Quant Analytics: from scratch to pro · Backtesting done right · Pro
Reading a backtest report: expectancy, win rate, profit factor and their traps
A backtest report is a set of summary numbers, each of which can mislead alone. Win rate and payoff only mean something together as expectancy. Profit factor, Sharpe ratio and trade count each need an uncertainty attached. This lesson shows how to read them and how to ask how long a record must be.
In this lesson
- Compute expectancy in R from a win rate and a payoff ratio
- Compute profit factor and the break-even win rate for a payoff
- Attach a standard error to expectancy and say how many trades are enough
- Compute a probabilistic Sharpe ratio and a minimum track record length
Read first: Overfitting, trial counts and the deflated Sharpe ratio, Walk-forward, out-of-sample and the locked holdout, Risk-adjusted return: the Sharpe and Sortino ratios
A report gives the reader a dozen numbers. Each one hides a way to look good. This lesson reads the main ones in pairs, attaches an uncertainty to each, and ends with the question that matters most: is the record long enough to mean anything?
Win rate and payoff belong together
The win rate is the share of trades that make money. The payoff is the size of an average winner relative to an average loser. Measuring results in R, the amount risked on the trade, gives the expectancy, the average result per trade:
expectancy = win rate x average win - loss rate x average loss
The profit factor is gross profit divided by gross loss: (win rate x average win) / (loss rate x average loss). It is above 1 exactly when expectancy is above zero, before costs.
A rule that wins 40% of its trades, with winners of 2.5R and losers of 1R, has an expectancy of 0.4 x 2.5 - 0.6 x 1, which is 0.40R. Its profit factor is 1.67. Winning less than half the time is no fault when the payoff pays for it. The break-even win rate for a payoff of 2.5 is 1 / (1 + 2.5), or 28.6%.
The trap runs the other way. A rule that wins 90% of trades with winners of 0.2R and losers of 2.5R has an expectancy of a loss of 0.07R, and a profit factor of 0.72. It feels good to run and loses money. Rules that win often and lose rarely but heavily are the usual source of these figures.
Trade count and uncertainty
An average over a few trades is noisy. For the 0.40R rule, the result per trade has a standard deviation of 1.71R. The standard error of the expectancy is that divided by sqrt(n):
| Trades | Standard error | t-statistic |
|---|---|---|
| 50 | 0.24R | 1.65 |
| 100 | 0.17R | 2.33 |
A t-statistic of 2 needs (2 x 1.71 / 0.40)^2, which is 74 trades. That assumes the trades are independent and that the 0.40R is the true value rather than the best of many searches. Overlapping trades in correlated markets are fewer independent bets than they look. The profit factor is more fragile still: one large winner can lift it, and it has no standard error in the report.
Sharpe ratio, with its sample length
The Sharpe ratio is mean excess return over its standard deviation. Report it with the sample length. The probabilistic Sharpe ratio (Bailey and López de Prado) is the probability that the true Sharpe ratio exceeds a benchmark, allowing for skewness g3 and raw kurtosis g4 (3 for a normal distribution):
PSR = phi((SR - SR*) x sqrt(T - 1) / sqrt(1 - g3 x SR + (g4 - 1)/4 x SR^2))
Here SR is per period, not annualised. The minimum track record length inverts it: the number of observations needed before the estimate can be called above the benchmark at a given confidence:
MinTRL = 1 + (1 - g3 x SR + (g4 - 1)/4 x SR^2) x (z / (SR - SR*))^2
where z is the normal quantile for the confidence level. Negative skew and fat tails lengthen it. When the number of trials is known, set SR* to the expected best of N unskilled trials; the result is the deflated Sharpe ratio.
Worked example
A daily Sharpe ratio of 0.08 (1.27 annualised, 0.08 x sqrt(252)), over 500 days, skew -1, raw kurtosis 6. The adjustment term is 1.043. The z-score is 0.08 x sqrt(499) / 1.043, which is 1.71, and the probabilistic Sharpe ratio against zero is 95.7%.
For 95% confidence (z of 1.645 one-sided), the minimum track record length is 1 + 1.043^2 x (1.645 / 0.08)^2, which is 461 days. The record of 500 days clears it, with little room. If the rule was picked from many trials, the benchmark rises and the record needed grows.
A reading checklist
- Win rate and payoff stated together, and the expectancy in R.
