Kelly has the suspicious charm of something that fits on a napkin. If a bet pays even money and wins with probability \(p\), bet \(2p - 1\) of your bankroll. If \(p = 0.60\), the full Kelly fraction is 20%.

That answer is correct for the toy problem. It is also where many real systems quietly walk onto a wet floor.

In a production betting, trading, or allocation system, the hard part is rarely typing the formula. The hard part is deciding whether the probability that goes into the formula is calibrated, stable, and measured against the same odds the system will actually trade. A five-point error in probability can become a ten-point error in position size at even odds. A model that looks excellent by accuracy can still be ruinous if its 60% forecasts only win 53% of the time.

Thesis: Kelly sizing is not primarily a courage problem. It is a calibration problem. The fraction should be treated as the last step of a probabilistic forecasting system, not as a magic multiplier placed on top of a backtest.

This is not investment advice. The examples below are deliberately simplified: independent binary bets, even-money payouts, no transaction costs, no market impact, and no liquidity constraints. The simplification is useful because it isolates the core failure mode: overconfident probabilities create oversized positions with very clean arithmetic.

Start With the Napkin, Then Put It Down

For a binary bet that wins \(b\) units per unit wagered and loses the stake when wrong, the expected log-growth from betting fraction \(f\) is

\[G(f; p, b) = p\log(1 + bf) + (1-p)\log(1-f).\]

Maximizing this expression gives the classical Kelly fraction

\[f^\* = \frac{bp - (1-p)}{b}.\]

At even money, \(b=1\), so \(f^\* = 2p-1\). The connection to information theory comes from Kelly’s original 1956 paper: the growth rate of capital can be linked to the information rate of a channel when bets are sized using the available signal.1 Thorp later made the criterion famous in blackjack, sports betting, and securities markets.2 MacLean, Thorp, and Ziemba’s edited volume remains one of the broad references on both the good and bad properties of Kelly-style capital growth.3

The formula has an appealing moral clarity. If you know the probability and the payoff, there is a unique fraction that maximizes long-run log wealth. Bet less and you leave asymptotic growth on the table. Bet more and you lower growth, increase drawdowns, and can eventually turn a positive edge into negative geometric growth.

But the word “know” is doing almost all of the work, and it is not a small word.

A Good Ranker Can Still Mis-Size

A classifier can be accurate and still useless for sizing. Suppose a model ranks opportunities well: its top bucket wins more often than its middle bucket, and its middle bucket wins more often than its bottom bucket. That is discrimination or ranking quality. Kelly sizing needs something stricter. It needs calibrated probabilities.

A probability forecast is calibrated when events assigned probability 60% actually happen about 60% of the time, events assigned probability 70% happen about 70% of the time, and so on. This idea has a long history in forecast verification, including Brier’s scoring rule for probability forecasts and Dawid’s work on calibrated Bayesian forecasters.4, 5

Modern machine learning makes this distinction practical. Guo, Pleiss, Sun, and Weinberger showed that modern neural networks can be poorly calibrated even when they are accurate, and that temperature scaling can often improve probability calibration with a simple held-out adjustment.6 The lesson carries directly into trading and betting: the model score is not yet a stake size. It is an input into a calibration layer, which is then an input into a sizing rule.

Kelly sizing as a forecasting pipeline A model score passes through calibration, odds comparison, sizing, and monitoring before it becomes a position. Model score ranking signal Calibration score to probability Market odds is there edge? Size fraction of capital outcomes update calibration, drawdown limits, and model trust the fraction is an output of the forecasting system, not a replacement for it
Kelly appears at the end of the pipeline. Before sizing comes probability calibration; after sizing comes monitoring.

Overconfidence Charges Interest

The easiest way to see the problem is to write down the wrong-growth function. If the true probability is \(p\), but the model reports \(\hat p\), then the system may bet

\[\hat f = 2\hat p - 1\]

at even odds, while realized log growth is still

\[G(\hat f; p, 1) = p\log(1+\hat f) + (1-p)\log(1-\hat f).\]

The growth function is asymmetric around the optimum. Betting too little gives up some growth. Betting too much can destroy it. If the true probability is 53% and the model says 60%, full Kelly recommends a 20% stake when the true growth-optimal stake is only 6%. That is not a small rounding error. It is a different risk regime.

