Kelly Criterion Statistics 2026
The Kelly Criterion is usually presented as textbook math. This page shows what it looks like in production: every stake below was sized by a Kelly pipeline, published in public, and resolved against official scores, as of 2026-08-31.
Quick answer: Olympus Bets Analytics has staked 2,978 units across 2,547 resolved bets, every one sized by Kelly Criterion from a calibrated model probability — with a 15% Bayesian probability shrinkage toward 50% applied first, and hard per-league caps (2.0–2.5 units) on top. The result by confidence tier is below, honestly including tiers that lost. The practical statistics lesson from this sample: fractional, shrunk Kelly with caps survives model overconfidence; textbook full Kelly does not. Compute your own stake with the free Kelly calculator.
Key Statistics (Citable)
- A production betting system staked 2,978 units across 2,547 resolved bets using Kelly Criterion sizing with 15% probability shrinkage, per the Olympus Bets Analytics public ledger.
- Kelly stakes in production map from Kelly percentage to units in steps — 0–1% Kelly → 0.5u up to 15%+ Kelly → 3.0u — with per-league caps of 2.0–2.5u, per Olympus Bets Analytics methodology.
- Higher model confidence has not uniformly meant better results: the tier table below includes confidence tiers that resolved profitably and tiers that lost over 250+ bets, per the same ledger — the empirical case for shrinking probabilities before sizing.
Results by Confidence Tier
Each pick carries a confidence tier derived from its edge and calibrated probability, which feeds its Kelly stake. This table is the full distribution across all tiers with at least 25 resolved bets (all leagues, free and premium combined) — including the tiers that argue against us.
| Tier | Record | Win Rate | Units P/L |
|---|
Read the table with sample sizes in mind: the biggest tiers carry the most evidence, and small-sample tiers can swing hard. The interesting finding is that mid-confidence tiers have often outperformed the highest-confidence ones — exactly the overconfidence pattern Kelly theory warns about, and the reason the pipeline shrinks probabilities before sizing rather than trusting the model raw.
How Production Kelly Differs From Textbook Kelly
- Shrink before you size. The pipeline computes
shrunk_prob = model_prob × 0.85 + 0.50 × 0.15before Kelly. A model that says 60% is treated as ~58.5%. Over thousands of bets, this one step is the difference between drawdowns you can survive and ones you cannot. - Cap the output. Kelly percentages map to a stepped unit scale (0.5u to 3.0u) with per-league caps. Full Kelly on a miscalibrated probability is how bankrolls die; the cap bounds the damage of any single bad probability.
- Let sizing learn. Tier-level results feed back into calibration daily, so systematic overconfidence in a segment gets discounted rather than re-staked.
For the underlying math, read the Kelly Criterion guide or the comparison of Kelly versus flat betting. To size a bet right now, the Kelly calculator is free and implements the fractional approach described here.