Browse › Residual Autocorrelation
Analytics module · as of
Within-game errors barely change tick to tick
mlb: median within-game lag-1 residual autocorr = model 0.981 (n=222 games), market 0.980 (n=225); share>=0.9 model 0.95 vs market 0.96. soccer_intl: median within-game lag-1 residual autocorr = model 0.962 (n=51 games), market 0.949 (n=...
confirmednull (a finding)not testabledescriptivepending
Residual Autocorrelation
What it means
Because the outcome is fixed and the probability path is smooth, consecutive in-game rows are nearly identical, so the ~78,986 in-game MLB rows carry far fewer than 78,986 independent observations. This is a guardrail about how to count evidence, not a claim about any team -- the same near-1.0 autocorrelation holds for the market path too.
Caveats & confounds
descriptive_only, no edge/ROI claim; share of games with autocorrelation >= 0.9 is 0.955 for the model. Games with too few rows or flat residuals are skipped.
Method. lag-1 sample autocorrelation of within-game signed residual (prob_t - terminal outcome), per (game_id, side) series ordered by ts