What the Kelly Criterion Actually Says
The Kelly criterion is a mathematical result from information theory. Given a known probability of winning and a known payoff ratio, it identifies the stake size that maximises the long-run geometric growth rate of your capital.
The logic is clean. Bet too little, and you leave growth on the table. Bet too much, and variance destroys you faster than your edge can rebuild you. There is a precise point between those two outcomes - the growth-optimal stake - and Kelly tells you where it is.
The criterion works. On its own terms, for the problem it was designed to solve, it is correct.
The problem is the phrase "known probability of winning."
Why "Known Edge" Is the Load-Bearing Assumption - and Why It Fails
In practice, nobody knows their edge. You estimate it.
You measure your historical win rate. You calculate your expected payoff. You run a backtest, observe the results, and derive a number. That number is your estimated edge - and it has a systematic flaw built into it.
Models overfit to the data they are trained on. Backtests look cleaner than live performance because they are computed in-sample, where the curve is smoothed by hindsight. The edge you measured is almost always larger than the edge that survives contact with the market. Entry costs bite. Slippage occurs where the model assumed none. Market conditions shift in ways the training data did not capture.
The result is predictable: your estimated edge is an upper bound, not a point estimate. The true edge is somewhere below it, and you do not know by how much.
Betting full Kelly on an inflated estimate is not aggressive sizing. It is applying a formula derived for a known quantity to an unknown one - and the formula does not warn you when you do this. It just runs the math on the number you gave it.
The Asymmetry That Runs Against You
Here is the part that is easy to miss.
The relationship between stake size and long-run geometric growth is not linear. It is concave - a peak with a falling slope on both sides. Betting below the true Kelly optimum costs you growth. Betting above it costs you more than you gained.
The asymmetry is structural. Moving from half-Kelly to full Kelly does not double your growth rate. The growth gain is modest. But the drawdown exposure scales far more steeply. When your estimated edge is higher than your true edge - which is the typical direction of the error - you are not sitting at the optimum. You are sitting past it, on the descending slope, in the region where every additional unit of stake is costing you more in expected drawdown than it is adding in expected growth.
And drawdown has a compounding cost that growth does not. A loss of a given percentage requires a larger percentage gain to recover. The larger the drawdown, the longer you spend recovering rather than compounding. Full Kelly, applied to an overestimated edge, reliably produces larger drawdowns than the single-bet math implies.
What the Fraction Is Actually Doing
Fractional Kelly - sizing at a half or a quarter of the full criterion - is standard practice in systematic trading and quantitative portfolio management. The reason is not conservatism.
The fraction is doing epistemic work that your edge estimate cannot do. It is the rational response to the gap between the edge you measured and the edge that is actually there. If your estimate is inflated by a factor you cannot observe, a fractional stake means you are betting closer to the true optimum even without knowing where the true optimum is.
The discipline also compounds across a sequence of trades. A sizing method that survives a run of overestimated edges - as most edges will be, at least initially - outperforms one that does not, regardless of what any individual bet's expected value looked like on paper. Staying in the game is itself a structural edge, and fractional Kelly is what protects it.
Vulcan Trading's Polymarket Flow engine applies fractional-Kelly sizing as a structural feature of the engine, capped at a defined fraction of bankroll per signal, in paper mode. The cap is not a setting. It applies regardless of how strong a given signal appears, because a strong-looking signal from an estimated edge is still an estimated edge. The engine treats its own model uncertainty as a first-class input to the sizing logic.
The reasoning behind that cap is the reasoning behind this piece. Polymarket Flow runs in paper mode on Vulcan Trading - the logic is there to examine, separate from any result it produces.