How Gridiron Edge Works

The Problem With Standard NFL Analytics

Most publicly available NFL analytics stop at EPA — Expected Points Added. EPA is a useful play-level metric, but it has a structural flaw: it treats a 10-yard completion against the 2024 San Francisco 49ers defense the same as a 10-yard completion against the 2017 Cleveland Browns. The opponent doesn't exist in the math.

That's fine for descriptive purposes. It fails for prediction.

A team that ran up gaudy EPA numbers against a weak schedule looks identical to a team that ground out the same numbers against a murderers' row. When those two teams meet, standard EPA-based models give you the wrong answer more often than they should. Gridiron Edge was built specifically to close that gap.

Defense-Adjusted EPA (DAEPA)

The core metric is DAEPA — Defense-Adjusted Expected Points Added.

The adjustment uses an iterative opponent-weighting process conceptually similar to PageRank: a play's value is scaled by the quality of the opponent it was executed against, and opponent quality is itself a function of the quality of theiropponents. The system runs until convergence — typically around a dozen iterations — producing ratings where each team's offensive and defensive efficiency is expressed in a common unit that accounts for the full strength-of-schedule context.

The result is three per-team DAEPA components:

• Offensive DAEPA — efficiency generating EPA, adjusted for defensive quality faced
• Defensive DAEPA — efficiency suppressing EPA, adjusted for offensive quality faced
• Special Teams DAEPA — adjusted field position and scoring value from the kicking game

These three components are converted to the points-per-game scale and summed into PWR— each team's expected margin in points vs. an average team on a neutral field. A team at PWR +3.5 projects to win by about 3.5 points on a neutral field against a PWR 0 opponent. PWR translates into a plain-English tier (Elite, Contender, Playoff, Average, Below Avg, Rebuild) so the number reads quickly without losing the underlying precision.

Game Context Weighting

Raw EPA includes plays that should carry little analytical weight — a prevent defense giving up a meaningless 15-yard gain with 90 seconds left in a 28-point blowout tells you almost nothing about either team's real capability.

Gridiron Edge applies a per-play context weight before aggregating DAEPA. Three conditions are handled:

Garbage time discounting: Plays in late-game blowouts receive reduced weight using a smooth function driven by win probability and score differential — not a binary cutoff. A 3-point game in the fourth quarter is barely discounted; a 28-point game with two minutes left is heavily discounted.

Clock-kill flagging: Run plays by a leading team in the fourth quarter designed to kill time rather than gain yards are identified and excluded from offensive efficiency calculations.

Prevent defense flagging: Pass plays against a defense in prevent coverage — trailing by multiple scores late — are flagged and discounted. Prevent defense statistics are notoriously misleading as indicators of pass defense quality.

Recency Weighting and Season Blending

Gridiron Edge applies exponential recency weighting to all play-level data, with a half-life tuned to NFL roster stability patterns. Recent games carry more weight; earlier games decay.

Cross-season blending uses a per-component Bayesian shrinkage framework on the points-per-game scale. Heading into Week 1, a team's prior shrinks the prior-season Off PR, Def PR, and ST PR toward zero by empirically-measured α coefficients (see the next section). Coaching changes and roster overrides are layered in as additive points-scale deltas — not as multiplicative scale factors, which is how earlier versions of this model accidentally produced no-op roster adjustments. As the new season's data accumulates, the posterior shifts toward current performance; the prior is effectively washed out by midseason.

Empirically-Calibrated Per-Component Shrinkage

The α (year-over-year shrinkage) and β (component win-weight) coefficients quoted above are measured, not assumed. The calibration ran in May 2026 over the full 1999–2025 play-by-play dataset.

Year-over-year stability (α coefficients).Measured as Pearson R between each team's prior-season component rating and the same franchise's next-season rating, across 829 team-pair observations from 1999 through 2025.

The ST finding is counter-intuitive: blocked kicks and return touchdowns feel like luck. They are, on the play level. But over a 17-game season, ST is dominated by kicker FG accuracy and punter net yardage — both highly stable skills tied to specific players. Defense, by contrast, depends on edge-rusher health, turnover luck, scheme chemistry, and opponent quality — all of which churn year-to-year.

