How our model works

1. Data before narrative

Every prediction starts with provider data normalized into Postgres. If a stat is missing, the UI shows unavailable — never a made-up placeholder. Data quality and confidence are always visible.

2. Elo ratings

Team strength begins with Elo ratings updated from historical results. Elo provides a stable baseline for how strong each side is relative to the league.

3. Logistic regression (1/X/2)

Win probabilities combine Elo with lightweight features: recent form, home advantage, rest days, head-to-head recency, and league position. A logistic model converts these features into home/draw/away probabilities.

4. Bivariate Poisson (goals)

Expected goals use a Bivariate Poisson model derived from the same feature set. This produces an expected goals range and supports both teams to score and underdog threat estimates.

5. Confidence buckets

6. Predictions Center (Top 10 picks)

Each UTC day we rank upcoming allowlist fixtures (status NS/TBD) that have a fresh pre-match model output. A fixture must pass configurable thresholds — by default: max outcome probability ≥ 55%, confidence ≥ MEDIUM, and data quality ≥ PARTIAL.

Ranking score = model_probability × confidence_score × data_quality_score. Confidence scores: HIGH 1.0, MEDIUM 0.85, LOW 0.65. Data quality: COMPLETE 1.0, PARTIAL 0.75. We show the top 10 after sorting. There are no betting odds, no bookmaker data, and no "value edge" language — only transparent model estimates.

Tune thresholds in deployment with TOP_PICKS_MIN_MODEL_PROBABILITY, TOP_PICKS_MIN_CONFIDENCE, TOP_PICKS_MIN_DATA_QUALITY, and TOP_PICKS_LIMIT.

7. AI explanation layer

The LLM receives structured model outputs and trusted context only. It explains probabilities, key factors, and scenarios — it does not invent statistics or override the model. Outputs are validated with Zod and cached until meaningful inputs change.

8. Limitations

Cold-start leagues and partial data reduce confidence. The MVP model is intentionally interpretable rather than a black-box neural network. Predictions are analytical estimates, not guarantees.