Powerball · Prediction lab
Can anything beat random — out of sample?
This is an experiment, not a service. Eight models pick 5 numbers for every draw using only past data (no lookahead), and each is scored against the same seeded random baseline. Results are split across three time periods and corrected for testing many models at once. On a fair lottery, the expected outcome is: noise.
| Model | Avg matches | vs random | Period 1 | Period 2 | Period 3 | Signal? |
|---|---|---|---|---|---|---|
| Random baseline Uniform random selection — the bar every model must clear. | 0.360 | 0.000±0.012 | 0.35 | 0.36 | 0.36 | baseline |
| Frequency Most-drawn numbers in the past window. | 0.380 | +0.020±0.012 | 0.40 | 0.37 | 0.37 | noise |
| Recency Numbers drawn most recently. | 0.380 | +0.020±0.012 | 0.43 | 0.37 | 0.35 | noise |
| Gap Numbers absent the longest. | 0.263 | -0.097±0.012 | 0.05 | 0.37 | 0.36 | noise |
| Pair co-occurrence Greedy set built from the most co-occurring pairs. | 0.371 | +0.011±0.012 | 0.38 | 0.38 | 0.35 | noise |
| Sum pattern Set whose sum tracks the historical average sum. | 0.382 | +0.023±0.012 | 0.42 | 0.34 | 0.39 | noise |
| Bayesian blend Frequency smoothed with recency (a prior-weighted mix). | 0.387 | +0.028±0.012 | 0.41 | 0.38 | 0.38 | noise |
| Ensemble Average of normalized frequency, recency, and gap scores. | 0.382 | +0.022±0.012 | 0.41 | 0.36 | 0.38 | noise |
How the experiment stays honest
- No lookahead: a model's pick for a draw uses only draws before it (same harness as the backtesting lab).
- Period splits: matches are also reported per time period so a lucky stretch can't masquerade as skill.
- Multiple-comparison correction: with 8 models, a ~3σ difference is required before anything is flagged (Bonferroni-style, α = 0.05/8).
- Reproducible: dataset version 1357a99f, calc version 1, fixed seeds.
Every model's verdict is printed next to its numbers. If none is flagged, the honest conclusion is: no model beat a random baseline out of sample — which is exactly what a fair lottery predicts.
What this doesn't mean
- Historical frequency and patterns do not change the probability of any future draw — every combination is equally likely.
- This tool does not predict, guarantee, or improve winning odds. Anyone claiming otherwise is selling something.
- Simulated and backtested results describe the past and model a fair random process; they are experiments, not advice.
- Lottery tickets are a form of entertainment, not an investment. Play within your means.