EDGE Bet Model dashboard

NRL Margin Predictor

The margin predictor is an operator-run workflow. dbt builds point-in-time
features, Python trains and predicts, and the FastAPI website reads the
prediction output.

Train and Evaluate

cd analytics/dbt_nrl
dbt build --select ml_nrl_match_training_examples

cd ../..
python ml/evaluate_margin_model.py --holdout-season 2025
python ml/train_margin_model.py --train-through-season 2025

Model artifacts are written to:

nrl/models/lgbm_v1.pkl   ← trained LightGBM model
nrl/models/metrics.json  ← evaluation metrics

Weekly Prediction

bash scripts/weekly_2026_round.sh <ROUND>
bash scripts/stage_nrl_datasets.sh weekly

cd analytics/dbt_nrl
dbt build

cd ../..
python ml/predict_weekly_margin.py --season 2026 --round <ROUND>

predict_weekly_margin.py writes predictions to gold_nrl_predictions and
exports a Parquet artifact. The FastAPI bet-advisor website uses the separate
logit/Elo backtest pipeline in nrl-bet-advisor/data/backtest.py for the
game-day value board, rather than the LightGBM margin model.

DQ Checks

The dbt ML layer includes checks for:

  • one training row per completed match
  • one upcoming row per unscored match
  • exactly two team-form and ladder feature rows per fixture
  • no future match or ladder data in point-in-time features
  • latest-prior match and latest-prior ladder selection
  • rolling feature ranges such as last-3/last-5 counts and win rates
  • target correctness for home_margin