Why win rate is not enough

Win rate is the share of closed trades that made a profit. It answers how often a strategy won, but not how much it won or lost. A strategy can win often and lose money when its occasional losses are large. Another can lose often and remain profitable when its winners are much larger.

What win rate measures

If 60 of 100 closed trades have positive profit and loss (PnL), the win rate is 60%. Breakeven trades require an explicit convention. Tradee counts closed trades with positive PnL as wins.

The numerator and denominator contain no information about payoff size. Two strategies can report the same 60% win rate and produce opposite results. One might average +1R on winners and -1R on losers. Another might average +0.4R and -3R.

Win rate also hides the order of outcomes. Ten losses in a row and ten alternating losses create the same win rate, but different drawdown and operational demands.

Expectancy connects frequency and payoff

Expectancy is the sample's average result per closed trade:

expectancy = win rate x average win - loss rate x average loss

In this form, average loss is a positive magnitude. Tradee stores average loss as a negative number, so its equivalent calculation is:

expectancy = win_rate * avg_win + loss_rate * avg_loss

Expectancy can be positive below a 50% win rate or negative above it. It is still an estimate from the observed sample, not a promise about the next trade.

Profit factor adds another view:

profit factor = total gross profit / absolute total gross loss

A profit factor above 1 means gross profit exceeded gross loss in that sample. The result still depends on the included trades, costs, and test assumptions.

Expectancy = win rate × average win - loss rate × average loss

Frequency and payoff must be read together after costs.

The breakeven win-rate formula

Let:

• W be the average net winner, greater than zero.
• L be the absolute average net loser, greater than zero.
• p be the win rate.

At breakeven, expected wins equal expected losses:

pW - (1 - p)L = 0

Solving for p gives:

breakeven win rate = L / (W + L)

If the average winner is $100 and the average loser is $100, breakeven is 50%. If the average winner is $200 and the average loser is $100, breakeven is 33.3%. If the average winner is $50 and the average loser is $100, breakeven is 66.7%.

Use net average outcomes when possible. If W and L are gross and every trade has an average cost C, the corresponding formula is (L + C) / (W + L). Variable costs require trade-level calculation because one average cost can hide differences in size and liquidity.

The gap between observed and breakeven win rate is easier to interpret than the observed win rate alone. A 55% result is strong against a 40% breakeven rate and weak against a 54% breakeven rate. Sampling uncertainty still applies to both cases.

Two Tradee backtest cards with the same trade count

Consider two samples with 100 closed trades and the same starting budget. Both are a worked example chosen to make the arithmetic clear. They are not a recorded backtest, and the figures in this article's images are the same constructed numbers.

The scalp needs a breakeven win rate of 120 / (20 + 120) = 85.7%. Its observed 80% is not enough. The trend follower needs 70 / (180 + 70) = 28%. Its observed 38% is above that threshold.

On the Backtests hub, sorting by win rate puts the losing scalp first. Sorting by profit puts the trend sample first. Neither sort proves that the ranking will persist outside the sample.

Sampling uncertainty changes the meaning

An observed win rate is a point estimate. Ten trades with eight wins produce 80%, as do 100 trades with 80 wins. The second estimate is more precise if trades behave like independent observations with a stable win probability.

For illustration, a 95% Wilson interval is about 49.0% to 94.3% for 8 wins in 10 trades. It is about 71.1% to 86.7% for 80 wins in 100 trades. The NIST handbook explains why simple symmetrical approximations can be inaccurate for small samples or rare outcomes.

These intervals are only a diagnostic for trading results. Trades can cluster by market regime, share overlapping exposure, and change after volatility shifts. Those dependencies violate the simplest binomial assumptions and can make the effective sample smaller than the trade count suggests.

A strategy selected after many trials has another uncertainty source. The best result can win the research contest by chance. Bailey, Borwein, López de Prado, and Zhu describe this selection problem in their paper on backtest overfitting. A held-out period and a complete trial log make that process more visible.

Why chasing win rate can hurt

A common way to raise win rate is to close winners earlier and allow losing trades more room. This changes both sides of the breakeven formula. Average win falls, average loss grows, and the required win rate rises. The displayed win rate can improve while expectancy moves below zero.

Costs create the same effect. A small average winner is more sensitive to commission, spread, slippage, and funding. Cost assumptions can turn many small wins into negative net outcomes even when the gross hit rate stays unchanged.

A reproducible comparison

Use the same procedure for every candidate so a favorable strategy does not receive favorable assumptions.

1. Fix the asset, timeframe, date range, starting budget, sizing rule, commission, and slippage assumptions.
2. Freeze each strategy version before comparing results.
3. Record every closed trade's net PnL in order.
4. Calculate trade count, win rate, average win, absolute average loss, breakeven win rate, expectancy, and profit factor.
5. Record maximum drawdown, exposure, turnover, best trade, and worst trade.
6. Repeat the calculation with higher cost assumptions.
7. Evaluate the unchanged versions on a held-out date range.
8. If a strategy survives, compare its decisions on incoming data through a paper test.

For the two 100-trade examples above, every value can be reproduced from four inputs: number of wins, number of losses, average win, and average loss. A real ledger needs trade-level data because position sizes and costs vary. Keep the ledger and formulas with the result.

Reading a Tradee result

Tradee backtests place win rate beside return and versus buy-and-hold. The hub also supports filters for profit, Sharpe, and drawdown. On a strategy result, read trade count, average trade, best and worst trade, exposure, and drawdown with the win rate.

On a bot backtest, inspect the aggregate and its instances. One instance can dominate the combined result. On Execution and Analytics, paper and output books remain modeled. Their win rate does not establish exchange fills, queue position, or realized slippage.

If the goal is to test a historical rule set, start with a first backtest. If the goal is to compare the frozen bot with incoming data, use a paper pool and keep the distinction in paper trading vs backtesting.

FAQ about win rate

Can a 40% win rate be profitable?

Yes. If the average net winner is large enough relative to the average net loser, the breakeven rate can be below 40%. The observed sample can still fail to persist.

What win rate should I target?

There is no universal target. The relevant threshold depends on average payoff, costs, sizing, and the uncertainty around all estimates. A high target alone can reward fragile payoff shapes.

Methodology and sources

The formulas use closed-trade arithmetic and treat average loss as a positive magnitude unless stated otherwise. The interval examples use the Wilson score method and are rounded to one decimal place. NIST documents confidence intervals for binomial proportions. The multiple-testing caution follows the peer-reviewed discussion of backtest overfitting. This article is educational and does not provide personalized investment or trading advice.