Accuracy and performance
Accuracy is correct draw calls divided by settled predictions. Void matches and pending matches are excluded. Every percentage shows the number of predictions behind it.
Record by period
| Period | Published | Correct | Incorrect | Void | Pending | Settled | Accuracy |
|---|---|---|---|---|---|---|---|
| Today | 4 | 0 | 0 | 0 | 4 | 0 | No data |
| Last 7 days | 18 | 6 | 8 | 0 | 4 | 14 | 43% of 14 |
| Last 30 days | 136 | 41 | 90 | 1 | 4 | 131 | 31% of 131 |
| All time | 226 | 56 | 143 | 1 | 26 | 199 | 28% of 199 |
Periods are based on kickoff date in Africa/Lagos. Void matches (postponed, cancelled, awarded, or with no result after 3 days) are shown but never counted as wins or losses.
Hit rate as results come in
The solid line is the share of draw calls that were correct so far. The dashed line is the average draw chance we published for those matches. If the model is well calibrated the two should drift together as the sample grows. Early on they will not.
By competition
| Competition | Published | Settled | Correct | Accuracy |
|---|---|---|---|---|
| Premier League | 30 | 26 | 6 | 23% of 26 |
| La Liga | 18 | 17 | 7 | 41% of 17 |
| Serie A | 6 | 6 | 2 | 33% of 6 |
| Ligue 1 | 12 | 11 | 1 | 9% of 11 |
| Championship | 45 | 40 | 7 | 18% of 40 |
| Eredivisie | 5 | 5 | 1 | 20% of 5 |
| Primeira Liga | 10 | 10 | 1 | 10% of 10 |
| Série A (Brazil) | 30 | 21 | 6 | 29% of 21 |
| NPFL | 18 | 16 | 6 | 38% of 16 |
| Premier League | 19 | 19 | 7 | 37% of 19 |
| Botola Pro | 10 | 10 | 4 | 40% of 10 |
| Premier League | 2 | 2 | 1 | 50% of 2 |
| Ligue 1 | 16 | 11 | 4 | 36% of 11 |
| Ligue 1 | 4 | 4 | 2 | 50% of 4 |
| CAF Champions League | 1 | 1 | 1 | 100% of 1 |
By estimated draw chance
| Published estimate | Settled | Average estimate | Actual draw rate |
|---|---|---|---|
| Below 26% | 21 | 26% | 29% of 21 |
| 26% to 28% | 96 | 27% | 20% of 96 |
| 28% to 30% | 44 | 29% | 34% of 44 |
| 30% and above | 38 | 32% | 42% of 38 |
A calibrated model has an actual draw rate close to its average estimate in each row. Rows with few settled predictions are noisy.
Probability quality (live predictions)
| Measure | Our estimates | Baseline: competition average for every match |
|---|---|---|
| Brier score (lower is better), n=199 | 0.1997 | 0.2008 |
| Log loss (lower is better), n=199 | 0.5883 | 0.5910 |
These are measured only on matches we picked because the model liked them, so the baseline here is not a like-for-like contest. The backtest below compares on all eligible matches.
Backtest on past matches
Run with model draw-v1.0. Each week was predicted using only earlier matches.
| Measure | Model | Baseline |
|---|---|---|
| Brier score, n=5,218 | 0.1925 | 0.1938 |
| Log loss, n=5,218 | 0.5726 | 0.5759 |
Improvement over baseline (Brier skill): 0.65%. A positive number means the model scored better than the baseline. Gains in draw prediction are usually small.
Matches at or above our publishing threshold (26%): 961 of 3409 ended level (28%), against 26% for all eligible matches.
| Estimate range | Matches | Average estimate | Actual draw rate |
|---|---|---|---|
| 15% to 18% | 50 | 17% | 12% |
| 18% to 21% | 208 | 20% | 16% |
| 21% to 24% | 564 | 23% | 20% |
| 24% to 27% | 1623 | 26% | 26% |
| 27% to 30% | 1661 | 28% | 28% |
| 30% to 33% | 858 | 31% | 29% |
| 33% to 36% | 229 | 34% | 35% |
Ranges with fewer than 20 matches are hidden. Covers 2023-11-02 to 2026-11-14.
Estimates based on historical data. Historical patterns do not guarantee any result, and no prediction here is a sure bet.