# R&D result — prediction engine first

## What is now demonstrated

The multi-league dataset contains **25,195 complete matches** from seven leagues and 15 granular count targets (45 lines). Version 2 no longer chooses a model from one season:

- selection is the mean result from three leakage-safe walk-forward folds: 2020/21, 2021/22 and 2022/23;
- each fold trains only on seasons before its validation season;
- 2023/24 (2,364 matches per target) is a separate temporal confirmation, never used to select family, hyperparameters or the feature-removal policy; it is used only to decide calibration retention and assign evidence tiers;
- the already-seen 2024/25–2025/26 period is excluded from every reported metric;
- all feature vectors are produced before the current match updates history; the automated invariant passed for all 91,840 match-market feature sets.

Negative below means lower (better) Brier than the rolling team/opponent baseline on 2023/24. “Significant” uses an equal-league Student-t interval across the seven competitions, so within-league matches are not falsely treated as fully independent evidence.

| Target | Walk-forward selection | Mean Brier delta vs rolling | Significant lines vs rolling |
|---|---|---:|---:|
| Away shots | Poisson | -0.0193 | 3/3 |
| Home shots | Negative Binomial | -0.0174 | 3/3 |
| Away corners | Negative Binomial | -0.0129 | 3/3 |
| Home shots on target | Negative Binomial | -0.0128 | 3/3 |
| Home corners | Negative Binomial | -0.0104 | 3/3 |
| Away shots on target | Poisson | -0.0084 | 3/3 |
| Total shots | Poisson | -0.0080 | 3/3 |
| Total fouls | Negative Binomial | -0.0072 | 3/3 |
| Home fouls | Negative Binomial | -0.0065 | 3/3 |
| Away fouls | Negative Binomial | -0.0055 | 3/3 |
| Away yellow cards | Negative Binomial | -0.0042 | 2/3 |
| Total yellow cards | Negative Binomial | -0.0033 | 1/3 |
| Home yellow cards | Poisson | -0.0030 | 3/3 |
| Total shots on target | Poisson | -0.0022 | 0/3 |
| Total corners | Poisson | -0.0006 | 0/3 |

Eleven of the 15 targets beat the rolling baseline significantly on all three confirmation lines under the stricter league-clustered interval. The strongest recurring areas are team shots, team corners, team shots on target and fouls. `total_corners` and `total_shots_on_target` remain weak and must not inherit the confidence of their side-specific models.

## What the ablations say

A common regularized-Poisson diagnostic was repeated across all four temporal folds so that information value is not confused with model capacity.

- recent 5/10/20 form is the strongest recurring group for shots, shots on target, fouls and total corners;
- match context (standings proxy, opposition strength and rest) is strongest for cards and total fouls;
- style proxies are strongest for away corners and home shots on target;
- opponent adjustment is repeatedly useful, but smaller after recent/context/style features are present;
- H2H is usually small and inconsistent;
- referee history is useful for some cards/fouls folds, but cannot yet receive a universal weight because six of seven free league feeds have no named referee; missing `UNKNOWN` values are now kept at zero sample and never pooled as a fictional common referee;
- team×referee interactions are sparse and are not treated as established signal.

Feature removal was allowed only under a predeclared conservative rule: remove at most one group when all three selection folds improve and the mean gain is at least 0.0001. This removes `team_rate` for away corners, away fouls and total fouls, and `context` for home shots; the other 11 targets keep the complete feature set. The 2023/24 confirmation was not consulted for these choices.

## Calibration

Platt calibration is fitted on pooled out-of-fold predictions from the three selection seasons and tested only on 2023/24. It is retained per line only when **both** Brier and log loss improve on confirmation. Twenty-one of 45 line calibrators pass; all other lines keep raw model probabilities. Parameters used prospectively are then refitted on the four development out-of-fold seasons and versioned in the model artifact.

## What is not demonstrated

- No profitable ROI or yield is claimed: the historical dataset has match statistics, but not timestamped granular bookmaker lines and prices.
- Predictive skill versus rolling history is not proof of bookmaker edge.
- “Worldwide” coverage is not validated. Historical validation currently covers E0, SP1, D1, I1, F1, N1 and P1.
- Lineups and injuries are absent from this free historical dataset and are not silently imputed.
- Referee-dependent confidence outside E0 remains limited until a richer feed is accepted.
- The model is not commercially validated until immutable forward predictions and real pre-kickoff odds have sufficient settled sample size.

## Product consequence

The opportunity engine should prioritize lines for team shots, team corners, team shots on target and fouls when fresh compatible bookmaker markets exist. Cards remain useful but lower-evidence than the strongest side-stat models. Total corners and total shots on target stay tier C.

The engine still searches every supported market and produces Best Available, Conservative, Balanced, Aggressive and Best Value. `Strong Value` remains a separate flag and cannot activate until real-odds validation passes.