Skip to content

Findings / Effective Sample Size

How independent is our data, really?

mlb: 78,986 residual rows carry the independent information of at most ~227 games.

ESS from the already-committed lag-1 within-game residual autocorrelation (rho); because the game outcome is terminal-constant the probability path is smooth, so consecutive rows are near-duplicates rather than independent draws. The distinct-game count is the conservative anchor and the AR(1) figure is an approximation, so the reported ess_anchor takes the minimum of the two.

CorpusRowsDistinct gamesResidual autocorr (rho)Effective sample (AR1)Honest anchorCI must widen by
MLB78,9862270.9809market 0.979976122718.7xAR1 estimate 10.2x
Soccer (Intl)9,003510.9624market 0.94891725113.3xAR1 estimate 7.2x

How to read this — rows within one game are near-duplicates because the outcome is fixed and the win-probability path is smooth, so 78,986 MLB rows carry the independent information of at most ~227 games. Every confidence interval on a within-game analysis must therefore be widened by the factor in the last column.

Formula

ESS_AR1 = n_rows * (1 - rho) / (1 + rho); CI inflation = sqrt(n_rows / ESS)

AR(1) is a first-order approximation of a smooth terminal-constant path, not a full independence model; the distinct-game count is the conservative anchor. DESCRIPTIVE_ONLY -- this deflates our OWN sample sizes and makes no edge or ROI claim.

This exhibit deflates our own numbers on purpose — no edge or ROI is claimed anywhere here, only the honest count of what was actually measured.