DATA_PROVENANCE.md

Data Provenance and Methodology

This document explains exactly what data ships with this repository, how it was generated, and how the reproduction numbers relate to the production headline result (Brier 0.5783 over 97,000 real matches).

What is shipped

data/derived_sample.csv is a fully synthetic dataset: every row is procedurally generated by export_dataset.py from latent random team strengths, NOT sampled, aggregated, or derived from any third-party row.

It is not:

• A copy, subset, or statistical aggregate of football-data.co.uk CSV rows.
• A copy, subset, or statistical aggregate of any other third-party live-odds/match-data provider integrated elsewhere in the monorepo pipeline.
• An export of the production PostgreSQL sport_intelligence.fixture table.

The only thing carried over from the production pipeline is the feature schema (25 column names and their semantics, taken from app/sport_intelligence/features/db_extractor.py FEATURE_NAMES_V2) - this is this project's own original feature-engineering design, not third-party data, so publishing the schema and synthetic values under it carries no third-party licensing risk.

Why synthetic, not a real-data aggregate

The project's own ingestion code flags the football-data.co.uk licensing question as unresolved:

# ml-service/app/ingestion/datasets/football_results_2018_2026.py, line 37
LICENSE = "Verificare commercial use con football-data.co.uk"

A separate live-odds source integrated elsewhere in the monorepo pipeline carries an explicit is_public_redistribution_allowed=False marker in its own ingestion source configuration and must never be redistributed in any form (raw or aggregated).

Rather than resolve an unverifiable third-party licensing question by guessing, this repository sidesteps it entirely: the shipped sample contains zero third-party data, in any form, at any aggregation level. Anyone who wants to train on the real historical data can point export_dataset.py's feature schema at their own football-data.co.uk export (see "Bring your own data" below).

The Brier score: exact definition and scale

Both this repository's run_benchmark.py and the production app/sport_intelligence/training.py (compute_metrics) / app/sport_intelligence/models/metrics.py (brier_score_1x2) use the same, exact formula - the classic multi-class (3-outcome) Brier score, summed across classes per sample, then averaged across samples:

brier = mean_over_samples( sum_over_k in {home, draw, away} of (p_k - y_k)^2 )

where p_k is the calibrated predicted probability for outcome k and y_k is the one-hot ground truth (1 for the actual outcome, 0 otherwise).

Range: [0.0, 2.0] per sample (0.0 = perfect probabilistic prediction, 2.0 = maximally wrong with full confidence). This is the standard, textbook multi-class Brier formulation for a 3-outcome problem (Brier, 1950) - not a project-specific variant.

Reference points on this scale:

• Uniform-prior baseline (predicting 33/33/33 for every match): 2 x (1/3)^2 + (2/3)^2 = 2/9 + 4/9 = 6/9 ≈ 0.667.
• Production headline result: 0.5783 (a real, if modest, edge over the uniform baseline - consistent with the widely documented difficulty of football 1X2 prediction).

Reconciling 0.5783 vs. the project's separately-stated "<0.21" target

Project planning documents elsewhere state a target of "Brier < 0.21" for the Sport Intelligence model. This is not a contradiction - it is a different, but closely related, scale convention:

0.5783 / 3 ≈ 0.1928  (< 0.21)

Dividing the 3-class-summed Brier by the number of classes (3) yields the per-outcome average squared error, which is the convention used by some sports-analytics literature and dashboards that report Brier "per probability" rather than "summed across all three simultaneous probabilities for one match." Both are legitimate ways to report the same underlying calibration quality; this repository is explicit about using the summed (0.0-2.0 range) convention throughout, matching the project's own metrics.py source of truth, and shows the reconciling division here so the two numbers that appear in different project documents do not read as inconsistent.

What run_benchmark.py actually reproduces

run_benchmark.py reproduces the production methodology, not a bit-exact copy of the production number:

1. It fits the same class of meta-learner (LGBMClassifier, identical hyperparameters to app/sport_intelligence/training.py fit_meta_learner), with a LogisticRegression fallback if LightGBM is not installed - exactly mirroring the production fallback chain.
2. It fits 3 per-class IsotonicRegression calibrators, exactly as production does.
3. By default (in-sample mode, matching app/sport_intelligence/training.py train_full_pipeline, which fits and evaluates compute_metrics on the same (X, y) used for training - i.e. the production headline number is an in-sample metric, not a held-out validation score), it evaluates on the same rows used for fitting.
4. A --holdout flag is also provided, which instead does a 75/25 stratified split and evaluates strictly out-of-sample, for readers who want the more conservative, generalization-aware variant.

Because the shipped sample is synthetic and much smaller (a few thousand rows) than the real 97,000-match production dataset, the magnitude of the printed Brier score will differ from 0.5783 - a small synthetic sample lets a 200-tree LightGBM model fit the in-sample rows far more tightly than a real 97k-row noisy dataset would allow, producing a much lower (better-looking) in-sample number on the synthetic sample. This is expected and is not evidence of a bug: the point of run_benchmark.py is to prove the pipeline mechanics are correct and standalone-runnable, not to force-match a number computed on data this repository intentionally does not ship.

Trade-off: in-sample vs. held-out evaluation

Reporting an in-sample Brier score (as the production pipeline does) is a genuine methodological trade-off, not an oversight:

• Pro: it reflects the exact number the production system reports and monitors over time (drift_monitor.py compares against this same convention), and for a well-regularized model with isotonic calibration fit on the same fold, the gap to true out-of-sample performance is typically modest for large-N datasets (97k rows).
• Con: in-sample metrics are optimistic estimates of generalization by construction; a held-out or walk-forward backtest (see app/sport_intelligence/backtesting.py in the main monorepo) is the more rigorous number for claims about future predictive performance.

This repository ships both variants (run_benchmark.py default vs. run_benchmark.py --holdout) specifically so a reader can see the difference in magnitude directly, rather than being presented with only the more favorable number.

Bring your own data

To benchmark on real historical results instead of the synthetic sample:

1. Download historical CSVs directly from https://www.football-data.co.uk/data.php for the leagues/seasons you are interested in, subject to that site's own terms.
2. Compute the 25 FEATURE_NAMES_V2 columns from your own data (Dixon- Coles probabilities, Elo probabilities, market-implied probabilities, rolling form, days rest, head-to-head, lineup rating) - the feature engineering logic is documented in this repository's README and, in full, in the main portfolio monorepo's app/sport_intelligence/features/ and app/sport_intelligence/training.py.
3. Save the result as a CSV with the same column names as data/derived_sample.csv and pass it to run_benchmark.py --data your_file.csv.