on each half.
Limits
The intervals treat each window as independent. They are not: the four assets move together and the 5-minute, 15-minute and hourly windows overlap in time, so the real uncertainty is wider than shown.
Polymarket changed the 5- and 15-minute rule on Aug 7, 2026 (a 60-second Chainlink average instead of the last price), which is why this study starts that day. Our models took that rule into account from Aug 18; the first 11 nights use the earlier model version. The book's figures are not affected.
The book mid comes from our collector's snapshot. When the live feed goes quiet, a snapshot can rest on a slower price request, and the window record keeps no trace of it.
The pooled Brier score here is not the nightly average shown on /performance: a mean of nightly scores weights each night equally, a pooled score weights each window equally.
Related
Data analytics on Polymarket crypto markets · not financial advice · 18+. Figures from our own records of every window; outcomes are Polymarket's official resolutions.