NERVANALYTICAThe Foolish BaileyGitHub ↗

What actually happens to offense in October

Postseason offense really does collapse: about 40 points of , 1.26 runs per game, across 26 postseasons. The question is why, and whether any regular-season statistic tells you which offense will beat the collapse.

The decomposition

-40Total Octoberdrop-26Pitcher qualityfaced-14Unexplained “chill”

wOBA points relative to the regular season, 70,057 postseason plate appearances, 2000–2025. Two-thirds of the famous October offense collapse is just the pitchers being faced.

Applying the regular-season model to the pitchers each club actually faced explains 66% of the drop. October removes fifth starters and mop-up relievers from the sample; what remains is aces on full rest. Most of the postseason pitching advantage is a fact about roster composition.

The postseason also shifts toward home-run scoring (37.9% of regular-season runs score on homers vs 41.7% in October). The premise of the “HR-reliant teams are built for October” argument is therefore true. Its conclusion is not: across 252 team-postseasons, home-run-light offenses show no opponent-adjusted underperformance (t ≈ 1.0), and home-run share adds nothing to series prediction (p = 0.79).

Team add nothing to series-winner prediction either (likelihood-ratio p = 0.63 over 229 series).

One unresolved result

One pre-specified test came back positive: regressing postseason performance on each team’s regular-season slope times the quality jump it faced gives a of 0.92 ± 0.75 (permutation p = 0.020). The estimate stays between 0.75 and 1.01 across era subsamples, 2020 exclusion, and the leak checks.

Why this is reported as an anomaly, not a finding
The power problem
Twenty-six postseasons can barely detect even full 1:1 transfer (minimum detectable coefficient at 80% power: 1.08). Any significant estimate from this design is fragile by construction.
The consistency problem
Measured reliability says the slope estimates are ~88% noise, so classical attenuation predicts a coefficient near 0.12 even if the trait fully transferred. A coefficient near 1 from a mostly-noise predictor does not fit any standard measurement model.
  • The estimate survives leak controls, era splits, 2020 exclusion, and a permutation test.
  • It contradicts the measured reliability, sits below the design's detection floor, and never appears in series outcomes.
  • Verdict deferred to the committed out-of-sample test: see the pre-registration.

Recent team-postseasons, expected vs actual

SeasonTeamPARegular wOBAExpected Oct wOBAActual Oct wOBAvs expected
2025TOR737.332.292.354+62
2025LAD672.333.299.311+13
2025SEA462.324.289.307+18
2025DET320.319.279.271-8
2025MIL311.325.297.267-30
2025CHC272.327.300.296-5
2025NYY258.339.307.302-5
2025PHI155.331.287.287+0
2025BOS107.325.302.253-49
2025SD103.314.297.244-53
2025CLE98.294.296.253-43
2025CIN79.311.284.256-28
2024LAD617.339.291.333+42
2024NYM577.323.292.317+25
2024NYY545.333.296.330+34
2024CLE380.308.293.298+5
2024DET262.301.290.294+4
2024SD244.325.284.299+15
2024KC216.309.288.263-25
2024PHI149.327.303.277-26
2024ATL145.316.293.288-5
2024MIL106.321.314.329+16
2024BAL71.326.311.214-96
2024HOU70.323.302.229-73
2023TEX672.341.305.340+35
2023ARI637.319.307.319+11
2023PHI470.331.309.348+39
2023HOU423.333.307.335+28
2023MIN216.329.296.305+9
2023ATL141.361.302.238-64
2023BAL113.322.321.301-20
2023LAD105.343.305.229-76
2023MIL78.311.308.345+37
2023TOR72.326.293.250-43
2023TB67.337.304.207-97
2023MIA64.314.295.206-89
How to read this table
What this data is
Expected October wOBA is what an average-slope version of the same offense would hit against the exact pitchers this team faced. The last column is over/under-performance in wOBA points. Note how large it runs in a few hundred plate appearances.
  • Expected October wOBA tracks the pitchers actually faced.
  • A few hundred October plate appearances land mostly at random around that expectation; no measured roster trait shifts them.