Does success predict more success?
That is the hot-hand question: whether a recent run of success makes another success more likely.
Competitive Sports: Entropy to the Mean
The longer the winning streak, the worse the performance.
The hot-hand debate asks whether winning continues. This research asks what happens to performance while it does. Across seven seasons of college basketball, the decline exceeded what a same-skill memoryless process produced and remained after accumulated overperformance was included in the same model.
Full 20-page research paper. Shared publicly before journal peer review.
Regression to the mean is the statistical shadow.
The dissolution of sustained order is the phenomenon.
Beyond the hot hand
This starts where the hot-hand debate stops.
For forty years, streak research has focused on whether success predicts more success. This paper asks the inverse question: among teams that keep winning, does the quality of performance deteriorate as the streak gets older?
What happens to performance while the winning continues?
The outcome is held fixed: the team is still winning. The test asks whether the quality of that performance declines as the streak ages.
Featured paper
Entropy to the Mean
Winning Streaks Carry a Rising, Measurable Cost: Evidence from Seven Seasons of College Basketball
Using 69,022 Division I team-game rows across seven seasons, the paper asks whether the same team performs progressively worse relative to expectation as a winning streak ages.
The raw within-team decline is substantial, but the paper does not claim the raw slope. It constructs a same-skill memoryless benchmark through the identical pipeline, measures the mechanical floor, and makes its claim only about what survives.
Some apparent decline appears automatically because of how streaks are selected and compared. The study recreates that effect in a same-skill process with no aging and claims only the extra decline left over.
Left column
Wins already in the streak
Consecutive Division I wins carried into the game
Bars and right column
Points better or worse than expected
Average opponent-adjusted point margin per game
0wins
+0.23pts/game
1–2wins
+0.09pts/game
3–5wins
onset
−0.55pts/game
6–8wins
−1.75pts/game
9–12wins
−1.56pts/game
13+wins
deep streak
−3.10pts/game
How to read it: the first two groups are essentially flat. Performance moves below expectation at 3–5 prior wins and is roughly 3.1 points worse than expected at 13 or more.
Pooled out-of-fold estimates. The depth bins and expected shape were locked in writing and hashed before the held-out season was loaded.
01
The decisive horse race
Streak duration and accumulated overperformance each predict decline when tested alone. When forced to compete in the same model, duration remains strong and accumulated overperformance collapses.
The model compares two explanations: performance falls because the streak is old, or because the team had been playing unusually well. When both explanations compete, streak age still predicts the decline; the earlier overperformance does not.
- Depth, joint model
- −0.2837 · t = −6.8
- Accumulated residual
- −0.0053 · t = −1.2
- Memoryless null center
- +0.0117
What the word means
Entropression
The excess, duration-indexed compression of sustained ordered competitive performance: the progressive adverse deformation of performance channels with accumulated ordered duration, beyond the deformation generated by an appropriate own-construction memoryless process.
Entropression is the extra drop in performance that appears as a winning streak gets older, after subtracting the drop that would show up even if every game were independent and nothing were getting harder over time.
It names an observed form
Entropression describes a measurable pattern: sustained ordered performance compresses as accumulated ordered duration rises.
It does not name the mechanism
The causal generator remains open. The paper does not claim that a regression coefficient is a physical quantity.
Statistics are readouts
The ordered performance state undergoes the compression. Statistical channels register that compression as measurable deformation.
