Strategies Lab

Powerball strategy backtests

Walk-forward EV simulations across 10 strategy presets. Calibrated against 3,777 historical draws.

Forest plot — mean EV ± IQR vs baseline

Baseline: quick mix(-1.818)

No strategy IQR clearly separates from baseline — differences are sampling noise

Bar tip = mean EV. Whiskers = interquartile range (p25 to p75) from 20,000 backtested picks per strategy. Dashed line = quick_mix baseline. Strategies whose IQR straddles the baseline are not distinguishable from random under walk-forward CV. All strategies have negative mean EV (lottery house edge is structural).

Efficient frontier — win rate × mean EV

Different question from the forest plot: not "is it better" but "what is each strategy optimizing for?".

Pareto-efficient (1)Dominated by baseline

Baseline: quick_mix @ win 3.94%, EV -1.818

Each dot is one strategy. Vertical + horizontal slate lines mark the baseline position; dots in the upper-right quadrant Pareto-dominate baseline on both win rate and EV. The red EV=0 line is the unreachable breakeven — house edge keeps every strategy below it. Use S2 forest plot for "is this strategy actually better?"; use this chart for "what is each strategy optimizing for?".

Draw shapes — what a typical draw looks like

Descriptive views of the public 12-month timeseries. None of these are predictive — they characterize the shape of the data so anomalies pop.

T1 · Calendar heatmap of sum z-scores

154 draws · sum mean 178.3 ± 40.6

z-score:

±3σ

Each cell is one draw date in the public 12-month window. Color = z-score of the sum of main numbers (blue = low sum, red = high sum). Day-of-week clustering of extreme cells would suggest schedule-driven bias — typically none is visible because draws are i.i.d. across the calendar.

T4 · Sum-of-mains distribution

154 draws · sum range 77–285 · observed mean 178.3 ± 40.6

Theoretical mean 175.0 · offset +3.3 (0.08σ)

Solid line = observed mean. Gray band = ±1σ around observed. Red dashed line = theoretical mean under uniform K-of-N draws. The shape should be approximately symmetric and unimodal (CLT). Material bimodality or skew would indicate format change or sampling artifact.

T9 · Joint distribution — sum × range spread

154 draws · 103 populated cells (of 20×20) · max 5 draws / cell

density:

log

Each cell is a (sum, spread) bin; color is log(count) so heavy-tail bins don't wash out the visualization. The diagonal-ish envelope reflects an analytic constraint: a high range spread (max − min close to N) forces moderate sums; low spread forces extreme sums. Under uniform draws this envelope is symmetric — material asymmetry would indicate bias.

Phase 2.5+ build — more strategy charts incoming

S1 notched boxplots (p25/median/p75 with mean diamonds), S4 efficient frontier (win-rate × mean-EV with breakeven references), strategy leaderboard with sparklines, and a strategy builder (free 3 runs/day) ship in subsequent Phase 2 rounds.

See top-3 cards in Frequency Lab →