Sport statistics

Horse racing

Synthetic multi-runner fields with independent performance and pricing models for calibration research.

Dataset synthetic-v1 · generated 2026-07-26T03:36:42.407Z

Events

2,000

Average margin

18.15%

Model version

1.0.0

Calibration significance

Expected vs observed by participant

Each synthetic participant’s margin-adjusted price-implied rate against its observed win frequency over 2,000 simulated events, with a 95% Wilson interval, a proportion z-test, and a whole-field chi-square goodness-of-fit. Because the pricing and performance models are independent, well-calibrated sides land inside the Wilson band.

χ² 15.19 · df 8approx p 0.0552 significant bandsSynthetic calibration · not a wagering signal
BandEventsExpectedObserved95% WilsonDifferenceSignificance
Runner 12,00022.96%24.00%22.18% 25.92%+1.04%within noise
Runner 22,00022.96%24.10%22.28% 26.02%+1.14%within noise
Runner 32,00013.92%14.90%13.41% 16.53%+0.98%within noise
Runner 42,00013.17%11.50%10.18% 12.97%-1.67%p 0.027
Runner 52,0008.80%7.20%6.15% 8.42%-1.60%p 0.011
Runner 62,0006.82%6.30%5.32% 7.45%-0.52%within noise
Runner 72,0004.62%4.90%4.04% 5.94%+0.28%within noise
Runner 82,0003.44%3.50%2.78% 4.40%+0.06%within noise
Runner 92,0003.31%3.60%2.87% 4.51%+0.29%within noise

Edge detection

14 discrepancies0 strong2 notable12 infoSynthetic · look-ahead signals are illustrative only

Edge detection

Mathematical discrepancies

14 synthetic discrepancy signals detected (0 strong, 2 notable). Illustrative only — look-ahead signals are not realizable live edges.

  • Runner 1: priced vs fair

    The priced probability sits above the de-vigged fair line — the built-in margin you pay on this selection.

    notable

    +4.07%

  • Runner 2: priced vs fair

    The priced probability sits above the de-vigged fair line — the built-in margin you pay on this selection.

    notable

    +4.07%

  • Runner 4: calibration drift

    Observed win rate differs from the price-implied rate by a statistically significant margin in this sample.

    info

    -4.00%

  • Runner 1: model vs marketlook-ahead

    Gap between the simulation's true probability and the priced implied probability. Look-ahead: the true value is unknown in real markets.

    info

    -3.39%

  • Runner 4: model vs marketlook-ahead

    Gap between the simulation's true probability and the priced implied probability. Look-ahead: the true value is unknown in real markets.

    info

    -3.38%

  • Runner 5: calibration drift

    Observed win rate differs from the price-implied rate by a statistically significant margin in this sample.

    info

    -3.16%

  • Runner 2: model vs marketlook-ahead

    Gap between the simulation's true probability and the priced implied probability. Look-ahead: the true value is unknown in real markets.

    info

    -3.15%

  • Runner 1: calibration drift

    Observed win rate differs from the price-implied rate by a statistically significant margin in this sample.

    info

    -3.03%

  • Runner 3: priced vs fair

    The priced probability sits above the de-vigged fair line — the built-in margin you pay on this selection.

    info

    +2.47%

  • Runner 4: priced vs fair

    The priced probability sits above the de-vigged fair line — the built-in margin you pay on this selection.

    info

    +2.33%

  • Runner 5: model vs marketlook-ahead

    Gap between the simulation's true probability and the priced implied probability. Look-ahead: the true value is unknown in real markets.

    info

    -2.22%

  • Runner 3: model vs marketlook-ahead

    Gap between the simulation's true probability and the priced implied probability. Look-ahead: the true value is unknown in real markets.

    info

    -2.11%

  • Runner 5: priced vs fair

    The priced probability sits above the de-vigged fair line — the built-in margin you pay on this selection.

    info

    +1.56%

  • Runner 6: priced vs fair

    The priced probability sits above the de-vigged fair line — the built-in margin you pay on this selection.

    info

    +1.21%

Synthetic, look-ahead discrepancies — illustrative only. They use the simulation’s known true probabilities and are not realizable live edges.

Real-world use case

  • Reload a sport's statistics page with a different batch seed and check whether a flagged band stays significant — one-off flags are usually noise.
  • Compare a favourite band's tight Wilson interval against a longshot band's wide one to see why long prices need far more events before a deviation means anything.
More in the breakdowns guide →