Poisson v2 – rolling evaluation
Report #8 · 23.08.2026 07:55:19 · source backtest #6
Development rows11 229
Validation rows2 279
Selected config—
Holdout readNE
Development rolling test
2019/20–2023/24. Každý zápas používá jen výsledky před jeho časem. Kalibrace je naučená na těchto rolling OOS predikcích; její validní kontrola je až níže na 2024/25.
| Model | Brier | Log loss | Calibration error | Počet |
|---|---|---|---|---|
| Base-rate | — | — | — % | 0 |
| Poisson v1 | — | — | — % | 0 |
| Poisson v2 raw | — | — | — % | 0 |
| Poisson v2 calibrated | — | — | — % | 0 |
| Poisson v2 Dixon–Coles calibrated | — | — | — % | 0 |
| Opening market | — | — | — % | 0 |
Validation 2024/25
Specifikace a kalibrátor jsou zde už zmrazené podle developmentu. Opening market zůstává pouze benchmarkem.
Dixon–Coles vychází pro agregovanou pravděpodobnost Over/Under 2,5 stejně jako kalibrovaný Poisson očekávaně: standardní tau korekce pouze přerozděluje nízké skóre uvnitř intervalu 0–2.
| Model | Brier | Log loss | Calibration error | Počet |
|---|---|---|---|---|
| Base-rate | — | — | — % | 0 |
| Poisson v1 | — | — | — % | 0 |
| Poisson v2 raw | — | — | — % | 0 |
| Poisson v2 calibrated | — | — | — % | 0 |
| Poisson v2 Dixon–Coles calibrated | — | — | — % | 0 |
| Opening market | — | — | — % | 0 |
Výběr hyperparametrů
Řazení je podle development Brier score, při shodě podle development log loss. ROI se do výběru nepoužívá.
| Config | Decay half-life | Shrinkage prior | Carry-over | Dev Brier | Dev log loss | Validation Brier | Validation log loss |
|---|
Rolling vývoj po sezonách
| Sezona | V2 raw Brier | V2 calibrated Brier | DC calibrated Brier | V2 raw log loss | V2 calibrated log loss |
|---|---|---|---|---|---|
| 1920 | — | — | — | — | — |
| 2021 | — | — | — | — | — |
| 2122 | — | — | — | — | — |
| 2223 | — | — | — | — | — |
| 2324 | — | — | — | — | — |
| 2425 validation | — | — | — | — | — |
Metodika a zmrazení
{
"M0": "Opening no-vig market baseline.",
"M1": "Frozen market-offset regularized logistic model using Poisson v2 and pre-match metadata features.",
"M2_linear": "Market-offset regularized logistic model with the frozen M1 features plus rolling match statistics.",
"M2_boost": "A deliberately small depth-1 boosted-stump residual model; no external boosting library, no hyperparameter search, fixed rounds, learning rate, leaf cap, and L2.",
"new_information": [
"shots",
"shots on target",
"corners",
"short versus long form",
"shot conversion",
"rest days",
"rolling tempo",
"statistical history and data quality"
],
"rolling": "Each development season is predicted from earlier development seasons; validation is fit on development only. Same-time matches are predicted before any result is added.",
"league_quality": "Stats are modeled relative to league-season baselines. Serie A shot fields retain an explicit measurement caveat and are not globally pooled.",
"selection": "Only development Brier\/log loss select M2-linear regularization. M2-boost configuration is fixed in advance. ROI is excluded.",
"holdout": "The later holdout was not queried, read, trained on, tuned on, or evaluated.",
"main_metrics": [
"Brier score",
"log loss"
],
"immutable_result": true
}
{
"m1_regularization_frozen": 100,
"m2_linear_selected_regularization": 100,
"m2_boost_configuration": {
"rounds": 12,
"learning_rate": 0.05,
"l2": 50,
"quantiles": [
0.25,
0.5,
0.75
],
"max_leaf_value": 0.15
},
"linear_primary_metric": "development rolling Brier score",
"linear_secondary_metric": "development rolling log loss",
"boost_hyperparameters_tuned": false,
"validation_used_for_selection": false,
"roi_used_for_selection": false
}