The unfitted market remains the strongest totals probability benchmark.
Loading the latest model data…
Reference forecasts · no official picks
Every model uses the same 1,900 whole-season holdout matches. Lower log loss and goal error are better. “Frozen live” marks the unchanged forward model; “best historical” is newer research only.
| Model | Inputs | Status | Probability error ↓ | Result accuracy | Avg goal error ↓ | Exact score |
|---|---|---|---|---|---|---|
| Closing 1X2 market | Average closing home/draw/away prices | Benchmark | 0.9556 | 55.9% | — | — |
| Poisson · shot form | Rest, experience, Elo, shots, and shots on target | Best historical | 0.9755 | 55.1% | 1.320 | 11.9% |
| Poisson · goals + shots | Frozen live form, rest, experience, and Elo | Frozen live | 0.9781 | 54.9% | 1.321 | 11.7% |
| Attack/defense Poisson | Time-weighted team attack and defense | Historical | 0.9992 | 50.6% | — | — |
| Dixon–Coles | Time-weighted score model with low-score correction | Historical | 0.9992 | 50.6% | — | — |
Shot form improved the earlier model, but not enough to beat the closing market. Venue-specific form and prior-season strength also weakened the primary probability score, so neither was promoted.
The totals tournament now covers 102 fixed classifier-and-feature combinations, including venue form and decayed prior-season strength.
The unfitted market remains the strongest totals probability benchmark.
Elastic-net logistic · market only
Extra trees · market + venue goal form; -16.65 units on 160 historical bets.
Historical Poisson · shot form at the predeclared 3% expected-return gate. DNB was tested only where the recorded closing Asian handicap was exactly 0.0. Draws push; opposing wins lose one unit.
| Held-out season | Bets | ROI | Read |
|---|---|---|---|
| 2021-22 | 24 | -13.8% | Negative |
| 2022-23 | 34 | +30.4% | Positive |
| 2023-24 | 26 | +8.9% | Positive |
| 2024-25 | 16 | -8.8% | Negative |
| 2025-26 | 26 | +0.7% | Positive |
The interval crosses zero and does not account for model screening. Results also reverse by season, while average closing prices are not proof of recommendation-time execution. The next valid step is a frozen forward DNB ledger with timestamped book prices.
The first card lists live-model inputs. The others separate completed free-data experiments from information that still needs a trustworthy timestamped source.
Most likely exact score: the single scoreline with the highest probability, not the average score.
Average goals: expected home plus away goals. The probability of 3+ goals comes from the same score distribution.
Moneyline view: home, draw, and away probabilities with fair no-vig American odds. Named sportsbook prices are timestamped captures for comparison only, may have moved, and are not guaranteed to be executable.
Draw no bet: a draw refunds the stake. Historical DNB evidence uses only recorded zero-handicap closing prices and is not a live recommendation.
Frozen live model: the forward-test specification is unchanged while it accumulates evidence. Better historical research is labeled separately to avoid rewriting the test.
Free-data policy: the page uses free historical files and a bounded free odds quota. Missing lineup, weather, or availability data is disclosed rather than inferred.