Prove it
Backtest this angle
No data
No backtest run yet
Results appear here — pick an angle above, tune it, then run it.
Edge · Bench
Build your own betting angle
Pick a starting point, tune it in plain language, then put it through a real backtest against graded history.
1. Start with an angle
Every angle runs on the same real history — pick the one closest to how you already think, or start from nothing.
Follow the games everyone's already watching — rivalries and marquee windows.
- · Leans toward divisional rivalry games
- · Leans toward a specific day of the week when you pick one
- · Ignores line movement and model math entirely
2. Tune it
Every control below maps straight onto real games and lines — nothing here is simulated.
3. Prove it
Compare two angles
Two angles that graded the exact same games, side by side — ATS record only.
Model A
50.5%
2931-2877-152 · n=5960
graded through Aug 29, 8:21 PM ET
Model B
48.5%
2819-2989-152 · n=5960
graded through Aug 1, 8:02 PM ET
How published models rank
Every published model with a graded spread record, shown with its sample size. Models under 30 graded games are marked early and not ranked.
Spread leaderboard
Every model with a graded spread record
58 models, good and bad. Ranked by record, then closing line value, then sample. A model under 30 decided games is listed as an early sample with no rank.
All live pregame grades through Oct 18, 2026
Sort
| # | Model | Record | Record % | CLV, pts | Training calibration gap, pts | n |
|---|---|---|---|---|---|---|
| 1 | NFL Spread (Adaptive spline regression) · 58 graded picks | 38–20 | 65.5% | -0.03 | — | 58 |
| 2 | NFL Spread (Adaptive spline regression) · 91 graded picks | 43–27–21 | 61.4% | +0.59 | — | 91 |
| 3 | NFL Elo Power Ratings Spread (Gradient boosted trees) · 72 graded picks | 43–29 | 59.7% | +0.01 | — | 72 |
| 4 | Momentum Spread Model (Elastic net regression) · 91 graded picks · Model 2 | 44–34–13 | 56.4% | +0.24 | — | 91 |
| 5 | NFL Spread (Support vector regression) · 91 graded picks | 40–31–20 | 56.3% | +0.52 | — | 91 |
| 6 | Full Feature Spread (Stacking Ensemble) · 71 graded picks | 39–32 | 54.9% | +0.01 | — | 71 |
| 7 | Momentum Spread Model (Gradient boosted trees) | 40–33–18 | 54.8% | +0.24 | — | 91 |
| 8 | Full Feature Spread (Stacking Ensemble) · 91 graded picks · Model 2 | 36–30–25 | 54.5% | +0.69 | — | 91 |
| 9 | NFL Spread (Randomized trees) · 91 graded picks | 36–30–25 | 54.5% | +0.28 | — | 91 |
| 10 | NFL Spread (Nearest neighbors) · 70 graded picks | 38–32 | 54.3% | -0.04 | — | 70 |
| 11 | NFL GNN Matchup · 72 graded picks | 39–33 | 54.2% | +0.08 | — | 72 |
| 12 | NFL ATS Gradient boosted trees (Custom Objective) · 72 graded picks | 39–33 | 54.2% | -0.01 | — | 72 |
Record % is the share of decided picks that won against the spread; pushes are not counted. CLV is how far the closing line moved our way, in points. The calibration gap is the average distance between the win chance a model gave in its training run and the rate that won; a smaller gap is better, and a dash means no training run is on file.
How the engine works
- Every angle you tune runs through the same real backtest engine — no separate math for different personas.
- It only scores an angle on games outside what it learned from, tested one season at a time.
- A 62% call should actually win about 62% of the time — we check that, not just whether a headline number looks good.
- "Rest," "weather," and the other controls are read straight from the real schedule and lines — nothing is simulated.
- An angle only counts as a real edge once it beats a coin-flip baseline over a real number of games.
- Small samples get labeled small — we would rather say "too early to tell" than oversell a lucky streak.