Methodology & Validation Notes
This page is written as a compact quantitative note. The public interface exposes only a constrained subset of the production stack, but the core validation principles are stated explicitly.
QuantSport builds probability baselines from structured sports data, archived market states, and temporal context. The goal is not narrative commentary but auditable probability estimation.
Model families differ by sport, but the shared rule is consistent: training, selection, and evaluation are all performed on temporally ordered data so that the public layer reflects out-of-sample behavior rather than hindsight optimization.
Historical matches, lineups, injuries, schedule load, market states, and archived settlements are normalized into sport-specific feature stores.
Football, NHL, basketball, and tennis do not share one monolithic feature recipe. Each pipeline is built around the physics and information structure of the sport.
Every material model update is evaluated in walk-forward fashion so that future information never leaks into earlier prediction rows.
Directional hit-rate alone is not sufficient. Log Loss, Brier Score, calibration drift, and divergence versus archived market prices are monitored continuously.
The ingestion layer synchronizes raw event data, archived prices, roster context, and settlement records into normalized intermediate tables. Rolling windows, exponential moving averages, opponent-strength adjustments, and schedule-based variables are then derived in sport-specific feature builders.
This means football can emphasize xG and match-state structure, NBA can emphasize possession efficiency and rest/fatigue, while tennis can emphasize surface-specific ELO, serve-return components, and tournament fatigue.
Training and evaluation are separated by time, not by random shuffling. A model only sees information that would have been available before the prediction timestamp of the fixture it is asked to score.
Walk-forward validation is used because it mirrors real deployment conditions. This prevents silent leakage from future matches, later market states, or ex-post settlement information into the training fold.
Raw model scores are not assumed to be calibrated probabilities by default. Probability quality is evaluated with Log Loss, Brier Score, and divergence analysis relative to archived market probability baselines.
Calibration gates determine whether a model state is suitable for publication. A model can be directionally useful while still failing probability-quality checks; in that case the public layer remains constrained until calibration improves.
The public site does not expose the full research warehouse. It serializes a narrower auditable layer: processed events, confirmed archived prices when available, settlement status, model metrics, and explicitly invalidated dates when a recommendation path is known to have been defective.
- Probability quality: Log Loss, Brier Score, and mean probability divergence
- Operational consistency: archive completeness, settlement timing, and publication-state sanity
- Temporal robustness: walk-forward holdout behavior by sport and market family
- Public-surface quality gates: filters that prevent weak or structurally invalid rows from entering the visible archive
The public archive is a transparency layer, not a guarantee of future opportunity. Historical out-of-sample behavior can still be noisy because market variance is inherent to the domain.
When a recommendation path is later found to have been structurally invalid, the affected day is excluded and annotated rather than silently rewritten.