forex-scalping.com

Compare Two Strategies Beyond the Win Rate

Strategy X wins 45 percent of the time, with an average win of 1.5R and an average loss of 1R. Its historical maximum drawdown is 8R. Strategy Y wins 35 percent of the time, with an average win of 3R and an average loss of 1R, and a maximum drawdown of 12R. Which is better?

Expectancy first. X: 0.45 times 1.5 minus 0.55 times 1, which is 0.675 minus 0.55, plus 0.125R per trade. Y: 0.35 times 3 minus 0.65 times 1, which is 1.05 minus 0.65, plus 0.40R per trade.

Over 100 trades, X earns about 12.5R, and Y about 40R.

Return relative to drawdown: X divides 12.5 by 8, which is 1.56. Y divides 40 by 12, which is 3.33. Y does better on both measures, although its drawdown is larger in absolute terms.

Now the experience. Y loses 65 percent of its trades, so losing streaks are longer. The probability of at least one run of ten losses in 100 trades at a 35 percent win rate is about 37 percent, versus about 10 percent for X at 45 percent. Y asks for more patience.

Costs also matter. If each trade has a cost of 0.1R, expectancy drops to 0.025R for X and 0.30R for Y. X nearly vanishes.

Finally, check sample sizes and robustness. If X comes from 400 trades and Y from 40, the figures carry very different weight.

Your turn: pick two setups from your journal and compare them on expectancy, drawdown, and ratio.

Nice comparison. Return per unit of drawdown also matters: X has 12.5R over 8R drawdown, Y has 40R over 12R. Y is better on both measures here.

But psychology matters too. A 35 percent win rate means long losing runs, for example ten losses in a row are quite likely over a few hundred trades. Traders who cannot follow the rules through those streaks often end up with worse results than the backtest of Y.

Take a scalping setup that fades a push into a level. Entry on the close of the signal bar, 3 pip stop, 3 pip target. Run it across a year of one-minute bars and the curve climbs in a smooth line. Win rate in the high sixties. It looks like something worth funding.

Then it goes live on a small account and the win rate sags toward the mid forties. Same rules, same pair, same sessions. What changed?

Part of the answer sits in how the tester resolved the bars. On any one-minute bar where the range was wider than 3 pips in both directions from entry, the tester had to pick which level came first. Many default tools follow a fixed convention, such as assuming the path goes open, then the nearer extreme, then the far extreme, then close. That convention is a guess dressed up as a fact. If it happens to favor the target side, every ambiguous bar becomes a win.

Here is how to test whether that is happening. Take the same rules and force the pessimistic ordering on every ambiguous bar: stop always hits first. Then force the optimistic one. The gap between those two curves is the zone of uncertainty. If the strategy only survives under the optimistic ordering, the edge was assembled out of assumptions rather than price.

A second check is to rerun a slice of the period on tick data and compare trade by trade. Even a single month will show how many bar-based fills were wrong, and in which direction they lean.

The uncomfortable part is that tight scalps are the strategies most exposed to this, because their levels live inside the bar. Wider swing logic barely notices.

What gap did you find between pessimistic and optimistic bar ordering on your own scalping rules?

The streak probabilities check out with the usual approximation. At a 35 percent win rate, the chance of a run of ten losses in 100 trades is about 1 minus e to the minus 100 times 0.65 to the tenth times 0.35, which is about 37 percent. At 45 percent, it's about 11 percent. So Strategy Y will put the trader through a ten loss streak in more than a third of hundred trade periods. That's psychologically hard, even knowing the expectancy is positive. Many traders abandon strategies like Y during exactly these streaks, which is why the experience matters as much as the numbers.

I compared two of my own setups this way last winter and the answer surprised me. One was a pullback entry with a 52 percent win rate and small winners, and the other was a breakout entry that won only 31 percent of the time but paid close to 3R. The pullback felt safer, yet after counting commission and slippage at 0.08R per trade its expectancy was nearly zero, while the breakout stayed clearly positive.

The difficulty was the experience of trading it. In my log the breakout had a run of 14 straight losses once, and I reduced size by half before it recovered, which cost me part of the gain. So I now treat the longest losing streak in the data as something I must be able to sit through before I commit real size.

I also stopped trusting any comparison based on fewer than about 100 trades per setup. With 40 trades I could make either one look better by choosing the start date.

The cost sensitivity is a strong point in favour of Y. At 0.1R per trade, X's expectancy falls from 0.125R to 0.025R, an 80 percent reduction, while Y's falls from 0.40R to 0.30R, a 25 percent reduction. Strategies with larger average wins relative to costs are more robust to cost increases. If real world costs turn out to be higher than estimated, X could easily become unprofitable while Y would still work. That's a good reason to prefer strategies with more room between their gross edge and their costs, as the cost stress thread discusses.

Return relative to drawdown is a useful ratio, but it's based on historical maximum drawdowns, which are single observations and can understate future risk. Running a Monte Carlo simulation for each strategy would give a distribution of drawdowns, and comparing the 95th percentile values would be more informative. Y's lower win rate means its drawdown distribution is likely wider, with a longer tail. That might change the comparison somewhat, though with an expectancy more than three times higher, Y would likely still come out ahead on most risk adjusted measures.

The sample size point is crucial and often overlooked. If X's statistics come from 400 trades and Y's from 40, X's numbers are much more reliable. Y's expectancy of 0.40R from 40 trades has a standard error of perhaps 0.3R, so its true value could be much lower. Comparing strategies with very different sample sizes requires caution. Has anyone here compared two strategies from their own journal this way? I'd be interested in whether the one with the lower win rate turned out to be better, and whether you were able to stick with it through its losing streaks.