epa-bidirectional-spread
NFL · custom
epa-bidirectional-spread-81e29bbe Track record is temporarily unavailable.
Closing-line record unavailable.
Logistic regression on 8 EPA-derived features (off/def EPA/play, pass EPA, rush EPA, success rate, explosive rate, RZ TD rate, 3rd-down conversion) vs spread_line, trained walk-forward and betting both home AND away based on model direction. Rich nflverse PBP features (2009-2024, 4-week rolling, leakage-free). Key result: bidirectional model achieves 0.5161 ATS across 3563 graded games (2010-2024 walk-forward), vs schedule-only baseline of 0.5054. Confidence-filtered at p>=0.55 achieves 0.5264 (n=870) — crosses 0.524 break-even in historical average but is volatile per-season. Does NOT clear 0.524 reliably in the most recent 3 seasons (2022-2024 mix: 0.561, 0.459, 0.476 unfiltered).
⑂ Remix this model
Turn knobs to put your own spin on epa-bidirectional-spread. Every change re-runs the backtest over real graded nfl games.
Backtest record
—
0-0 on 0 graded games
ATS win rate vs the original
—
Not enough graded games yet — 0 of the 10 this comparison needs.
Lens
Balanced
Every graded game.
See the full run
You're seeing the headline record, the last 0 graded games and the sparkline. Unlock every graded game, the per-slice record, CLV proof and your remix's upcoming picks with Pro — just $1.00 first week.
Draft tracking
No future draft picks are declared in this model's latest version.
ROI equity curve
No settled per-bet backtest data is available for this model yet.
Tune it
Pick a dial to see how tuning this model changes its win rate on the same graded games.
Model calibration
10 probability bins from settled public-model picks. The diagonal is perfect calibration.
No settled public-model confidence results match these filters yet.
CLV history is temporarily unavailable.
Inputs
No feature slugs declared on this card.
Assumptions
- Logistic regression on 8 EPA-derived features (off/def EPA/play, pass EPA, rush EPA, success rate, explosive rate, RZ TD rate, 3rd-down conversion) vs spread_line, trained walk-forward and betting both home AND away based on model direction. Rich nflverse PBP features (2009-2024, 4-week rolling, leakage-free). Key result: bidirectional model achieves 0.5161 ATS across 3563 graded games (2010-2024 walk-forward), vs schedule-only baseline of 0.5054. Confidence-filtered at p>=0.55 achieves 0.5264 (n=870) — crosses 0.524 break-even in historical average but is volatile per-season. Does NOT clear 0.524 reliably in the most recent 3 seasons (2022-2024 mix: 0.561, 0.459, 0.476 unfiltered).
Known limitations
- Auto-generated by nflbets Lab from grind ledger evidence; inspect benchmark artifacts before making supported thesis claims.
Version history
- v1.0.0 Jul 5, 2026
Lineage
No ancestors or forks. This model stands alone — yet.
Discussion
Loading thread…
Sign in to join the discussion.