What happened
An explainer reintroduces the Kelly criterion, derived from John L. Kelly Jr.'s 1956 paper, and lays out the formula f* = (bp − q) / b, giving b = odds − 1 and f* = (odds × p − 1) / (odds − 1) for fractional odds.
Why it matters
Because the rule maximizes expected log growth rather than expected profit, the author argues that betting beyond the optimal fraction, or using an overconfident model, pushes long-run growth negative and can lead to ruin even when the edge is positive.
What to watch
Because probability estimates are imperfect, the author says the sizing hinges on calibration — checking predicted probabilities against actual outcome frequencies — and recommends using Half Kelly or smaller with a cap rather than Full Kelly.
WHO IT HITSThis matters to anyone sizing positions with a model-generated probability — quantitative traders, sports bettors and data scientists building betting or allocation strategies — because the guidance is to trust a model's calibrated probabilities, not its accuracy, when deciding how much to stake.
Summaries like this, in your inbox every morning.
The piece traces the Kelly criterion back to John L. Kelly Jr.'s 1956 paper, originally framed in communication theory to describe how information contributes to capital growth, and notes that the idea later migrated to casinos, sports betting, investing and portfolio optimization. What ties these settings together is a single shift in objective: rather than maximizing expected profit per bet, the rule maximizes the expected logarithmic growth rate of capital, which is the compounding growth rate. The author illustrates this with the classic coin-flip example of 60% wins and 40% losses at even money, where the expected value is positive (0.2 per unit staked) yet betting the entire bankroll every round eventually wipes you out. The Kelly fraction there works out to 20%, and betting 30% or 40% instead pushes expected log growth down and eventually below zero — the rule's sharpest lesson is where over-betting turns positive expectations into long-run decline.
From there the explainer moves to practical safeguards. Because the true win probability is never known, using the theoretical fraction, called Full Kelly, is risky; the article presents a spectrum from Full Kelly (c = 1.0) through Half Kelly (0.5), Quarter Kelly (0.25) and 1/8 Kelly (0.125). The machine-learning connection is where the argument sharpens: what Kelly consumes is a probability, not a label, so a model that merely classifies correctly can still be unsafe if its probabilities are misaligned, and the author points to metrics such as Log Loss, Brier Score, ECE and the calibration curve as ways to check alignment. The same caution carries into decimal odds, where the bookmaker's margin — visible when implied probabilities sum above 100% — must be accounted for before any edge is assumed real.
The later sections lay out an implementation path: a Kelly fraction function with fractional and cap parameters, an expected-log-growth function, and a minimal backtest that splits data in time order, verifies calibration on a validation period, and then sizes only bets with a genuine edge, before evaluating bankroll path, maximum drawdown, ROI and number of bets. The author's closing caution is that a Kelly strategy failing these checks is dangerous even if its formulas are correct, and that the real question when pairing the rule with machine learning is how far the model's probabilities can be trusted. Whether that trust is warranted hinges on the calibration evidence in each specific case, and for practitioners sizing real positions with model output, that is the practical test to watch.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
