Six NFL seasons of walk-forward testing, never trained on a week it had not already lived through: 7.3% more accurate than the strongest simple predictor of fantasy scoring, measured on walk-forward backtest across the 2020 to 2025 seasons. Every figure is published with its sample size, its measurement date and the document it came from.
Loading the detailed accuracy record.
Six NFL seasons of walk-forward testing, never trained on a week it had not already lived through: 7.3% more accurate than the strongest simple predictor of fantasy scoring, measured on walk-forward backtest across the 2020 to 2025 seasons. Every figure is published with its sample size, its measurement date and the document it came from.
Measured 2026-07-31 on walk-forward backtest across the 2020 to 2025 seasons. Source: engine/eval_history.jsonl (fan081-no-injury-background).
Fanademy error
4.2850
Strongest naive error
4.6232
Improvement
7.3%
Mean absolute error in points-per-reception (PPR) fantasy points per player-week. Lower is better. The baseline is a trailing four-game average shrunk toward the position median—the strongest simple predictor tested, not the most flattering comparison.
The underlying interval is not calibrated well enough to label. A stated floor would be beaten far more often than its name implies, and a points range built by adding per-stat percentiles is invalid because percentiles do not add. Rather than show a number we cannot stand behind, we show the projection alone.
Every model change is gated on a full walk-forward backtest and only ships if it beats the incumbent. Each gated release appends a row here. The log starts at the first gated release, so it is short by design and grows one row at a time.
The per-release log is available to members. Create an account to read it.