Kelly Position Sizing Framework

Kelly is a position sizing framework: it tells you the optimal fraction of capital to risk on a trade when you have a measurable edge. In trading, the key is translating "fraction to bet" into risk per trade (and then into shares/contracts using your stop).

1) The Core Kelly Formula (Single Trade)

Let:

p = probability of winning
q = 1 − p
b = payoff ratio = (average win) / (average loss), measured in R multiples if you use fixed risk per trade

Then the Kelly-optimal fraction of capital to risk is:

f* = (bp - q) / b = p - (1-p) / b

Interpretation:

If f* ≤ 0: you should not take the trade (edge not positive).
If f* > 0: risk approximately f* of equity as "risk" (the amount you lose if stopped out), not as notional exposure.

Quick Example

Suppose your backtest shows:

win rate p = 0.45
average winner = 2R, average loser = 1R → b = 2

f* = 0.45 - (0.55 / 2) = 0.45 - 0.275 = 0.175

Full Kelly would suggest risking 17.5% of equity per trade—which is far too aggressive for real trading because your estimates are noisy and outcomes are not IID.

2) Use "Fractional Kelly" in Real Trading (Standard Practice)

In practice you apply a haircut:

f = λ · f* where λ ∈ [0.1, 0.5] is common

Typical choices:

0.25 Kelly for aggressive systematic strategies with stable edges
0.10–0.20 Kelly for discretionary or noisy edges
0.50 Kelly is already very aggressive

Using the example above:

0.25 Kelly → f = 0.25 × 0.175 = 0.04375
So you'd risk ~4.4% of equity per trade (still high for many traders; many cap at 0.5–2% regardless).

3) Convert Kelly "Fraction" into an Actual Position Size (Shares/Contracts)

Kelly gives you risk budget per trade:

$Risk = f · Equity

Then position size is set by your stop distance:

Position Size = $Risk / Stop Distance ($ per share/contract)

Example:

Equity = $100,000
Fractional Kelly f = 0.01 (1% risk)
Stop distance = $2/share

$Risk = 0.01 × 100,000 = $1,000
Shares = 1,000 / 2 = 500

This makes Kelly usable in any stop-based trading system.

4) Estimating Inputs Correctly (The Real Hard Part)

(a) Use R-Multiples

Define each trade in R units:

Loser ≈ −1R (by construction if you honor stops)
Winners measured in R (e.g., +1.6R, +2.3R, etc.)

Then:

p = fraction of trades with R > 0
b = average(R | win) / average(|R| | loss)

(b) Beware Regime Dependence and Sample Error

Kelly is extremely sensitive to estimation error—especially in p and b. Practical safeguards:

Use walk-forward / out-of-sample estimates
Use Bayesian shrinkage (pull p toward 0.5, b toward 1 when data is sparse)
Use caps: max risk per trade (e.g., 1–2%), max gross leverage, max sector exposure

5) A Trading-Ready "Kelly Workflow"

Define your trade unit (setup + stop + exit rules) so you can compute R.
Backtest / journal to estimate p and b (preferably by regime).
Compute f* = p - (1-p)/b. If f* ≤ 0, skip the trade.
Apply fractional Kelly: f = λf* with λ conservative.
Apply hard risk caps (e.g., min(f, 1%)).
Convert to shares/contracts using stop distance.
Monitor edge drift: recalc p, b on rolling windows.

6) Extensions You'll Want if You Trade a Portfolio

Single-trade Kelly assumes independent bets. Real portfolios require constraints:

Correlated positions: scale down when trades move together (otherwise you overbet).
Multiple simultaneous bets: use "portfolio Kelly" or simpler proxies:
- reduce λ as the number of correlated positions rises
- cap total portfolio risk (sum of per-trade risks) to a fixed limit
- volatility targeting at the portfolio level

7) When Kelly is a Bad Idea

Avoid relying on Kelly when:

You can't estimate p and b with reasonable stability
Your payoff distribution has fat tails (gap risk, short options) and losses exceed modeled "average loss"
You routinely violate stops (then the loss side is unbounded)

A Practical Default (If You Want Something Deployable)

Compute Kelly from rolling out-of-sample R-multiples.
Use 0.10–0.25 Kelly.
Enforce:
- max risk per trade: 0.5–2% (depending on style)
- max total open risk: 3–8% across positions (style-dependent)
- correlation haircut: reduce risk when trades cluster (same sector/factor)

Need help calculating your Kelly? If you provide (1) your typical win rate, (2) average win in R, (3) average loss in R (or stop size), and (4) how many positions you hold concurrently, you can compute your Kelly and develop a conservative fractional-Kelly sizing rule tailored to your trading style.