Method
Matty The Bookie publishes NFL predictions generated from a separate model pipeline. The website consumes exported prediction artifacts and presents the model output in a public, readable format.
Picks are model projections, not guarantees. The site is designed to show confidence, margin, odds, and matchup context so each pick can be evaluated on its own terms.
For Sports Geeks
These are scoreboard-style reads from the published evaluation window. The headline 65.3% accuracy number is Straight-Up (SU) winner accuracy, not ATS. It measures whether the model picked the outright winner across completed historical games.
Straight-up accuracy
Straight-up accuracy counts whether the predicted winner won the game outright. A 65.3% SU mark means the model selected the actual winner in roughly two out of three evaluated games.
Applied as: Baseline SU winner quality. This is separate from spread picks and is not an ATS hit rate.
Vegas favorite accuracy
Vegas favorite accuracy shows how often the market favorite won outright over the same completed-game window.
Applied as: Market benchmark. The model is judged against the no-vig favorite, not just against random guessing.
Model/Vegas agreement
Agreement rate measures how often the model and no-vig Vegas favorite picked the same side.
Applied as: Signal routing. High agreement means many games are market-consensus reads rather than contrarian edges.
Model accuracy in disagreements
This is the model's SU accuracy only in games where it picked against the no-vig Vegas favorite.
Applied as: Contrarian caution flag. Model disagreement games are reviewed separately before treating them as betting edges.
Disagreement moneyline ROI?
Disagreement moneyline ROI measures historical returns when the model picked against the no-vig Vegas favorite and the model side was bet at listed moneyline odds.
Applied as: Market-edge check. It tracks flat one-unit ROI when backing the model side only in favorite-disagreement games.
Best spread ROI
Best spread ROI is the top corrected spread backtest row from the current evaluation artifacts.
Applied as: ATS research filter. It identifies where the spread model has tested best, not a guarantee for every spread pick.
Best spread edge bucket ROI
The best spread edge bucket groups games by absolute model-vs-market spread edge and reports the best qualifying ROI bucket.
Applied as: Spread-edge bucket sizing. This helps separate small theoretical edges from ranges that have historically performed better.
IT and Math Nerds
These diagnostics explain the probability model underneath the picks. They are used to compare feature sets, tune probability calibration, and avoid models that look confident while being poorly calibrated.
AUC
AUC measures how well predicted probabilities separate winners from losers across the evaluated games. Higher is better.
Applied as: Feature selection and rank quality. Better AUC means the model is doing a better job ordering stronger teams above weaker teams.
Brier score
Brier score is the average squared error of the win probability forecast. Lower is better.
Applied as: Probability calibration. It rewards probabilities that are close to reality and punishes confident misses.
Log loss
Log loss penalizes incorrect high-confidence predictions more harshly than modest misses. Lower is better.
Applied as: Overconfidence control. It helps prevent the model from assigning extreme probabilities unless the underlying evidence supports them.
F1 score
F1 balances precision and recall for the home-win classification task.
Applied as: Classification balance. It helps catch models that look accurate only because they over-pick one class.
Margin MAE
Margin MAE is the average absolute difference between predicted margin and actual game margin.
Applied as: Spread modeling. Lower margin error improves the quality of model-vs-market spread comparisons.
Margin RMSE
Margin RMSE is another margin-error score that penalizes large misses more than MAE. Lower is better.
Applied as: Blowout sensitivity. It highlights whether large margin misses are hurting the model.
Evaluation starts with the 2018 season window and excludes live games at and after Week 7 so current unpublished outcomes do not leak into the scoring set.