Observed frequency

For each main number, we count filtered draw records containing that number once. Observed percentage = number of those records ÷ total filtered draw records × 100. Because a draw contains several main numbers, percentages across the full main-number table do not add to 100%.

Special balls are counted in their own table and denominator. They are never mixed with main-number frequencies.

Digit games preserve order

NY Numbers and Win 4 records are normalized to three and four digits. Leading zeros are preserved. Frequency is calculated separately for each digit position, so repeated digits are not collapsed. Each midday or evening session is a separate draw record.

Pairs and repeated combinations

For main-number games, each distinct unordered main-number pair is counted once per draw. The special ball is excluded. Observed percentage is the share of filtered draws containing that pair. For digit games, the table instead shows repeated exact combinations, preserving digit order.

Date windows and weekdays

The 90-, 180- and 365-day windows are calendar-day windows ending at the latest draw date returned by the source, not at today’s date. The page states its actual range and latest source draw. Archive windows therefore end at the archive’s last draw. Weekdays are derived from the source’s draw date without shifting it through the visitor’s time zone. Weekday and session filters change the denominator.

Game rules must stay separate

The beta limits Powerball to records on or after October 7, 2015 and Mega Millions to records on or after April 8, 2025. Those boundaries keep the supported number pools consistent with this version’s validation. This is not yet a complete lifetime rule-history engine. Earlier Mega Millions periods are excluded rather than blended into current-pool statistics.

Projected combinations

The Command Center can assemble a research combination from the selected verified sample. The current model considers observed main-number frequency, pair co-occurrence, odd/even composition, high/low composition and the central historical sum range. The same verified sample produces the same research set so results can be audited and compared over time.

A projected combination is a historical-pattern research output, not a forecast of the next winning draw. In a fair random lottery, a valid combination does not gain better mathematical odds merely because its numbers or structure appeared more often historically.

Model versioning and frozen projections

The current combined projected-combination methodology is identified as combined-v1.0.0. A frozen Performance Ledger record stores the model version, game, generated set, selected research window, draw-day filter, verified data range, sample size and creation timestamp. Frozen records are intended to preserve what the model produced before later results are evaluated; changing the methodology requires a new model version rather than rewriting earlier records. When verified data later contains a draw dated after the frozen record’s verified-through date, Ledger v1 evaluates the first eligible subsequent draw and stores the main-number match count and special-ball result without replacing the original projection. Ledger v1 is browser-local while account-based persistence is under development.

Forward performance reporting

Evaluated frozen projections are aggregated separately from historical backtests. The Forward Performance view reports evaluated and pending records, average main-number matches, the full 0–5 match distribution, special-ball hits and results by model version. For supported five-main-number games it also shows the mathematical expected average number of main matches under uniform random drawing, k²/N, using the configured rule pool. Observed differences from that benchmark are descriptive and are not labeled predictive advantage without adequate evidence and statistical testing.

Statistical uncertainty

Forward 2+ match, 3+ match and special-ball hit rates are accompanied by 95% Wilson confidence intervals. PatternPicks also labels evidence depth by evaluated sample size so an early result is not displayed with the same apparent certainty as a larger body of forward observations. Confidence intervals describe uncertainty around observed rates; they are not future winning probabilities.

Ordered-digit games

Pick 3 and Pick 4 use a separate ordered-digit engine rather than the five-main-number model. The first implementation measures digit frequency independently at each position, verified sample size, digit sums and repeated-digit structures such as doubles, triples and quads. Its baseline research projection selects the most frequently observed digit at each position. This historical construction does not increase the mathematical probability of a future fair draw; digit-specific models are validated separately. The combined ordered-digit model is versioned as digit-combined-v0.1. Its baseline weighting is 45% long-term positional frequency, 30% trailing-90-day positional frequency and 25% target-session positional frequency, followed by a structural check against the historical median digit sum and modal odd/even composition. These weights are hypotheses to be evaluated rather than claims of optimality. The current digit walk-forward benchmark compares the combined model, long-term positional frequency, trailing 90-day positional frequency, target-session positional history and a deterministic random-digit control. The Digit Structure Lab separately reports sum interquartile range and median, odd/even composition, unique/double/triple/quad structures, and session-level average sums and repeat rates. Each historical target is evaluated using only prior draws, after a minimum 20-draw training history, for up to the latest 100 eligible targets. Metrics include average exact-position matches, straight hits, box hits and complete positional misses. The trailing model falls back to the prior training sample when fewer than 10 verified draws exist in the preceding 90 days.

Distinct research models

The walk-forward benchmark separates several hypotheses instead of treating one heuristic as momentum. Long-term frequency uses all prior training draws; trailing 90-day momentum uses only draws in the 90 calendar days before each target (falling back to the prior training sample when too few observations exist); target-weekday history uses only prior draws sharing the target draw’s weekday when sufficient history exists; structure-balanced and combined models retain their documented structural logic. All are compared with the same deterministic random control on the same target draws.

Walk-forward validation and random control

PatternPicks evaluates research models with walk-forward tests: each target draw is evaluated using only draws dated before that target. A deterministic pseudo-random valid-number control is tested against the same targets. This comparison is intended to expose whether apparent analytical performance differs from ordinary random matching in the tested sample; it does not establish future predictive advantage. Short samples can vary materially by chance.

Historical Pattern Fit

When a customer analyzes a five-number set in Pattern Lab, the beta Historical Pattern Fit combines three descriptive components: 40% observed prevalence of the set’s odd/even structure, 35% closeness of the set’s sum to the selected sample’s average relative to its observed sum range, and 25% average selected-number appearance frequency normalized to the sample. The 0–100 result measures similarity to historical sample characteristics only; it is not a probability, forecast, expected return, or claim of improved odds.

Validation and missing data

Dates, main-number counts, ranges, uniqueness and special-ball ranges are checked before analysis. Invalid records and duplicate draw identities are excluded and counted. Missing digit-game sessions are not filled with invented numbers. We do not yet certify that every scheduled draw is present. A visible source date and delayed-data message help expose freshness limits.

No predictive odds claims

A high historical percentage does not mean that number is more likely to appear next. In a fair independent draw, previous outcomes do not change the next draw’s probabilities. The beta does not supply calibrated forecasts, winning-number guarantees or a validated PatternScore. Any future model must be evaluated separately, with transparent out-of-sample testing.

Source and verification

Draw records originate from the State of New York’s public datasets. Complete, fresh windows are served from the Pattern Picks database; other windows use a direct public-source fallback. The analytics status identifies which connection supplied the records. Certified records maintained by the lottery prevail in a discrepancy. Review the data-source directory or export the filtered records to inspect the underlying draws.