NFL Line Board

Six books, one screen — spreads, totals, moneylines, props & kickoff weather

best available price in row Fair line & edge % (vs Pinnacle, else consensus) green = beats fair red = worse than fair

🧮 Bet calculator (single or parlay)

Stake $ Odds

Bankroll $ Your win %

Comma-separate odds to price a parlay (e.g. -110, +120, +150). Kelly uses YOUR win % estimate (not the book's) and the first odds entered; half-Kelly is the safer stake.

Hit Load live odds to see the board.

ℹ️ How this tab works & what's validated (tap to expand)

🎯 Player usage & opportunity — built from this season's play-by-play (same import as the Model tab, no extra download). Touches/targets by field-position zone, each player's target share and red-zone role within their own team, and "TD debt": each zone's league-wide historical TD rate applied to this player's own touches there, compared to what they've actually scored. Positive debt means their opportunities haven't paid off yet (a regression-to-the-mean signal); negative means they've outscored their expected rate. This intentionally leaves out defensive coverage-shell splits (single-high/two-high) — that needs PFF/Next Gen Stats-tier charting with no free, legitimate source.

⚠️ Receiving Yards, Rushing Yards, and Receptions get a ridge regression "Model projection" (season average, recent-5-game usage, opponent's allowed rate at that position, home/away). Originally fit on 2023 and validated clean against the untouched 2024 season: beat a naive trailing-5-game average by 4.3% (receiving yards), 2.7% (rushing yards), and 2.6% (receptions) — real but modest. Retrained since on 2023-2025 plus this season's confirmed games (same fixed settings from that validation), but this exact coefficient set hasn't itself been re-tested against a fresh unseen season — there isn't one available until 2026 wraps. Anytime TD gets a separate, simpler Poisson estimate from recent scoring rate, shrunk toward a rough league-average rate so a real zero-TD sample doesn't collapse to a hard, misleading 0% — not this regression, not independently validated the same way. Passing stats, targets, and carries have no projection at all — not tested, so nothing's shown rather than guessed. The "Line" defaults to the live book price when you've pulled this game's props (tap "Pull live prop line"), otherwise a rough recent-average estimate — either way it's yours to adjust with +/-.

ℹ️ How the model works & what's validated (tap to expand)

⚠️ Unvalidated as of this refit. This is a ridge regression (not hand-picked), retrained on 2023-2025 nflverse play-by-play now that 2025 is complete. The previous version graded out at -1.2% ROI on the untouched 2025 season — the closest to breakeven of everything tested, but not a proven edge. That number no longer applies to these exact coefficients since 2025 is now inside the training set instead of held out. There's no clean unseen season to test against until 2026 finishes. Treat every signal below as experimental research, not instruction, until this gets validated the same way.

🐛 Bug found and fixed — a feature was removed as a result. The live ATS cover% calculation had its sign backwards: it was reading nflverse's spread_line as "negative = home favored" when nflverse's own data dictionary confirms the opposite (positive = home favored). This made every team's displayed ATS record wrong (verified directly against 2026 game data) and has now been corrected — the ATS column below is accurate. The model briefly included an 11th feature (season-to-date ATS cover% differential) validated at a claimed ~1% RMSE improvement — that validation was built on the same buggy sign. Re-ran the identical 2023→2024 walk-forward check with the corrected sign: result flipped to -0.2% (slightly worse, not better). The claimed improvement was an artifact of the bug, not a real signal, so the feature has been removed — the model is back to the 10 features below.

Predicted spread = a fixed set of coefficients (learned by ridge regression, never hand-adjusted) applied to opponent-adjusted, recency-weighted play-by-play efficiency: EPA/play split by pass and rush (both sides of the ball), success rate, CPOE, pass-rate-over-expected, pressure rate, rest-days differential, and live wind — plus one flat, separately-measured adjustment: a small penalty (0.56 pts, from a real but modest effect measured across 2022-2025) when the confirmed starter (Sleeper's depth chart) has thrown ~0 passes for that team this season, tagged 🩹 in the QB line below. The raw output is then anchored to the closing line as a Bayesian prior (25% model / 75% market — the market is the far more efficient estimate by every backtest run against it) and hard-capped at a 3.0-point disagreement, so "My spread" can never show as an implausible blowout number even when the underlying model is noisy. Source: nflverse play-by-play (nflfastR) + Sleeper for starting QBs.

ℹ️ How the totals (O/U) model works — and why it's not validated (tap to expand)

⚠️ Did not beat the market in backtesting. Tested the same clean walk-forward way as every other model here: trained on 2023, graded on the untouched 2024 season. Predicting the total purely from the market's own posted total line beat every version of this regression, including the raw regression alone. This is genuinely weaker evidence than the spread model (which at least showed a small real improvement from its ATS-cover-% feature) — totals markets appear to be at least as efficient as spread markets against this simple a feature set.

Predicted total = a fixed set of coefficients applied to each team's own season-to-date scoring average and points-allowed average (trailing, games already played only — no leakage), plus live wind. Anchored to the market total the same way the spread model anchors to the closing spread (25% model / 75% market, capped at a 4.0-point disagreement) so it can't show an implausible number — but unlike the spread model, that anchor is doing essentially all the work here; the regression itself isn't shown to add anything on top of the market. Kept visible for context and because the data's already being pulled for other things, not because it's demonstrated to be useful. Treat it as a number to look at, not a signal to bet on.

Scoreboard — model vs. market, accumulated every time you load past results

Model vs market Week

Sort

Tap a game to load it into the Kelly calculator. 🔒 Mortal Lock marks the model's single most confident BET of the week (biggest gap vs. market among this week's BET signals).

Past results — cross-check the model's pick against how the game actually went

Pulls final scores from the odds feed's scores endpoint (last 3 days back — that's the feed's own limit, so a Thursday game can roll off by the following Monday). The "model's pick" here is recomputed with today's team ratings, not the live prediction from before kickoff, since past predictions aren't saved anywhere — treat this as a rough cross-check, not a clean backtest. Graded against your own captured closing line where available; games with no captured line are shown but not graded.

📊 Team ratings — opponent-adjusted EPA, success rate, CPOE, pressure (read-only, import only)

This table is read-only — every number comes from the import button above, nothing is hand-typed, and the formula's coefficients are fixed (fit offline against 2023-2025 results, not adjustable here). Each team's rating is an exponential moving average (recent games weighted more) of their play-by-play efficiency, filtered to exclude garbage-time snaps (win probability between 5% and 95%), then adjusted for the strength of opponents actually faced. Pass and rush EPA are tracked separately on both offense and defense. Starting QB is identified from Sleeper's live depth chart (catches trades/signings/returns before a first game is even played) and shown for reference — the regression found no incremental signal from QB-specific EPA once team-level splits were included, so it's display-only and doesn't affect the prediction. Rest-days differential and live wind (from the Weather tab's forecast) are also in the formula; home-field edge is learned as part of the regression rather than a flat constant. This import pulls the entire current season's play-by-play file (a few MB early in the season, growing to 40-50MB by December) plus the small season schedule file and Sleeper's player list — noticeably heavier than a typical import, especially late in the season. Load live odds first so wind can be pulled from each game's forecast. Early-season weeks are seeded from 2025's final ratings instead of a blank league average — tuned on one season transition and confirmed on a separate held-out one (early-season ROI improved slightly, no downside on the full season); the seed's influence fades on its own as real 2026 games accumulate, and it's a fixed table that needs manually refreshing once 2026 ends.