Sharpe Ratio and the Backtest Metrics That Actually Matter

A backtest report is a wall of numbers, and most people read it the same way: eyes go to total return, then to win rate, then to the equity curve, and the decision is made before the third row. That reading order is close to the worst one available. Total return is the number most easily manufactured by a bull market, and win rate is the number most easily manufactured by a wide stop.

This guide covers how to read backtest results in an order that resists that: what the Sharpe ratio measures, what a good one looks like in a backtest, why the deflated Sharpe ratio exists, the six numbers to read in order, and the metrics that mislead on their own.

What the Sharpe ratio measures#

The Sharpe ratio, introduced by William Sharpe in the 1960s, is return per unit of risk. Take the strategy's return above a risk free rate, divide by the standard deviation of its returns, and annualise. A strategy that earns 12% a year with 8% volatility has a Sharpe of 1.5. One that earns 30% with 40% volatility has a Sharpe of 0.75. The second made more money and is the worse strategy by this measure, because it took far more risk to do it.

That is the whole point of the ratio: it makes a calm strategy and a wild one comparable, and it stops a large return from hiding the size of the swings that produced it.

Two caveats before you trust one in a backtest. It treats upside and downside volatility identically, so a strategy with occasional large winning months is penalised for them; Sortino (downside volatility only) and Calmar (max drawdown) tell a truer story in that case, and the metrics reference covers both. And a Sharpe from a short backtest is a guess with a wide error bar, whatever it says.

Sharpe ratio metrics: what is a good Sharpe in a backtest?#

The rule of thumb in the docs is that above 1 is good and above 2 is strong. That is fine for a first triage and dangerous as a final answer, because the same Sharpe can mean very different things depending on where it came from.

Consider two illustrative results, both showing a Sharpe of 1.4. Strategy A has 38 trades over eight months, all in one rising market. Strategy B has 310 trades over five years that include a crash, a recovery and two long sideways stretches. The number is identical. The evidence is not. Strategy A's Sharpe is one draw from a wide distribution, and the next eight months could easily produce 0.2. Strategy B's Sharpe has been tested by enough trades and enough regimes that a repeat somewhere near 1.4 is plausible.

Two adjustments help. Costs first: a Sharpe of 1.4 before fees, spread and slippage is not a Sharpe of 1.4, and a backtest that reports gross returns is reporting a number that never existed. Then sample shape: the probabilistic Sharpe ratio (PSR) asks how confident you can be that the true Sharpe is above zero, given how short the sample is and how fat tailed the returns are. A raw 1.4 on a short, skewed sample can carry a PSR that says "maybe," and that is the honest reading.

So a good Sharpe is one computed after costs, on a sample long and varied enough that its error bar is narrow, and that survives the correction in the next section.

The deflated Sharpe ratio#

Here is the problem the raw Sharpe cannot see. Before you found this strategy, how many did you try?

If you tested one idea and it produced a Sharpe of 1.2, that is interesting. If you tested two hundred variations (lookbacks, stops, timeframes, tickers) and this was the best one, a Sharpe of 1.2 is roughly what the winner would show even if none of the two hundred had any edge. The best of many random tries always looks good, and the raw Sharpe reports that luck as skill.

The deflated Sharpe ratio, from work by David Bailey and Marcos López de Prado, corrects for this. It estimates the Sharpe you would expect the best of N zero skill trials to reach, given the spread of results across the trials, the sample length and how far from normal the returns are, and credits your strategy only for what it delivers above that bar. The output is a probability that the observed Sharpe beats the lucky maximum. On the platform a deflated Sharpe at or above 95% means the result survived the search, and a large drop from PSR to DSR is the signature of search luck.

Two consequences follow. A modest Sharpe from a first and only test can outrank a higher Sharpe that survived a long parameter sweep. And a tool that does not count the attempts cannot honestly grade the winner, which is why the deflated figure sits beside the raw one instead of behind it. The deflated Sharpe lesson works through the arithmetic, and the validation guide covers how it fits with the other gates.

Backtest results interpretation: the six numbers to read in order#

Every result card carries the same six headline metrics. Reading them in the right order is most of backtest results interpretation, because the early numbers tell you how much to trust the later ones.

1. Total trades. Sample size first, always. Under roughly 30 trades, every other metric on the card is noise, and a beautiful curve on 14 trades is 14 anecdotes.

