The ScratchOffsNY Smart Score is a versioned model. Every significant change is documented here with date, headline rationale, and a one-paragraph explanation. Older versions are not retroactively rerun — historical rankings reflect the model that was live at the time.
Second mining pass over the 88-day realized-outcome panel found that games with exactly one top prize remaining return 3-4 cents less per dollar in every price tier, and the final top prize is rarely claimed (0.24% per prize-day claim hazard vs 0.62% when 4 or more remain). The ranking now applies a -5 adjustment when only one top prize is left. Walk-forward backtest: realized 28-day per-day rank correlation +0.53 to +0.54, top-5 realized payout +0.5pp.
Validated the model against 88 days of realized window payouts (actual prize value claimed per dollar spent). The daily ranking now blends 30% of the day’s payout-rate percentile plus a price-tier lifecycle decay adjustment ($1-$3 games penalized past 60% sold, $10+ games credited late-life). Backtest: realized 28-day rank correlation +0.31 to +0.50; top-5 realized payout 75.4% to 76-78%.
Replaced hand-tuned factor weights with a weekly walk-forward optimizer that fits weights to realized prize-claim outcomes. Brought rank Spearman correlation from 0.71 to 0.78 in backtest.
Added a per-retailer vending-machine eligibility signal based on the official NY State Gaming Commission equipment list. Stores with self-service vending received a small positive composite bump (+1.2 average).
Replaced naive remaining-prizes-per-pack estimator with a Cox proportional-hazards model fit on six years of historical top-prize claim timing. Improves prize-availability density factor accuracy.
Added a hidden-Markov state classifier over 30-day claim volume to detect regime shifts (e.g. holiday spikes). Smart Score now down-weights extreme momentum during anomalous regimes.
Pruned 4 collinear factors with low mutual-information vs. realized outcomes. No accuracy loss, smaller model surface, faster daily refresh.
Introduced multiplicative interaction factors (payout × freshness, velocity × value, top-prize × density). Captures cases where multiple weak signals compound.
Added an estimator for in-flight pack physics (how prizes cluster across packs of 60 tickets). Improves the Density factor for games where pack-level data is observable.
Statistical test for under-reporting periods. Late-arrival claim data is now retroactively folded into the historical record, fixing a slow-bleed bias in older games.
First version of the Cox hazard preview. Replaced the linear top-prize depletion estimator. Top-prize hazard now drives the Density factor.
When citing a Smart Score for a specific game on a specific date, the model version that was live can be inferred from this changelog (latest entry whose date is on or before the cited date). For academic / journalism use, please request the exact model + factor weights snapshot via the contact form on /about/team.
Data sourced from nylottery.ny.gov and NY Open Data. Independent, non-commercial project. Play responsibly — NY Problem Gambling · 1-877-8-HOPENY.