Model performance
How the models are tracking — records, CLV, and the +EV subset. CLV is the honest long-run judge.
Winners
—
Totals
6-10
38%
Spread
8-8
50%
Avg CLV
-0.4%
beat the close
Shop gain
+0.26%
best vs avg book (ML)
By model
| Model | Moneyline | ML CLV | Totals | Tot CLV | Run line | RL CLV |
|---|---|---|---|---|---|---|
| nfl_elo | 12-4 · 75% | -0.4% | — | — | 8-8 · 50% | -0.8% |
| nfl_totals | 0-0 | — | 6-10 · 37.5% | 0.7% | — | — |
nfl_elo = Elo — spread + ML · nfl_totals = weather totals. CLV > 0 = beating the close (all in win-prob).
+EV picks — flagged plays only
the subset the betting card flags| Model | +EV picks | Record | Win% | Avg odds | P/L (u) | ROI | CLV |
|---|---|---|---|---|---|---|---|
| nfl_elo | 11 | 5-6 | 45% | -101 | -0.90u | -8.2% | -0.3% |
| nfl_totals | 13 | 4-9 | 31% | +101 | -4.82u | -37.1% | 0.9% |
Only the picks that cleared the +EV threshold. P/L is flat 1-unit stakes; avg odds is the mean price of those picks. Record, ROI, and P/L are outcome-based and noisy at this sample; CLV is the honest judge — does the model beat the close on the games it actually flags. Needs hundreds of picks to trust.