2. Max drawdown. The largest peak to trough drop in equity, and the pain metric. Could you have sat through this loss live, at your sizing, without abandoning the system? If not, the return was never available to you. It is also one path through one ordering of trades; the Monte Carlo lesson explains why the drawdown you should size for is usually larger.

3. Profit factor. Gross profit divided by gross loss. Above 1.0 means the strategy made money; around 1.5 and above is healthy. Be suspicious of very high values on few trades, since one outsized winner can push a 12 trade profit factor to 4 and mean nothing.

4. Win rate, read together with expectancy. Win rate alone is meaningless. What matters is whether win rate multiplied by payoff ratio (average win divided by average loss) clears 1. A 40% win rate with winners twice the size of losers is a good trend following system. A 75% win rate with losers four times the size of winners is a slow leak. Expectancy, average profit per trade in units of risk, combines the two, and positive expectancy is the whole game.

5. Sharpe ratio. Now the risk adjusted score, read with everything above in mind: how many trades it rests on, whether costs are in, and what the deflated figure says about the search that produced it.

6. Total return. Last, and only once the first five have told you whether it is trustworthy. A large return with a huge drawdown is fragile; a large return over a rising market may be the market, not the strategy.

Here is the order applied to two illustrative strategies:

Read top to bottom, A loses. Its 22 trades cannot support any of the numbers beneath, and a Sharpe of 2.3 on 22 trades has an error bar that swallows the whole scale. Strategy B has evidence behind it: a long ledger, a drawdown that is uncomfortable but survivable, and a profit factor a 44% win rate can honestly produce when winners run. Read bottom to top, A wins on every line, which is exactly why the order matters.

Metrics that lie on their own#

Each of these is a real, correctly computed number that misleads when it is the only one you look at.

Win rate. Set a 1% target and a 15% stop and you will win most of your trades for years, right up to the month you give it all back. That is how a backtest looks reliable and trades like a time bomb. Always pair it with payoff or expectancy.

Total return. A long only crypto strategy tested on a bull market returns a lot. So does buying and holding. Return without a comparison to doing nothing, and without a bear market in the sample, measures the market's mood, not the rules.

Sharpe from a short or one regime sample. A Sharpe earned in one market mood is a description of that mood. Walk forward analysis, covered in its own guide, is the test that checks whether the number holds across eras.

Any metric averaged across directions. A strategy that is profitable long and a loser short can show a healthy blended card. The ledger split by direction separates a genuinely two sided edge from one that only works one way.

This is why a TextToQuant result does not hand you the headline numbers and leave. Every run returns a letter grade built from four pillars (profit, risk, consistency and edge), so a strategy cannot earn a high grade by maximising return while ignoring sample size or drawdown, and a verdict that can be DISCARD. Underneath sit the validation layer (walk forward out of sample, Monte Carlo, deflated Sharpe, cross market robustness, regime and context analysis), a robustness panel and an overfitting verdict. The reading a result lesson walks through a real card in this order; if you have not run one yet, start with backtesting without coding. The same card comes back for crypto, US stocks, the Egyptian Exchange, forex and metals, and the free tier is enough to try it.

FAQ#

What is a good Sharpe ratio for a trading strategy?#

As a first triage, above 1 is good and above 2 is strong, but only after costs and only on a sample long enough to trust. A Sharpe of 1.5 on 30 trades in one rising market is weaker evidence than a Sharpe of 0.9 on 300 trades across a crash and a recovery. Check the trade count and the costs, then look at the deflated Sharpe to see whether the number survives the attempts that produced it.

What is the difference between Sharpe ratio and profit factor?#

Profit factor is gross profit divided by gross loss over the whole ledger: it tells you how much the winners outweighed the losers in money terms, and nothing about how bumpy the ride was. The Sharpe ratio divides return by the volatility of returns: it tells you how much reward you got per unit of risk, and nothing about the size of individual trades. A strategy can have a strong profit factor and a poor Sharpe if the profit came in a few large, irregular bursts. Read both.

Why is my win rate high but my backtest loses money?#

Because win rate ignores the size of wins and losses. A strategy that takes small profits and wide stops wins often and loses big, and if win rate multiplied by payoff ratio comes in under 1, it bleeds no matter how many trades it wins. Check the average win against the average loss, and check expectancy in R. If the losers are several times the winners, the high win rate is the problem, not the consolation.