This is why fractional Kelly is so common in practice. A half-Kelly rule is not just a psychological compromise, although it often helps with drawdowns. It is a rough defense against estimation error. Baker and McHale study Kelly betting under parameter uncertainty and show that shrinking bet size can improve out-of-sample performance when probabilities are estimated rather than known.7 Busseti, Ryu, and Boyd take a different route by adding drawdown constraints to Kelly gambling and deriving a convex risk-constrained formulation.8

Fractional Kelly is therefore better understood as a model-risk budget:

\[f_{\text{trade}} = \lambda f^\*(\tilde p), \quad 0 \le \lambda \le 1,\]

where \(\tilde p\) is a calibrated probability, not the raw model score. The fraction \(\lambda\) should not be chosen by vibes. It should reflect calibration error, non-stationarity, dependence between bets, liquidity, tail loss, and the drawdown a human or institution can actually survive.

A Tiny Betting Desk

The live experiment below creates binary opportunities with a real but noisy edge. A synthetic model ranks those opportunities, then becomes overconfident by stretching its logits. A held-out calibration sample estimates how much the logits should be shrunk before sizing. The chart compares four policies:

  • full Kelly using the raw model probability,
  • half Kelly using the raw model probability,
  • full Kelly after a held-out calibration step,
  • a guarded calibrated rule that subtracts a crude uncertainty reserve before sizing.

The audit badge is not a performance claim. It is the same deterministic 69-check suite exported by the script: even-money Kelly arithmetic, conservative guarding, side selection, path wealth multiplication, drawdown accounting, quantile ordering, calibration-ECE improvement in the default run, deterministic replay, and bounded trade-rate/drawdown summaries.

Raw full Kelly Raw half Kelly Calibrated Kelly Guarded calibrated

Deterministic simulation, 360 paths of 260 even-money bets. Dots are median terminal wealth per $1; whiskers show the 5th to 95th percentile on a log wealth scale. The audit badge checks implementation contracts; it does not turn the toy model into investment advice or certify live-market execution.

The static-site audit runner treats the lab as an executable model-risk memo:

node - <<'NODE'
const lab = require("./assets/js/kelly-lab.js");
const EXPECTED_AUDIT = { checked: 69, failures: [], ok: true, passed: 69 };
const EXPECTED_DEFAULT = [
  "0.33",
  "0.098968",
  "0.014779",
  "rawFull,rawHalf,calibrated,guarded",
  "rawFull=0.007685/1.748980/332.108884/0.965865/0.997831",
  "rawHalf=0.355387/4.928479/65.590730/0.723760/0.995363",
  "calibrated=0.395256/6.898316/126.832009/0.752628/0.993088",
  "guarded=0.681797/4.634650/30.212029/0.572983/0.623643"
].join(":");

function fixed(value, digits = 6) {
  return Number(value).toFixed(digits);
}

function sameJson(actual, expected) {
  return JSON.stringify(actual) === JSON.stringify(expected);
}

const audit = lab.runAudit();
const result = lab.evaluate({
  overconfidence: 180,
  calibration: 800,
  edge: 32,
  cap: 25
});
const summaryShape = result.summaries.map((row) => `${row.key}=${[
  fixed(row.q05),
  fixed(row.median),
  fixed(row.q95),
  fixed(row.medianDrawdown),
  fixed(row.tradeRate)
].join("/")}`).join(":");
const defaultShape = [
  fixed(result.slope, 2),
  fixed(result.rawEce),
  fixed(result.calibratedEce),
  result.summaries.map((row) => row.key).join(","),
  summaryShape
].join(":");

if (!sameJson(audit, EXPECTED_AUDIT) || defaultShape !== EXPECTED_DEFAULT) {
  throw new Error(JSON.stringify({ audit, defaultShape }, null, 2));
}

console.log(`${audit.passed}/${audit.checked} Kelly sizing checks passed`);
NODE

The point is not that a toy simulator settles portfolio theory. It does not. The point is more operational: when the model is overconfident, the raw Kelly policy often has the most impressive in-sample story and the ugliest realized wealth distribution. Calibration turns a ranking model into a decision model.

In the default deterministic run, the fitted calibration slope is about 0.33: the model’s logits need to be cut by roughly two-thirds before they behave like probabilities. Expected calibration error falls from about 9.9 percentage points to about 1.5 percentage points, while the raw full-Kelly policy’s 5th-percentile terminal wealth is less than one cent per starting dollar. The exact numbers are toy numbers, but the shape of the failure is real: an overconfident model can turn a positive ranking signal into a fragile sizing rule.

Risk Board Before Victory Lap

A serious Kelly-like system needs more than a backtest equity curve. I would ship it with at least seven diagnostics.