- Number of trades, and independent bets behind them.
- Costs included, with the cost per trade.
- Sharpe ratio with sample length, skew and kurtosis.
- Number of trials behind the headline result.
- Out-of-sample or holdout result shown apart from the in-sample one.
Common mistakes
- Quoting a win rate without the payoff, or a payoff without the win rate.
- Reading a profit factor from 20 trades as a property of the rule.
- Annualising a Sharpe ratio before applying the formulas above, which need per-period values.
- Ignoring skew. A smooth record of small gains and a rare large loss has a high Sharpe in a sample that lacks the loss.
- Comparing reports with different cost, sizing or sampling conventions.
In practice
Ask of every report: what is the expectancy, how many trades stand behind it, how many variants were tried, and how long must the record be? Research on FinUties is where studies built on methods like these are published, with their method and limits stated.
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| win rate | 0.4 | 40% |
| average winner in R | 2.5 | 2.5R |
| expectancy per trade in R | p_win * w_r - (1 - p_win) * 1 | 0.40R |
| profit factor | (p_win * w_r) / ((1 - p_win) * 1) | 1.67 |
| break-even win rate for a payoff of 2.5 | 1 / (1 + w_r) | 28.6% |
| expectancy of a 90% win rate rule | 0.9 * 0.2 - 0.1 * 2.5 | 0.07R |
| its profit factor | (0.9 * 0.2) / (0.1 * 2.5) | 0.72 |
| standard deviation of the result per trade | sqrt(p_win * (w_r - expect)**2 + (1 - p_win) * (-1 - expect)**2) | 1.71R |
| standard error of expectancy over 50 trades | sd_r / sqrt(50) | 0.24R |
| t-statistic with 50 trades | expect / se50 | 1.65 |
| t-statistic with 100 trades | expect / (sd_r / sqrt(100)) | 2.33 |
| trades for a t-statistic of 2 | (2 * sd_r / expect)**2 | 74 trades |
| daily Sharpe ratio | 0.08 | 0.08 |
| the same, annualised | sr_d * sqrt(252) | 1.27 |
| adjustment for skew -1 and kurtosis 6 | sqrt(1 - (-1) * sr_d + (6 - 1) / 4 * sr_d**2) | 1.043 |
| probabilistic Sharpe z-score | sr_d * sqrt(500 - 1) / adj6 | 1.71 |
| probabilistic Sharpe ratio against zero | phi(z_psr6) | 95.7% |
| minimum track record length, 95% confidence, benchmark zero | 1 + adj6**2 * (phi_inv(0.95) / sr_d)**2 | 461 |
Check yourself
A rule wins 40% of its trades. Winners average 2.5R and losers lose 1R. What is the expectancy per trade?
0.4 x 2.5 - 0.6 x 1 = 0.40R. Each trade is expected to add 0.4 times the amount risked.
A rule wins 90% of the time, with winners of 0.2R and losers of 2.5R. Is that a good record?
No. Expectancy is 0.9 x 0.2 - 0.1 x 2.5 = a loss of 0.07R per trade, and the profit factor is 0.72. A high win rate says nothing alone.
What win rate is needed to break even when winners are 2.5 times the losers?
1 / (1 + 2.5) = 28.6%. Below it the rule loses before costs, above it the rule gains.
The 0.40R rule has a per-trade standard deviation of 1.71R. How many trades give a t-statistic of 2?
n = (2 x 1.71 / 0.40)^2 = 74 trades. With 50 trades the t-statistic is 1.65.
A daily Sharpe ratio of 0.08 over 500 days, skew -1, kurtosis 6. What is the probability that the true Sharpe is above zero?
The adjustment is 1.043, the z-score is 0.08 x sqrt(499) / 1.043 = 1.71, and the probabilistic Sharpe ratio is 95.7%.
Key terms
Sources
- The Sharpe Ratio Efficient Frontier. Journal of Risk 15(2), 3-44 (David H. Bailey, Marcos López de Prado), 2012. Accessed 9 Nov 2026
- The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality. The Journal of Portfolio Management 40(5), 94-107 (David H. Bailey, Marcos López de Prado), 2014. Accessed 9 Nov 2026
- The Statistics of Sharpe Ratios. Financial Analysts Journal 58(4), 36-52 (Andrew W. Lo), 2002. Accessed 9 Nov 2026