Our Mission

Every day, millions of dollars flow through New York’s scratch-off lottery — and most players pick tickets based on gut feeling, cover art, or whatever’s at eye level behind the counter.

We believe players deserve better. ScratchOffsNY analyzes every active game, every retailer, and every prize claim to surface the games that give you the best mathematical shot. We don’t sell tickets. We don’t take a cut. We just show you the data.

What We Track

How Our Rankings Work

Smart Score — Our proprietary ranking system that evaluates every scratch-off game on multiple dimensions. It goes far beyond simple payout rate to consider factors like where a game is in its lifecycle, how the odds have shifted since launch, and overall risk-reward balance. The result is a single 0–100 score that tells you which games are worth your money right now.

Payout Rate — The percentage of ticket revenue returned to players as prizes. An 85% payout means you lose about 15 cents per dollar on average. Higher is better, but payout rate alone doesn’t tell the full story — that’s why Smart Score exists.

Overall Odds — The lottery’s published chance of winning any prize on a single ticket. Sounds good on paper, but most “wins” are break-even or less. We focus on odds of winning more than you paid.

% Remaining — How much of the total prize pool is still unclaimed. A game with 40% of prizes remaining but only 30% of tickets unsold is mathematically better than one with 60% remaining and 70% unsold.

Key Insight — Data shows that game selection can improve your expected return by 30–47% compared to picking at random. Choosing the right store adds another edge on top of that.

Analysis Tools

Monte Carlo Simulations — We run 10,000 randomized trials per game to show what you’ll actually experience buying 10 tickets, not just the theoretical average. This reveals variance — the real-world difference between a game that pays back steadily and one that’s all-or-nothing.

Live Odds Tracking — As tickets sell and prizes get claimed, the true odds shift. We track these changes daily and show you whether each prize tier is getting easier or harder to hit compared to the original game launch.

Store Intelligence — Not all retailers are equal. We analyze sales volume, activation patterns, and delivery schedules for every licensed retailer in New York to help you find stores with optimal inventory turnover.

Historical Trends — Track how any game’s value has changed over time. See whether your favorite game is improving or declining, and make buying decisions based on trajectory, not just a snapshot.

Our Data Sources

Everything on ScratchOffsNY is built on official, publicly available data:

• NY Lottery — Live game data, prize structures, and remaining prizes via nylottery.ny.gov
• NY Open Data — Retailer locations, weekly sales data, and prize claim records via data.ny.gov
• US Census Bureau — Demographic and income data for geographic analysis
• Player Community — Verified win/loss reports submitted by real players on our platform

Responsible Play

Let’s be clear: the house always wins in the long run. Even the highest-scoring games on our platform have a negative expected return. No strategy eliminates the house edge — but playing smart means losing less and enjoying more.

We encourage setting a budget and treating scratch-offs as entertainment, not investment. If you spend $20/week on the best-ranked game instead of a random one, data shows your expected loss drops from about $6–$7 down to $3–$4. That’s real savings over a year.

If you or someone you know needs help with problem gambling, call 1-877-8-HOPENY or text HOPENY (467369).

Data Sources & Methodology

ScratchOffsNY is built entirely on official public data. We don’t scrape, guess, or crowdsource our numbers.

Contact

Questions, feedback, or partnership inquiries? Reach us at win@scratchoffsny.com

Data Sources & Methodology

ScratchOffsNY is built entirely on official public data. We don’t scrape, guess, or crowdsource our numbers.

Game Data: Prize structures, remaining prizes, and payout rates come directly from nylottery.ny.gov and are refreshed daily via automated pipeline.

Store Data: Retailer sales, payouts, and delivery patterns come from NY Open Data (Socrata API) — the same dataset the state publishes for transparency. Updated weekly.

Smart Score Algorithm: A weighted 0–100 composite of 69 factors across value, edge, prize density, playability, momentum, confidence, upside, heat, freshness, volatility, efficiency, bankroll resilience, timing, depth, acceleration, burst probability, competition, liquidity, claim lag, nonlinear pair interactions, macro demand, seasonality, geography, store concentration, event-study drift, top-prize and $600+ field reachability, mid-tier odds lift, and distribution-timing signals (pull risk, weekend-skew demand, activation volatility, 7-day vs 14-day burn regime, and 7-day claim intensity). Factor weights are continuously optimized via autoresearch — automated experiments that test thousands of weight combinations against historical outcomes.

Store Smart Score: Ranks retailers on payout, volume, consistency, momentum, freshness, inventory activity, statistical significance, hot-streak regime, pack turnover, region context, reporting quality, and cashing-hub stabilization. The public score preserves raw payout as an annotation but uses a buy-quality payout factor so stores are not over-ranked merely because customers cash prizes there.

Score Versioning & Cashing Stabilization: Every store record includes a score_version and factor_set_hash so downstream caches can detect when the factor set changes. The current store model promotes county, region, competition, segment, paycheck-cycle and day-of-week priors into weighted factors, adds zero-new-data stability factors, and adds cashing_adjusted_payout plus cashing_stability to separate true buy-quality from prize-cashing hub behavior in NY Open Data.

Big-Win Recency & Exposure-Adjusted Claims: Documented prize claims at each retailer are scored two ways. Big-claim recency applies a half-life decay by prize size — a $5M win 18 months ago beats a $10K win last week, but a $5M win five years ago has fully decayed. Exposure-adjusted residual compares each store's claim count to the expected count for similar stores (matched on volume tier, retailer segment, and county). Stores claiming more winners than the cohort baseline get a positive residual; stores below baseline get a negative one. Both ship as candidate factors today.