Component win-weights (β coefficients). Fit by OLS regression on 745 team-seasons from 2000–2024: season_wins ~ β₀ + β_off · off_pr + β_def · def_pr + β_st · st_pr. R² = 0.78 — the three components explain 78% of season-win variance; the remaining 22% is schedule strength, injury luck, turnover variance, and tactical matchups.

The β values for Off / Def / ST sit within ~25% of each other — well inside the standard-error envelope. This is why PWR sums the three components directly with equal weights rather than applying a fancy β-weighted formula; the ranking changes from β-weighting would be smaller than the noise in the α shrinkage.

Why this matters for the rankings. The previous version of these rankings summed components on a z-score×100 scale, which made special teams artificially equal in weight to offense or defense. On the native points-per-game scale, offense and defense range about ±10 points league-wide; special teams ranges only about ±1.4 points. Summing on the points scale produces a defensible total where the JAX-style "ST-strong team accidentally ranks top-5" failure mode goes away structurally.

Roster Adjustment and Injury Impact

The roster adjustment layer quantifies player impact via Player Importance Scores (PIS)— a per-player measure of how much that player's presence or absence shifts a team's expected DAEPA. PIS is derived from the player's historical DAEPA contribution in similar roles, weighted by snap share and positional leverage.

This is particularly important during the off-season: free agent departures, draft picks, and trade additions are all reflected in the team's projected ratings before the first preseason snap. Active injury adjustments at game time use the same PIS framework — when a player is ruled out, their contribution is removed and redistributed to their replacement using historical replacement-level baselines by position.

The Win Model

Win probability prediction uses a hybrid approach based on data availability:

Spread-available games (~96%): The opening line is the single most information-rich predictor of NFL game outcomes. For these games, Gridiron Edge converts the opening spread directly to a win probability using a normal distribution calibrated to NFL margin-of-victory distributions. DAEPA serves as a cross-check and edge-identification tool rather than the primary predictor.

No-spread games (~4%): A logistic regression model trained on DAEPA features generates the win probability for games without an opening line.

The value of DAEPA in the spread regime isn't in replacing the line — it's in identifying where the model disagrees with the market and why. A significant divergence between DAEPA-implied probability and spread-implied probability is the edge signal.

Walk-Forward Validation

Every performance metric published by Gridiron Edge comes from a strict walk-forward backtest: train on all data through season N, predict season N+1, advance one year, repeat. No future data ever influences a past prediction. No hindsight.

The backtest covers 5,934 games from 2003 through 2024 — the full era of modern NFL data with consistent play-by-play tracking.

• 66.5%Win prediction accuracyvs. DVOA benchmark ~66%
• 0.616Log losspenalizes confident wrong predictions
• +0.61ppAverage CLVvs. closing line, 5,733 matched games

What This Means for Prediction Markets

Kalshi, Polymarket, and DraftKings all price NFL outcomes as independent markets. Sometimes they disagree — and those disagreements are signal. Gridiron Edge generates an independent probability estimate from DAEPA and the opening spread, compares it to each market's implied probability, and surfaces the largest gaps.

Edges are categorized as STRONG (≥12pp), HIGH (≥8pp), MODERATE (≥5pp), or LOW(<5pp) based on the magnitude of the model-market divergence. Only edges with directional agreement between the DAEPA signal and the spread signal are elevated to HIGH or STRONG. Cross-platform confirmation — two or more markets agreeing on the direction of mispricing — strengthens conviction further.

Market Coverage

The DAEPA engine feeds edge detection across every NFL market type we can find pricing for:

• Moneyline — head-to-head winner, compared across Kalshi, Polymarket, and DraftKings
• Spread — point-differential prediction vs. the opening and closing lines
• Total — combined scoring over/under
• Player props — passing, rushing, and receiving stat lines (Kalshi + DraftKings)
• Win total futures — seasonal O/U per team, run via Monte Carlo over full schedules
• Division, conference, Super Bowl — playoff bracket probabilities from a season-plus-postseason simulation
• Season awards — MVP, ROY, OPOY, DPOY when Kalshi or Polymarket posts them

Not every market is live year-round. Preseason coverage emphasizes futures and win totals; in-season shifts to weekly games and player props. The backtest number above (66.5%) is the moneyline figure — spread, total, and prop performance is tracked separately as those product lines mature.