Regression Analysis

Regression to the mean is the statistical phenomenon where extreme performances tend to move back toward average over time. In NFL betting, identifying regression candidates creates significant edges.

Why Regression Matters for Betting

• Markets overweight recent performance - Public bettors chase hot teams
• Unsustainable stats inflate/deflate lines - Creates value opportunities
• Sharp bettors exploit regression - Understanding it is essential

These stats have low year-over-year correlation and high variance. Extreme values will likely regress:

These stats have high year-over-year correlation. Extreme values are more likely to persist:

Estimate where a stat will regress based on sample size and league average:

1. Early Season Regression Plays

2. Unlucky Team Buys

3. Point Differential Adjustment

Watch for these signs that a team is due for regression:

The stat-level correlations quoted above come from published play-by-play research. The bundled nflverse game log lets us demonstrate the same force at the team level, where anyone can recompute it. Pairing each franchise's season with its following season, 1999–2025 (822 pairs):

Data: nflverse games.csv bundled with this site (regular season, 1999–2025), computed by the author. Year-N averages computed from the same pairs.

Powerhouses gave back three wins on average; doormats picked up three and a half. Both groups moved toward the middle, not because talent evaporates overnight but because extreme records are part talent, part schedule, part one-score-game luck — and the luck components reset every January. The year-over-year correlations tell the same story in one number each: wins correlate 0.32 season to season, point differential 0.40. Real signal, heavily diluted.

The betting punchline is the third correlation: team cover rate (ATS%) correlates −0.05 year over year — nothing. Teams that covered 60%+ in a season covered 49.8% the next. Whatever regression the market misses in October, it has fully priced by the following September.

The regression estimator above shrinks an observed rate toward the league average in proportion to sample size. Worked in full for the classic case — a team converting 75% of red-zone trips into touchdowns through six games (say 24 of 32 trips), against a league average around 56%:

• A common shrinkage form is: estimate = (successes + K × league rate) / (trips + K), where K is a stabilization constant — for illustration, K = 50 trips.
• Estimate = (24 + 50 × 0.56) / (32 + 50) = (24 + 28) / 82 = 63.4%.

The tool's answer is not “75% is fake” but “your best guess going forward is 63%, not 75%.” The team is probably good in the red zone and probably running hot — both at once. That is what “regression to the mean” means in practice: partial credit, sized by evidence. (Exact formula math on an illustrative example; K varies by stat and must be estimated from data.)

The same logic explains the table above: a 13-3 record is strong evidence of a good team and weak evidence of a 13-win team.

• The stat-level correlations on this page (turnovers ~0.20, EPA ~0.55–0.65, etc.) are approximate values from public play-by-play research — the bundled game log has no play data, so we cannot recompute them here. The team-level table above is what this dataset can prove directly.
• Regression is not a pendulum. Extreme groups move toward the mean on average; individual teams routinely stay extreme (the 2003–2004 Patriots did not regress on schedule).
• Markets regress faster than teams. The −0.05 ATS persistence means the profitable window, if any, is inside a season — by the next year the line has moved.

Keep reading: turnovers & luck, Pythagorean wins, and one-score games — the three main luck reservoirs regression drains.