Loading…
Reading the draws and working out what chance predicts.
Crowd Calibration Lab
Lower-tier winner counts estimate ticket exposure. The first 70% of usable dates calibrate the popularity score; the final 30% are kept sealed until it is compared with a no-preference baseline.
Cash Pop1/15 era · since Apr 20220 draws
The selected evidence does not contain at least two parseable exposure dates, so no holdout score is reported.
Chronological Poisson holdoutpoisson-chronological-holdout-v170% train; 30% untouched test; baseline fitted on train only
For a lower tier with winners and per-ticket probability , estimated sales are ; several tiers are combined by a winner-weighted median. Jackpot winners then follow
Published research says birthdays and cultural favourites crowd tickets. Calibration asks the harder local question: how strongly, in this game and era? Missing sales are not a licence to guess, so the lab reports its usable denominator and is allowed to return no fit.
where is exposure, jackpot probability, the crowd multiplier, and the fitted scale.
Dates are ordered before fitting. The first 70% estimate ; the final 30% are never used during training. On those later dates, Poisson deviance compares the popularity model with a baseline that sets . A positive deviance improvement is evidence that the richer model predicts unseen sharing better—not merely that it can explain its training data.