Skip to main content

Estimating Edge & Odds

Updated Jun 24, 2026 at 2:35pm

Research Draft High 1,241 words

The Kelly criterion is only as good as the two numbers fed into it: the probability of winning (the edge) and the win/loss payoff ratio (the odds). The formula assumes these are known. In trading they are never known — they are estimated from a finite, noisy, regime-dependent sample of past trades. This node is about that estimation problem: how W and R are derived, how much sampling error they carry, and why the asymmetric penalty for getting them wrong (overbetting ruins, underbetting only slows) is the central reason practitioners deliberately shrink their inputs. The estimation gap — not the algebra — is what makes "full Kelly" dangerous in real markets.

How the inputs are estimated

For a trading strategy, the standard discrete-Kelly inputs come straight from a trade log or backtest:

  • W (win rate) = winning trades ÷ total trades.
  • R (payoff ratio) = average win ÷ average loss (gross profit per winner ÷ gross loss per loser, in dollars or R-multiples).
  • Kelly fraction f = W − (1 − W) / R, equivalently f = (p·b − q) / b where p = W, q = 1−W, b = R.

A commonly cited worked example: 120 trades, 66 winners (W = 0.55), average win \$850, average loss \$500 (R = 1.70) gives f\* = 0.55 − 0.45/1.70 ≈ 0.285, i.e. risk ~28.5% of capital per position (JournalPlus, Ryan O'Connell). That single number exposes the problem — a 28.5% bet is enormous, and it rests entirely on two sample statistics.

The continuous analogue, used for return-distribution strategies rather than discrete bets, is f* ≈ μ / σ² (excess mean return over return variance) — which is just the mean-variance form and inherits the same estimation fragility, since μ is the hardest moment to estimate from price data.

Why the estimates are fragile

Both inputs are sample means, so their uncertainty shrinks only with √N. Three structural problems make them worse than they look:

1. Win rate has wide confidence intervals at realistic sample sizes. With 100 trades a 55% observed win rate has a roughly ±10-point sampling band (binomial standard error ≈ √(p·q/N) ≈ 5%). The Kelly fraction is highly sensitive to W, so a band that wide on the input produces a very wide band on the bet size. 2. The payoff ratio is dominated by tails. R is the ratio of two averages, and the average win is pulled by a small number of large outliers. A few big winners (or letting one loss run) can swing R substantially, and that fat-tailed mean converges slowly. 3. The plug-in estimator is biased upward. Substituting sample estimates into the Kelly formula systematically overstates the true optimal fraction — the estimated Kelly portfolio is biased toward overbetting (quant literature on plug-in vs. shrinkage Kelly estimators). Backtest overfitting and survivorship in the trade sample push the same direction.

The asymmetry that governs everything

The reason estimation error is not a symmetric "it'll average out" problem is the shape of the long-run growth curve as a function of bet size:

  • Growth is maximized at exactly full Kelly (f\*).
  • Underbetting (f < f\) only slows* compounding — the curve is gentle on the left.
  • Overbetting (f > f\*) degrades growth fast, and beyond ~2× Kelly the expected long-run growth rate turns negative — guaranteed ruin despite a genuine positive edge (Thorp; Wikipedia).

Because overestimating edge pushes you right on this curve toward the ruin zone, and because the plug-in bias and overfitting already bias estimates upward, the expected-value-correct move is to bet less than the point estimate says. Thorp's classic result: half-Kelly captures roughly 75% of the growth rate while cutting variance and drawdown roughly in half, and protects against the negative-growth zone at a maximum cost of ~25% of growth. This is the formal justification for fractional Kelly (covered in detail in the sibling Full vs Fractional Kelly node).

How it's used in practice

Practitioners treat the formula output as a ceiling, not a target, and shrink the inputs before shrinking the fraction:

  • Haircut the inputs. Use a deliberately conservative W and R — lower the win rate and payoff ratio below historical values — rather than trusting backtest figures (Complete Trader's Edge, AvaTrade). This is shrinkage applied at the source.
  • Apply a fractional multiplier (commonly ¼ to ½ Kelly) on top, to absorb the residual model error and the upward plug-in bias.
  • Out-of-sample validation. Estimate W and R on a holdout / walk-forward period, never the same data used to design the strategy, because in-sample stats overstate edge.
  • Re-estimate as a rolling window so a decaying edge (regime change) lowers the bet automatically; a strategy's true W and R are not stationary.
  • Require a minimum sample (many traders want ≥30–100 closed trades) before sizing off Kelly at all; below that, the confidence interval on W is too wide to act on.
  • Bayesian / shrinkage Kelly — formally, blending the sample estimate toward a prior is equivalent to a fractional Kelly with full information, giving a principled way to discount thin or noisy evidence.

A counterintuitive but well-documented simulation result (Downey, summarizing Thorp): uncertainty alone shrinks the optimal bet only modestly (e.g. from 0.40 toward ~0.36 at moderate input uncertainty). Fractional Kelly is "overdetermined" — it is justified less by symmetric uncertainty and more by systematic overestimation of edge, risk-of-ruin aversion, and a preference for protecting downside (lower-percentile) outcomes rather than the median.

Strengths & limitations

The discipline's strength is that it makes the bet size an explicit function of measured edge — it forces you to quantify W and R rather than guess position size, and it correctly punishes a thin edge with a small bet. Its limitations are entirely in the inputs: Kelly assumes a stable, known distribution and one bet resolved before the next, which trading violates (overlapping positions, correlated holdings, fat tails, non-stationary edge). The single most common misuse is feeding in-sample / overfit backtest statistics into the raw formula and sizing at full Kelly — combining an upward-biased estimate with the unforgiving overbetting side of the growth curve, which is how a genuinely positive-edge system still blows up. Correlated simultaneous positions also break the single-bet assumption: their combined effective Kelly is larger than the sum of independently-sized bets suggests.

Sources