Model Health
Grading the model on process, not results. The best leading indicator is whether our flagged picks beat the closing line.
How to read this: “Beat close” is the share of flagged games where the market moved toward our pick after the opener — i.e. we'd have gotten a better number than the close. Above 50% over a few hundred recs is the signal that the model is finding real edges. It converges far faster than win/loss, which is mostly variance.
By score tier
Higher-scored picks should beat the close more often. If they don't, the score→edge link needs work.
| Tier | Recs | Beat close | Avg CLV |
|---|---|---|---|
| Strong (75+) | 0 | — | — |
| Lean (65-74) | 0 | — | — |
| Marginal (55-64) | 1 | 100.0% | +0.70 |
Recently graded recommendations
| Game | Pick | Open | Close | CLV |
|---|---|---|---|---|
Golden State Valkyries @ Atlanta Dream WNBA · score 58 | Atlanta Dream | -3.8 | -4.5 | +0.7 |
Calibration
When the model says 65%, do those picks actually win ~65%? “Won” = the favorite won outright. “Suggested” is the data-driven probability (shrunk toward the current value on small samples); adopt it once each tier has ~100+ games.
No graded outcomes yet. Results are captured automatically (free ESPN feed) as flagged games finish — the calibration table fills in from there.