A trader can lose seven of every ten trades and still have a profitable process. That statement is mathematically ordinary, not paradoxical. The important questions are how large the wins actually are, how stable those estimates are, and whether the trader survives long enough to realize them.
Start with normalized outcomes
Assume each trade begins with one unit of planned risk, called 1R. Ignore position size in dollars for now. The model has only two outcomes:
- A winning trade occurs 30% of the time and earns +3R.
- A losing trade occurs 70% of the time and loses −1R.
This binary model is intentionally clean. Actual trading produces partial wins, scratches, slippage, gaps, and losses beyond −1R. Simplicity makes the derivation visible; realism must be added before using the result.
Calculate expected R per trade
Substitute the assumed values:
E[R] = 0.90R − 0.70R = +0.20R
Each trade contributes +0.20R on average across the model's probability distribution. If initial risk were $100, the modeled expectation would be $20 per trade before any compounding. This is not a prediction that every trade pays $20. Every observed trade still returns either +$300 or −$100 in this model.
A 100-trade accounting identity
Use the expected frequencies as a simple ledger:
| Outcome | Expected count | R per outcome | Contribution |
|---|---|---|---|
| Wins | 30 | +3R | +90R |
| Losses | 70 | −1R | −70R |
| Net | 100 | +0.20R average | +20R |
The ledger illustrates expectation. It does not assert that a real 100-trade sample will contain exactly 30 wins.
Find the breakeven hit rate
Set expectancy to zero and solve for the win probability p:
At a 3R average win and 1R average loss, 25% is the breakeven win rate before costs. The modeled 30% rate has a five-percentage-point margin. That margin is not large. If implementation reduces average wins to 2.3R, breakeven rises to 30.3% and the apparent edge disappears.
Convert the edge to profit factor
The model earns about $1.29 in gross profit for each $1.00 in gross loss. Profit factor and expectancy describe the same model from different angles. Neither says how smooth the path will be.
Now calculate the noise around the edge
The second moment of this two-outcome distribution is:
Variance subtracts the square of the mean:
σ = √3.36 = 1.83R
Standard deviation of 1.83R sits beside an edge of only 0.20R. The payoff noise is roughly nine times the average. This is why short samples from a genuinely positive system can be negative.
How many trades make the mean obvious?
Under independent, stationary assumptions, the standard error of the sample mean is 1.83 / √n. A rough 95% interval stops crossing zero only when:
That is not a universal required sample size or a formal backtest verdict. It is a warning about signal-to-noise. Serial dependence, changing regimes, parameter selection, and fat tails can require much more evidence than the textbook formula.
Stress the two estimates that create the edge
| Win rate | Average win | Average loss | Expectancy |
|---|---|---|---|
| 30% | 3.0R | 1.0R | +0.20R |
| 28% | 3.0R | 1.0R | +0.12R |
| 30% | 2.7R | 1.0R | +0.11R |
| 28% | 2.7R | 1.1R | −0.04R |
| 25% | 3.0R | 1.0R | 0.00R |
Small changes matter because the edge is the difference between two much larger quantities: +0.90R of expected gross gains and −0.70R of expected gross losses. Estimation and execution errors do not need to be dramatic to consume +0.20R.
What the 30% example does and does not prove
It proves that low win rate is mathematically compatible with positive expectancy. It does not prove that any strategy advertising 3:1 reward-to-risk has an edge.
- The win rate and average payoff must come from a repeatable, executable process.
- The sample must be large and independent enough to support the estimates.
- Costs, slippage, gaps, and skipped signals must be included.
- The payoff relationship must persist outside the data used to choose the rules.
- Position size must be small enough to survive the ordinary losing path.
The equation is easy. Earning the inputs, and continuing to earn them as conditions change, is the entire problem.