We Froze the Model on 2024 Data. Then We Replayed the 2025 Season.
5,227 player-week projections. The model was trained only on 2016–2024 data, its weights frozen, then replayed against the full 2025 season as a held-out test — no 2025 data touched training, tuning, or feature selection. Each week's projection uses only the player's prior-game stats, so every number is what the frozen model would have projected with pregame information. This replay was generated after the season ended (July 2026); it is a holdout scorecard, not a live-season track record. Below is exactly how close the model got, week by week, position by position, and why each projection landed where it did.
How Close Are We? — Hit Rate by Position
What percentage of projections land within a tight band of the actual result? The production score is a TD-free composite (yards + receptions, position-weighted). For a QB scoring 12 (~250 pass yds + 30 rush yds), being within ±3 means our projection is roughly ±60 passing yards off — tight enough to act on.
Position
Sample
Within ±3
Within ±5
Within ±8
Median Error
90th Pctile
Each bar shows what percentage of player-week projections had an absolute error in that range. The green zone (0-3) is where the projection is directly actionable for lineup and game-plan decisions. Yellow is close enough to be useful with context. Red means we missed — and we show those honestly.
Point Accuracy — Every Stat Beats the Floor
How far off is the average projection for each stat? We compare against a 3-game rolling average — the simplest predictor that carries any signal. Every position × stat beats this floor.
Stat
Sample
Model MAE
Baseline MAE
Improvement
Boom / Bust Detection — Ranking the Right Players
Using only pregame information, could the model have flagged when a player was about to underperform — or break out? Bust = player falls well short of their recent average. Spike = player exceeds it. AUC 0.50 = coin flip; 0.70+ = a real, exploitable edge. This is where the model adds its most decision-relevant value.
Stat
N
Spikes
Busts
Spike AUC
Bust AUC
Verdict
Ranking Power — Spearman vs Recency
Does the model put players in the right order within their position each week? If you needed to rank your top 10 wide receivers for Sunday, would our ordering match reality? Pooled Spearman correlation across weeks 9-18. Recency = trailing-3-game average (the floor we must beat).
Position
Weeks
Spearman
vs Recency
Prediction vs Reality — Week-by-Week
Each dot is a player-week. X = our projection; Y = what actually happened. The diagonal line is a perfect prediction.
The Board — Predictions, Actuals & Why
Click any player to see the features that drove their projection. These are the same inputs a coach or analyst would evaluate: snap share, target share, role, trailing form, opponent.
Production Score is a TD-free composite (yards + receptions, position-weighted) used for ranking. It strips out touchdown randomness to isolate volume prediction. Click any player to expand the feature breakdown.
Full Data Explorer — Every Prediction, Every Player
All 5,227 player-week predictions from the 2025 holdout. Filter by position and week, search by player name, sort by any column. The error column shows how far off the projection was. This is the raw data behind every number on this page.