Metrics: how to tell a good strategy from a lucky one
“How much does it make?” is the most common — and worst-posed — question about a trading strategy. The right version is: how much does it make, what does that cost in risk, and how much can I trust it? Three families of metrics answer those three questions. Here are the ones we work with daily — including when each of them lies.
Return: how much — and how
- Net profit (%) — the final result over a period. Necessary, but by itself the least informative: it says nothing about the path the equity took to get there.
- Profit factor — gross profits divided by gross losses. Below 1 the strategy loses; fragile systems tend to sit at 1.1–1.3; values above ~2.5 on a large sample are usually a reason for suspicion rather than joy (overfitting).
- Expectancy — the average result per trade. Useful for comparing strategies with different trading frequencies.
Risk: what it cost
- Maximum drawdown — the deepest equity fall from peak to trough. For live deployment the single most important number: this is the pain you will actually have to sit through.
- Time under water — how long it took to climb out of the hole. Two systems with the same −20 % are two different experiences when one recovers in a month and the other in two years.
- Drawdown distribution — one backtest shows one max drawdown; Monte Carlo shows which drawdowns are typical for the strategy and which are possible. Percentiles instead of one number.
- Risk of ruin — the probability of the account falling below the point of no return. Managed by position sizing — see Kelly and sizing — and monitored against guardrails continuously.
Ratios: return per unit of risk
The real comparison of strategies happens here — return by itself is a meaningless number until you know how much risk produced it:
- Calmar — annual return divided by maximum drawdown. In plain words: how many times the worst fall you will live through pays for itself. Our favourite first filter for systematic strategies.
- Sharpe — return divided by total volatility. The classic, with one flaw: it punishes upside swings too, so it under-rates strategies with occasional large wins (typically trend following).
- Sortino — the same, but punishing only downside volatility. For asymmetric strategies the more honest sibling of Sharpe.
A detail that matters: we annualize the ratio metrics over 365 days, not the textbook 252 exchange days. The markets we trade do not sleep on weekends — and a metric computed on the wrong calendar quietly lies by tens of percent.
Penalized metrics: deliberate pessimism
A backtest always carries some luck — it sees only one history, and optimization gladly claims it. That is why our favourite metrics deduct the luck in advance:
- PROM (pessimistic return on margin, Pardo) — the return recomputed with pessimistically adjusted win and loss counts: under-weight the winners, over-weight the losers, and only then compute the return. A strategy that still looks good after this deliberate handicap is telling you something.
- Composite metrics — combinations of captured market movement, consistency of results and drawdown in one score. They punish strategies that earned everything in one lucky period.
Sample quality: when to trust the metrics at all
- Trade count. Below ~30 trades every metric is chance; reliability grows with the square root of the sample. A beautiful Calmar on twelve trades is not a result, it is an anecdote.
- Market exposure — the share of time in the market. Without it returns cannot be compared honestly: 20 % a year at half the time in the market is more per unit of exposure than 30 % at full time. Details in the sizing article.
- Stability over time. Metrics computed over the whole period mask a strategy that earned everything in year one and stagnated for three. That is why we read them by windows: consistency across periods weighs more than the grand total.
A metric as a target: beware the tunnel
In genetic optimization the chosen metric becomes fitness — the number evolution breeds by. And here the old truth about measurement applies: when a metric becomes a target, it stops being a good metric. Optimizing for net profit breeds gamblers with deep drawdowns; optimizing for Sharpe breeds timid systems afraid to earn. The defence is always the same: penalized metrics, reading the whole result landscape instead of one maximum — and before anything gets near production, walk-forward and Monte Carlo, which hold a mirror up to the metrics on data the optimization has never seen.
Summary: the metrics in one place
| Metric | Family | What it says | Trap |
|---|---|---|---|
| Net profit % | return | final result over a period | nothing about risk or the path |
| Profit factor | return | gross profits / gross losses | >2.5 on a large sample = suspicion |
| Expectancy | return | average per trade | masks the spread of outcomes |
| Max drawdown | risk | deepest fall from a peak | one run = one number, not a distribution |
| Time under water | risk | how long recovery takes | often not reported at all |
| Risk of ruin | risk | chance of falling past the point of no return | depends on sizing, not just the strategy |
| Calmar | ratio | annual return / max drawdown | cyclical — depends on the period |
| Sharpe | ratio | return / total volatility | punishes gains too (trend following) |
| Sortino | ratio | return / downside volatility | sensitive to few negative days |
| PROM | penalty | return discounted for statistical luck | conservative — undervalues good strategies |
| Composite score | penalty | captured move + consistency + DD | the recipe must be fixed upfront, not tuned |
| Market exposure | sample | share of time in the market | without it returns cannot be compared |
| Trade count | sample | sample size (min ~30) | below 30 everything is an anecdote |
Want to see your strategy in metrics that do not lie? Contact us →