Sparse-Data Gate: Stores with fewer than 7 active reporting days or under $5,000 in total settlements get an automatic limited data badge, and we display the lower confidence bound rather than the point estimate so a fluke spike on three days of data can’t catapult an unknown store to the top of the rankings.

Single-Month Damage Cap (Huber): The consistency factor uses a Huber-style cap on monthly residuals. A single catastrophic month (feed gap, store-level outage, holiday freeze) is clipped to a fixed maximum deviation before entering the stability stddev, so one bad month cannot crater an otherwise three-year-strong retailer.

Score-Change Rate Limiter: Day-over-day movement of a store’s published score is capped at ±10 points unless a real event justifies the jump — a documented regime shift (change-point), a fresh big claim, or a recent news win. This prevents jitter from autoresearch retunes from confusing returning users without hiding genuine breaking events.

Claim Concentration (One-Hit-Wonder): Each store gets a Gini-based diversification score over its recent prize claims. A retailer with one historical $1M jackpot and no other activity scores lower than a retailer with a steady mix of $5K–$50K wins at similar volume; ships as a candidate factor.

Robustness Guards: Two observability checks watch the rankings. A disparate-impact monitor tracks the top-100 composition by region/segment/county over a 14-day baseline and alerts when a slice moves more than 2σ. A Simpson’s-paradox guard re-ranks each top-100 store within three independent factor slices (payout-driven, stability-driven, opportunity-driven) and flags stores that surface high only because of a single factor.

Market-Normalized Momentum: Statewide daily settles are aggregated to a market shock index, then each store’s recent-vs-baseline ratio is divided by the market’s. Stores genuinely outperforming the macro trend during a jackpot week, holiday burst, or storm-suppressed window get a positive normalized-momentum signal; stores merely riding the wave do not.

Holiday-Week Lift & Restock-Phase Sweet Spot: Each store gets a holiday lift annotation comparing settles on US holiday windows (New Year, Memorial Day, July 4, Labor Day, Thanksgiving / Black Friday, Christmas, New Year’s Eve) to its ordinary baseline. Each store also gets a restock phase score — a bell-curve fit centered shortly after the most recent detected delivery spike, using the store’s own learned cadence. Together they pinpoint “buy after restock, before depletion” timing rather than a flat “recently restocked” flag.

Spike Quality, Depletion & Jackpot-Day Lift: Each detected delivery spike is graded on follow-through (sales 2–5 days after the spike vs the store’s median) and on depletion (sales 6–14 days after the spike vs the pre-spike baseline). Spikes with no follow-through look like accounting noise; spikes followed by collapse look like the pack burned out. Separately, a jackpot-day lift annotation compares settles on Powerball / Mega Millions draw days against ordinary days, surfacing stores whose foot traffic genuinely tracks jackpot cycles. All three ship as candidate annotations.

Segment-Aware Pack Turnover: The pack-cycle benchmark used inside the Turnover factor is no longer one-size-fits-all. We multiply the base pack-life expectation by a (segment × volume tier) factor so a vending machine isn’t penalized for averaging vending-typical turnover, and a high-volume grocery isn’t given undue credit for averaging grocery-typical turnover. Stores are rewarded only when they outpace their realistic peer group.

Factor-Ablation Sensitivity Matrix: Every nightly Smart Score refresh writes a sensitivity report: for each of the ten weighted factors, we recompute the ranking with that factor’s weight zeroed and measure top-50 turnover vs the live ranking. Any factor whose removal causes more than 50% of the top-50 to change identity is flagged as a single point of failure so we can rebalance weights before the model overcommits to one signal.

Score Flavors & Staged Rollout: The published score is the “control” flavor. A “candidate” flavor — including all the shadow factors (exposure-adjusted claim residual, big-claim recency decay, claim concentration, market-normalized momentum, holiday lift, restock-phase score, spike quality, depletion, jackpot-day lift, segment-aware turnover) — is available via the Try candidate score (beta) toggle on the Stores page (sets a 90-day cookie). Operators can also enable a per-tile rollout by adding zip prefixes to autoresearch/score_rollout_zips.json; a configurable holdout share keeps a stable fraction of users inside the rolled-out tile on control for clean measurement. Default behavior for everyone else is unchanged until cutover.

Prediction Engine: A self-improving, horizon-aware gradient-boosted tree model trained on walk-forward historical snapshots. The 1-day model focuses on change signals; 3-day and 7-day horizons add lifecycle, price tier, value, prize velocity, cannibalization, burst, day-of-week, and calibration features. The model validates itself against actual outcomes and auto-adjusts blend weights (Smart Score vs. Booster vs. Momentum) based on pairwise ranking accuracy. Separate from Smart Score rankings — they never cross-contaminate.

Autoresearch System: Three automated optimizers run continuously: (1) Game weight optimizer tests factor weight combinations to maximize rank correlation with future outcomes. (2) Store weight optimizer does the same for retailer rankings. (3) Prediction hyperparameter optimizer tunes the ML model’s trees, learning rate, train window, and blend weights. All results are logged and applied automatically.

Monte Carlo Simulator: We run 10,000 random trials per game to model realistic outcomes for different play budgets. Prize extinction estimates use statistical modeling based on claim velocity and remaining prize counts.

Live Autoresearch Stats

What kind of tips are useful?

Rare Finds — Spotted an ended game still on the rack? Tell us the store and game number.

Store Intel — Know a store that gets fresh inventory often, or one with dusty old tickets?

Strategies — Found a pattern or approach that works? Share it with the community.

Or just email us directly: win@scratchoffsny.com