How it works
Method & accuracy
Wal-Bel is a trained rating model, not a chatbot. It solves team strength from net EPA per play by ridge-regularised least squares, weighted toward recent games. The same inputs always produce the same ratings.
How accurate, honestly
walk-forward on 2025 — each week predicted before it was played| Measure | Wal-Bel | Vegas |
|---|---|---|
| Mean absolute error | 10.63 | 9.72 |
| Straight-up winners | 62.5% | 65.4% |
| Against the spread | 48.7% | — |
Vegas is the spread, so it cannot bet against itself. For reference, over 2015–2025 always backing the home team went 48.80%, always backing the favourite 48.58%. The model’s 48.7% is indistinguishable from either. That is market looks like from outside, and why nothing here is a betting edge.
Fitting ratings using 2025 itself — cheating outright — reaches only 10.16, and splitting pass and rush on both sides reaches 10.10. No team-strength model of this family reaches the market at 9.72. The gap is game-specific information: injuries, weather, and money. Rest, travel and quarterback changes were each tested and none helped.
What the data covers
Play-by-play 2015–2025. Next Gen Stats from 2016. PFR charting — pressure, coverage, contact — from 2018. There is no player tracking, no scheme labelling, and no proprietary grading of any kind.