Walk-forward evaluation
Backtesting
Every model is fitted on prior drawings only, then scored on the drawings that followed. Results are compared to uniform random selection over the same window and ticket count.
A fair drawing has no memory. Published Pick 3 research consistently finds no exploitable pattern, and the honest result of this analysis is that models perform within sampling error of random selection. Any apparent edge smaller than the confidence interval is noise.
Evaluation window
Compare models on identical conditions
All models evaluate the same target drawings with the same ticket count.
Live forecasts (not backtest)
Genuine pre-draw performance
Only forecasts sealed before their target drawing, scored after the official result.
Exact hits
0
0.00% rate
Any-order hits
0
0.00% rate
Prizes
$0
cost $37
Net
-$37
Live forecasts are sealed before their target drawing and evaluated afterwards. A small sample will show large apparent swings; the uniform baseline hit rate for 5 tickets is 0.005 exact and 0.03 any-order.
Computing
Running walk-forward evaluation
Each model is refitted across the evaluation window. This takes a few seconds on first run, then results are cached.