  1. Reliability curves by regime. A global calibration curve can hide that the model is calibrated in liquid markets and overconfident in thin ones.
  2. Expected calibration error and log loss over rolling windows. Accuracy alone is not enough because size depends on probability magnitude.
  3. Realized edge by predicted edge bucket. A 4% predicted edge bucket should not behave like a 0% bucket after costs.
  4. Turnover, capacity, and slippage by size bucket. The largest theoretical edges are often the least executable.
  5. Correlation between active bets. Kelly assumes a compounding process; clustered losses can make independent-bet sizing dangerously optimistic.
  6. Drawdown-at-risk under replay and block bootstrap. A path that is acceptable in expectation may be unacceptable to the person living through it.
  7. Kill-switch reasons. A system should record whether it did not trade because of missing features, stale calibration, low edge, high uncertainty, or risk budget exhaustion.

The calibration layer should be refit on a held-out or rolling sample, never on the same labels used to train the predictive model without correction. Platt scaling, isotonic regression, beta calibration, temperature scaling, and Bayesian binning are all reasonable candidates depending on sample size and model shape.9, 10 The method matters less than the discipline: probability forecasts must be audited as probabilities.

Where the Toy Stops Being a Market

Real markets are not a sequence of independent even-money bets. Odds move, liquidity disappears, constraints bind, taxes matter, and a strategy’s own trading can erase its edge. The edge estimate is also not exogenous. If many participants use similar data and similar models, the opportunity may vanish or become crowded exactly when the backtest says to scale.

Kelly also optimizes long-run logarithmic growth, not comfort, reputation, liability matching, or institutional survival. Samuelson objected to careless “large numbers” arguments in risk taking, and the later time-average literature kept the debate alive by asking what kind of averaging is relevant for a single path through time.11, 12 You do not need to take a side in every philosophical argument to get the engineering lesson: the objective function has to match the owner of the capital.

There is also a behavioral warning. Haghani and Dewey studied a simple biased coin game in which participants were told the coin had a 60% chance of heads; many still bet in ways that left large amounts of value on the table, and a substantial fraction went broke.13 Bet sizing is not intuitive even when the rules are known. It is much harder when the probabilities come from a model.

My Desk Copy

The Kelly criterion is a sharp instrument. It tells us that edge, odds, and capital growth are linked by a clean mathematical rule. But sharp instruments amplify upstream errors.

The practical sequence should be:

  1. Build the best probability forecast you can.
  2. Calibrate it on data the model did not train on.
  3. Compare it to executable odds after costs.
  4. Shrink for model risk and drawdown constraints.
  5. Monitor calibration and realized edge continuously.

Only then does the Kelly fraction deserve to become a trade size.

The formula is elegant. The system around it is where the intelligence lives.

References

  1. John L. Kelly Jr., “A New Interpretation of Information Rate”, Bell System Technical Journal, 1956. ↩

  2. Edward O. Thorp, “The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market”, in The Kelly Capital Growth Investment Criterion, World Scientific, 2011. ↩

  3. Leonard C. MacLean, Edward O. Thorp, and William T. Ziemba, The Kelly Capital Growth Investment Criterion: Theory and Practice, World Scientific, 2011. ↩

  4. Glenn W. Brier, “Verification of Forecasts Expressed in Terms of Probability”, Monthly Weather Review, 1950. ↩

  5. A. P. Dawid, “The Well-Calibrated Bayesian”, Journal of the American Statistical Association, 1982. ↩

  6. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger, “On Calibration of Modern Neural Networks”, ICML 2017. ↩

  7. Rose D. Baker and Ian G. McHale, “Optimal Betting Under Parameter Uncertainty: Improving the Kelly Criterion”, Decision Analysis, 2013. ↩

  8. Enzo Busseti, Ernest K. Ryu, and Stephen Boyd, “Risk-Constrained Kelly Gambling”, The Journal of Investing, 2016. ↩

  9. John Platt, “Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods”, in Advances in Large Margin Classifiers, 1999. ↩

  10. Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht, “Obtaining Well Calibrated Probabilities Using Bayesian Binning”, AAAI 2015. ↩

  11. Paul A. Samuelson, “Risk and Uncertainty: A Fallacy of Large Numbers”, Scientia, 1963. ↩

  12. Ole Peters and Murray Gell-Mann, “Evaluating Gambles Using Dynamics”, Chaos, 2016. ↩

  13. Victor Haghani and Richard Dewey, “Rational Decision-Making Under Uncertainty: Observed Betting Patterns on a Biased Coin”, 2016. ↩