Position sizing and risk of ruin for algo traders: fixed-fractional, Kelly and the ruin equation

A positive-expectancy EA can still go to zero

This is the part of algo trading that feels unfair. You can have a genuine edge, a profit factor of 1.5 over a thousand trades, and still lose the account, because you bet a size at which an ordinary losing streak was fatal. Edge determines whether the long run is profitable. Size determines whether you get to the long run. For a trading bot, size is one input in the settings and it is the most consequential number you will ever type.

Three ideas do all the work: fixed-fractional sizing, the Kelly criterion, and risk of ruin. Each has a formula, a worked example below, and a free calculator at /tools.

Fixed-fractional: risk a percentage, not a lot size

Fixed-lot sizing (always 0.10 lots) means your risk per trade shrinks as the account grows and balloons in a drawdown, which is backwards. Fixed-fractional sizing risks a constant percentage of current balance on each trade: lots = (balance × risk%) divided by (stop distance in points × value per point per lot).

Take a $10,000 balance, 1% risk and a 300-point stop on a symbol worth $1 per point per standard lot. Risk budget is $100; 300 points at $1 means $300 per lot at the stop, so the size is 0.33 lots. Lose, and the next trade risks 1% of $9,900. After ten straight losses you are down 9.6%, not 10%, and after twenty, 18.2%. The decay is geometric, which is exactly why fixed-fractional sizing makes literal ruin almost impossible and deep drawdown entirely possible.

Kelly: the size that maximises growth, and why you do not use it

The Kelly fraction is the risk per trade that maximises the long-run geometric growth of the account. In words: win probability minus loss probability divided by the payoff ratio. With p the win rate and b the average win over the average loss, Kelly = p − (1 − p) / b.

Worked example: win rate 55%, average win $42, average loss $38, so b = 1.105. Kelly = 0.55 − 0.45 / 1.105 = 0.55 − 0.407 = 0.143, or 14.3% of the account per trade. That is the growth-maximising size, and it is insane to run in practice. At full Kelly the expected maximum drawdown is severe and the estimate itself is built on a win rate and payoff ratio that have sampling error. Overestimate your edge by a third and full Kelly becomes over-betting, where growth turns negative.

The Risk Lab shows the Kelly percentage computed from your closed trades (it needs at least 20 trades and 15 trading days), your actual average risk per trade as a share of balance, and the ratio between them. A Kelly ratio above 1 means you are betting more than the optimum and the more you bet, the slower you grow. Most professionals run a quarter to a half of Kelly. Quarter Kelly in the example is 3.6%, and a more cautious developer would still halve that. The Kelly criterion calculator gives full, half and quarter figures.

Risk of ruin: the probability that matters

Risk of ruin is the probability that the account falls to a level you define as finished before the edge has time to work. For an even-payoff system with a fixed number of risk units, the classic formula is ((1 − edge) / (1 + edge)) raised to the number of units, where edge is win rate minus loss rate. Worked example: 55% win rate, so edge is 0.10, and 10 risk units (10% per trade). Ruin probability is (0.9 / 1.1) to the tenth, which is about 13.4%. With 20 units (5% per trade) it is 1.8%. With 50 units (2% per trade) it falls below 0.005%.

Real EAs do not have even payoffs or fixed units, so ConfirmedTrades estimates ruin by simulation instead. The Risk Lab's chance-of-ruin ladder reports how many of 1,000 Monte Carlo replays of your own trades ever dipped 10%, 25%, 50% or 80% below today's balance. The free risk of ruin calculator does a simpler simulated version from win rate, payoff ratio, risk per trade and your ruin threshold.

The Score uses this directly: 5 of the 25 risk-control points come from the probability of a 50% drawdown, scaling from full marks at 0% to nothing at 20%. An EA sized so that a 50% fall has a one-in-five chance is not being penalised for bad luck; it is being penalised for the setting.

Sizing from the worst case, not the average

Put the three ideas together into a procedure. Take the Monte Carlo 95th-percentile maximum drawdown at your current size. Decide the drawdown at which you would genuinely stop the EA. Scale risk per trade down until the simulated bad-luck drawdown sits comfortably under that line.

Say that at 2% risk per trade the 95th-percentile drawdown is 34%. Your stop-the-EA line is 25%. Drawdown scales roughly with risk per trade, so 1.4% risk brings the bad-luck figure to about 24%. You have given up some growth rate in exchange for a much higher probability of still running in a year. The compounding calculator will show you that the slower, surviving account usually ends up ahead anyway, because it never had to win 100% back from a 50% hole.

Sizing traps specific to EAs

• Martingale and grid lot progression. Doubling after a loss is the opposite of fixed-fractional, and the AI review flags it as high severity when lots rise after losses in sequence or when win rate above 65% pairs with losses twice the size of wins. It is sizing, not strategy, and it always finds the streak.
• Correlated positions. Five EAs each risking 1% on EURUSD, GBPUSD and gold are one 5% bet when the dollar moves. Size the book, not the ticket.
• No stop-loss means the stop distance is the margin call. Fixed-fractional sizing is undefined without a stop, and the review flags records where fewer than 15% of trades carry one.
• Backtest-derived win rates. Size off the forward-test numbers, degraded as described in backtest vs forward test.
• Tiny accounts and minimum lot. If 0.01 lots is already more than 1% risk at your stop distance, the account is too small for the EA, not the other way round.

Check it on the live record

Once the EA is connected, the Risk Lab and Score do this arithmetic on broker data rather than on your assumptions. Look at Kelly ratio, the chance-of-ruin tiles and Max DD (equity) together every few weeks; from backtest to live: what to monitor every week turns that into a routine.