Technical summary
Abstract
The hot hand literature asks whether success predicts success. This paper asks the inverse question: conditional on a team continuing to win, does the quality of its performance decay as the streak lengthens? Using 69,022 Division I versus Division I team-game rows across seven college basketball seasons, with point-in-time KenPom opponent adjustment and no future leakage, we estimate the effect of entering win-streak depth on opponent-adjusted margin residual within team-season fixed effects, with standard errors clustered on 2,306 team-seasons. The raw within-team estimate is -0.3207 points per game of depth (SE 0.0280, t = -11.5); an own-construction memoryless null, built at each team’s own rate and passed through the identical pipeline, reproduces a mechanical floor of -0.2238; the claim of this paper is the excess beyond that floor, -0.0969 per game of depth (0 of 200 null replicates as negative; lower-tail p at or below .005). The raw estimate carries a pre-registered threshold shape: flat through depths 1-2, onset near depths 3-5, deepening to roughly -3 points at depths 13 and beyond. The naive cross-team specification is opposite-signed, a rating-lag artifact the within-team design removes. The effect survives fatigue, recent-form, and opponent-preparation controls. A loss-streak placebo is opposite-signed but largely mechanical under its own null. In the decisive horse race, accumulated overperformance collapses to insignificance while duration survives. A pre-registered, hash-locked one-shot replication on the held-out 2021-22 season passes both frozen criteria (coefficient -0.4169, t = -4.6). A validated tennis instrument returns a concordant phase-tier residual in two universes, and an EPL construction replicates enter-depth compression locally. The duration-burden interpretation is [supported]; the causal mechanism is [open]; two pre-registered composition tests remain open and gate the strongest public version of the claim.
Cross-substrate program
One form, substrate-specific clocks and channels
The research program does not require identical statistics across sports. It asks whether the same law-of-form recurs through measurements appropriate to each competitive substrate.
College basketball
Opponent-adjusted margin compresses with entering win-streak depth within team-season, exceeds a memoryless floor, and survives a held-out replication with one disclosed venue caveat.
Tennis
A validated phase-tier instrument finds concordant late-streak pressure readouts in score-only and break-panel universes after path-erased and structure-matched nulls.
Soccer
An EPL construction replicates enter-depth compression across shot, shots-on-target, and goal channels, with lower-division evidence showing channel morphology rather than a copied basketball ladder.
Method and governance
The claim is defined by the attempts to kill it
Within-team identification
Each team-season is its own control, sharply reducing rating-lag and persistent between-team quality artifacts.
Own-construction nulls
The same streak-finding code, depth variable, and regression are applied to a memoryless same-skill process. Mechanical deformation is measured, not assumed.
Pre-registered replication
The held-out 2021–22 season was analyzed once under a hash-locked specification and passed both frozen criteria.
Recorded failures
Wrong predictions, null results, and retired claims remain in the record. The same discipline that supports the flagship killed weaker internal findings.
Interpretation ceiling
What remains open
The paper measures a footprint. It does not establish a thermodynamic identity or settle the causal mechanism that generates the duration burden.
The principal identification frontier is composition across streak lengths: whether whole-streak population differences alone can generate the observed within-streak ordering.
Open test
Within-episode separation
Does the duration effect survive population matching inside team-season?
Open test
Within-episode order
Can whole-streak population differences alone generate the ordering in matched synthetic worlds?
Open mechanism
Generative account
What converts accumulated ordered duration into adverse performance deformation?
Cite this work
How to cite
Castranova, J. (2026). Entropy to the Mean: Winning Streaks Carry a Rising, Measurable Cost: Evidence from Seven Seasons of College Basketball. Preprint, version 0.6. DOI: 10.5281/zenodo.21576772
@misc{castranova2026entropy,
author = {Castranova, James},
title = {Entropy to the Mean: Winning Streaks Carry a Rising,
Measurable Cost: Evidence from Seven Seasons of
College Basketball},
year = {2026},
month = jul,
note = {Preprint, version 0.6. Not yet peer reviewed.},
doi = {10.5281/zenodo.21576772},
url = {https://entropression.com/}
}Full 20-page research paper. Shared publicly before journal peer review.
About
James Castranova
Independent researcher and founder of DHH Labs. Castranova develops falsifiable instruments for studying sustained ordered competitive performance across sports, with frozen specifications, adversarial review, synthetic validation, and explicit claim-status boundaries.
jcastra@gmail.com