Position sizing basics: fixed fractional, vol targeting and Kelly

Two traders can run the same entry and exit rules on the same market and end the year in very different places. The difference is usually how much they put on each trade. Position sizing gets less attention than signals because it feels like bookkeeping, but it decides how deep your drawdowns go and whether a run of bad luck is survivable.

Here are the three sizing methods you'll meet most often in systematic trading, how each one works with real numbers, and where each one tends to go wrong.

Fixed fractional: risk the same slice of the account each time#

Fixed fractional sizing says: on every trade, risk a set percentage of current equity. "Risk" here means the amount you lose if your stop is hit, not the size of the position.

A worked example makes it concrete. The stock and the prices are made up; the arithmetic is the point:

Account equity          $50,000
Risk per trade          1%  ->  $500
Entry / stop            $40.00 / $37.50  ->  $2.50 risk per share
Position size           $500 / $2.50 = 200 shares  (worth $8,000)
Same rule, $1.00 stop   $500 / $1.00 = 500 shares

Two things fall out of this. First, the stop distance drives the size, so a sloppy or arbitrary stop gives you an arbitrary position. Second, because the percentage applies to current equity, positions shrink after losses and grow after wins. Here's what a run of ten straight losses does at two common risk settings:

1% per trade:  0.99^10 = 0.904  ->  down about 9.6%
2% per trade:  0.98^10 = 0.817  ->  down about 18.3%

Gain needed to get back to the start
  after an 18.3% drawdown:  1 / 0.817 - 1  =  about 22%
  after a 50% drawdown:     1 / 0.5 - 1    =  100%

Long losing streaks do turn up in systems that trade often, so the risk setting you pick is really a choice about how deep a hole you're willing to climb out of.

Fixed fractional has real weaknesses. The planned loss assumes your stop fills at the stop price, and gaps or thin markets can blow straight through it. It also says nothing about how volatile the instrument is beyond what the stop implies, and nothing about how good the strategy is. It's a risk budget, not an optimiser.

Volatility targeting: size so each position carries similar risk#

Volatility targeting sizes positions so the portfolio (or each position) aims at a fixed level of volatility instead of a fixed dollar exposure. The basic formula is position weight = target volatility ÷ forecast volatility.

Here's the rule applied to an equity index, aiming for ten per cent annualised volatility:

Forecast vol 20%  ->  10 / 20 = 0.50  (half the capital)
Forecast vol  8%  ->  10 / 8  = 1.25  (leveraged)
Forecast vol 40%  ->  10 / 40 = 0.25  (a quarter)

That middle line is why the Man Group researchers describe the method as taking larger exposures during low-volatility episodes. Most systems put a hard cap on the weight so a quiet period can't push leverage to silly levels.

The volatility forecast matters as much as the formula. In The Impact of Volatility Targeting, Harvey and co-authors from Man Group used the standard deviation of daily returns with exponentially decaying weights, mostly with a 20-day half-life, and required 270 trading days of data before scaling anything. Faster estimates react sooner but trade more: their turnover ranged from about five round trips a year for the most responsive estimates to under one a year for the slowest.

Their results across more than 60 assets were mixed in a useful way. Volatility scaling improved Sharpe ratios for equities and credit, but the effect on bonds, currencies and commodities was negligible. It did reduce the likelihood of extreme returns across asset classes, and it cut maximum drawdowns for the balanced and risk parity portfolios they tested. Their explanation for the equity result is the leverage effect: volatility tends to rise after prices fall, so the rule cuts exposure after declines. A separate NBER working paper by Moreira and Muir reached a similar conclusion for several equity factors, finding that portfolios taking less risk when volatility is high produced higher Sharpe ratios.

The caveats are practical. Your volatility estimate is always backward-looking, so a sudden shock hits you at full size before the estimate catches up. Low-volatility periods quietly push you towards your leverage cap. And a "same volatility" portfolio isn't a "same risk" portfolio if the instruments can gap or become illiquid.

The Kelly criterion: the fraction that maximises long-run growth#

The Kelly criterion comes from a 1956 paper by John Kelly in the Bell System Technical Journal, A New Interpretation of Information Rate. Kelly looked at a gambler with an edge who reinvests winnings and showed that betting a fixed fraction of capital lets money grow exponentially. He also pointed out that betting everything on each favourable bet maximises expected wealth, yet leaves the gambler broke with probability one if he keeps going. The answer is the fixed fraction that maximises the long-run exponential growth rate of capital, which in later literature is usually described as maximising the expected logarithm of wealth.

