AI model performance: October 2026
Figures are frozen on the publication date. Corrections are only added as dated notes at the bottom.
Key findings
- In October 2026, margin-free bookmaker odds were more accurate than the AI model in every market measured (4). On 1X2 the market's Brier score was 0.592 against 0.629 for the calibrated AI model (difference +0.037, 95% interval +0.032 to +0.043; 12,829 matches).
- Calibration helped most in over/under 2.5 goals: the Brier score fell from 0.524 for the raw model to 0.495 (improvement +0.028, 95% interval +0.023 to +0.033; 8,771 events).
- On 1X2, the Brier difference between the AI model and the market was lowest in “Third tier and below” (+0.030; 4,762 matches) and highest in “Top tier” (+0.046; 3,164 matches). The intervals overlap, so this difference between groups is not reliable.
- Calibration: on 1X2 the calibrated AI model was off by 0.72 percentage points on average (expected calibration error), the market by 0.62.
Markets this month
The same events for every forecast. Lower Brier is better. The difference is the calibrated AI model minus the market, with a 95% interval from resampling days.
| Market | Events | Brier raw | Brier calibrated | Brier market | Brier base rate | Difference (95% interval) | Log loss AI / market |
|---|---|---|---|---|---|---|---|
| 1X2 (full-time result) | 12,829 | 0.630 | 0.629 | 0.592 | 0.647 | +0.037+0.032 … +0.043 | 1.046 / 0.992 |
| Goals over/under 2.5 | 8,771 | 0.524 | 0.495 | 0.476 | 0.495 | +0.020+0.016 … +0.023 | 0.689 / 0.668 |
| Both teams to score | 11,492 | 0.521 | 0.495 | 0.489 | 0.494 | +0.006+0.003 … +0.009 | 0.688 / 0.682 |
| StatsAI Elo (1X2) | 1,202 | 0.598 | 0.600 | 0.578 | 0.639 | +0.021+0.014 … +0.030 | 1.004 / 0.973 |
By competition tier
A filtered view: the same compared events as in the tables above, split by competition tier. The overall result is above; this only shows where the gap between the AI model and the market is larger or smaller. Groups with at least 1,000 events are shown. 99.9% of events are classified.
Top tier is a country's first division, second tier its second division, and third tier and below everything lower (in England from League One down). Cups, national teams, confederation club competitions, youth, reserve and women's football form a separate group.
| Market | Tier | Events | AI model, calibrated | Market (odds without margin) | Difference (95% interval) |
|---|---|---|---|---|---|
| 1X2 (full-time result) | Top tier | 3,164 | 0.621 | 0.575 | +0.046+0.037 … +0.056 |
| Second tier | 2,570 | 0.644 | 0.606 | +0.039+0.030 … +0.048 | |
| Third tier and below | 4,762 | 0.640 | 0.610 | +0.030+0.018 … +0.040 | |
| Cups, national teams, youth and women | 2,332 | 0.601 | 0.562 | +0.039+0.028 … +0.049 | |
| Goals over/under 2.5 | Top tier | 2,473 | 0.493 | 0.477 | +0.016+0.010 … +0.022 |
| Second tier | 1,880 | 0.499 | 0.488 | +0.011+0.003 … +0.016 | |
| Third tier and below | 3,013 | 0.498 | 0.472 | +0.026+0.019 … +0.035 | |
| Cups, national teams, youth and women | 1,405 | 0.490 | 0.466 | +0.024+0.015 … +0.035 | |
| Both teams to score | Top tier | 2,954 | 0.496 | 0.488 | +0.008+0.003 … +0.013 |
| Second tier | 2,430 | 0.497 | 0.493 | +0.004−0.001 … +0.009 | |
| Third tier and below | 4,165 | 0.493 | 0.488 | +0.0050.000 … +0.011 | |
| Cups, national teams, youth and women | 1,942 | 0.497 | 0.488 | +0.010+0.002 … +0.019 |
By week
1X2 Brier score by week (Monday to Sunday). A hollow point means the week is only partly in the measurement.
Calibration
Expected calibration error (ECE) for 1X2 (full-time result): AI raw 1.25, calibrated 0.72, market 0.62 percentage points.
Expected calibration error (ECE) for Goals over/under 2.5: AI raw 9.81, calibrated 4.82, market 1.53 percentage points.
Expected calibration error (ECE) for Both teams to score: AI raw 9.58, calibrated 3.72, market 2.12 percentage points.
Expected calibration error (ECE) for StatsAI Elo (1X2): AI raw 1.81, calibrated 2.04, market 2.63 percentage points.
Method in brief
- Everything is measured: every resolved prediction of the month, good or bad. Filtered views are always labelled.
- The AI model and the market are compared on the same events; the market is pre-match odds with the margin removed proportionally.
- Calibration is out of sample: each day's curve is learned only from the 90 days before that day.
- 95% intervals come from resampling whole days. Groups with fewer than 1,000 events are not used for conclusions.
- Before publication two independent calculations must agree to four decimals, and the text is checked in all three languages.
- No profit or ROI: the measures are Brier score, log loss, accuracy and calibration.
Checked before publication: two independent calculations agree on 52 compared figures (largest difference < 0.000001).
Limitations
- 5,710 predictions on 751 matches have no result (postponed, abandoned or result not found) and are not measured.
- One month is a short period; conclusions about a whole season need several months.
- Odds come from the pre-match offer, not closing odds; the real market is probably even more accurate.
- A good Brier score does not mean profit at a bookmaker. This is statistics, not betting advice.
How to cite
StatsAI (2026). AI model performance: October 2026. StatsAI Football Prediction Research. Published 2 Nov 2026. https://statsai.bet/en/research/ai-model-performance-2026-10
The figures and data may be reused freely with attribution (CC BY 4.0). Data for this report (JSON, CC BY 4.0)