For an even-money bet that you win with probability p and lose with probability q = 1 − p, the Kelly fraction is p − q. Rotando and Thorp use the example of a 53% win rate: f* = 0.53 − 0.47 = 0.06, so you'd bet 6% of capital each time.

The growth-rate curve around that point is the part people miss. Taking the same 53% win rate from Rotando and Thorp's example, the expected log growth per bet is p × ln(1 + f) + q × ln(1 − f). Working that through:

Fraction bet          Expected log growth per bet
 3%  (half Kelly)      0.00135
 6%  (full Kelly)      0.00180
 9%  (1.5x Kelly)      0.00135
12%  (2x Kelly)       -0.00002  (just below zero)

Half Kelly keeps about three-quarters of the growth rate of full Kelly while staking half as much on each bet. Betting twice Kelly takes growth down to roughly zero. MacLean, Thorp and Ziemba state the general version: the growth rate falls to the risk-free rate when you bet exactly twice the Kelly wager, and it never pays to bet more than Kelly because risk rises while long-run growth falls. Rotando and Thorp make the same point for their example, where the growth rate reaches zero at a fraction of about 12%, and anything above that drives wealth towards zero over time.

Why full Kelly is dangerous with real trading data#

Kelly assumes you know your edge exactly. In trading you only have an estimate, usually from a backtest, and backtests tend to flatter the strategies that survive your testing.

Run the numbers on a modest mistake. Your backtest says you win 53 of every 100 even-money trades, so you size at Kelly, but the true rate is 51 in 100:

Kelly you used:        0.53 - 0.47 = 0.06
Kelly you needed:      0.51 - 0.49 = 0.02   (you're at 3x Kelly)
Growth at 0.06, p=.51: 0.51*ln(1.06) + 0.49*ln(0.94) = -0.0006 per bet
Growth at 0.02, p=.51: +0.0002 per bet

The strategy still has a real edge, yet at that size your capital tends to shrink over time. This is the overbetting problem the Kelly literature warns about, and a two-point error in the win rate is enough to trigger it.

MacLean, Thorp and Ziemba are blunt about this. Expected-log calculations are extremely sensitive to errors in the mean, they cite work finding errors in means about 20 times as important as errors in variances, and they conclude that estimates must be accurate and, to be safe, bet sizes should be reduced. They also list the main disadvantage directly: Kelly's suggested wagers can be very large, so the criterion can be very risky in the short term. Even half Kelly isn't a guarantee: in one of their examples, $1,000 can still become $145.

The usual response is fractional Kelly, which the paper describes as a blend of Kelly and cash: you give up some growth for more security, and treat the full Kelly figure as a ceiling rather than a target. If your inputs come from a backtest, it's reasonable to assume the true edge is smaller than measured and size from that.

Putting the three together in a trading system#

These methods answer different questions, so many systems combine them rather than choose one.

Volatility targeting decides relative size: how much of each instrument to hold so a quiet bond future and a jumpy small-cap stock contribute comparable risk. Fixed fractional sets the risk budget per trade or per day, in money you can afford to lose. Kelly, used cautiously, gives you an upper bound: if your sizing implies more than the Kelly fraction for your estimated edge, you're taking risk that can only lower long-run growth.

In code, that tends to look like a sizing function that computes a volatility-scaled target, clips it to a maximum leverage, checks the planned loss against a per-trade risk limit, and refuses to go above a fraction of an estimated Kelly bet. When we build trading automation at TheAILAB, this sizing layer sits in front of the order logic and has its own tests, because a bug here can lose more money than a bad signal.

Start smaller than the maths says you can. Check how your chosen fraction would have handled the worst losing streak in your backtest, then assume live trading will produce a worse one, and size so you'd still be trading afterwards.

This article is for education only and isn't financial advice. Trading involves risk of loss.

Sources

1. A New Interpretation of Information Rate — J. L. Kelly Jr. — Bell System Technical Journal, Vol. 35, No. 4 (July 1956) (checked 2026-10-14)
2. The Kelly Criterion and the Stock Market — Louis M. Rotando & Edward O. Thorp — The American Mathematical Monthly, Vol. 99, No. 10 (1992) (checked 2026-10-14)
3. Good and bad properties of the Kelly criterion — Leonard C. MacLean, Edward O. Thorp & William T. Ziemba (2010) (checked 2026-10-14)
4. The Impact of Volatility Targeting — Harvey, Hoyle, Korgaonkar, Rattray, Sargaison & Van Hemert — The Journal of Portfolio Management, 45(1) (2018) (checked 2026-10-14)
5. Volatility Managed Portfolios (NBER Working Paper 22208) — Alan Moreira & Tyler Muir — National Bureau of Economic Research (2016) (checked 2026-10-14)