On our data, at least 3:1. Running identical XAUUSD entries at every reward-to-risk from 2:1 to 5:1 with costs included, everything below 3:1 lost money despite having the highest win rates. Profit factor rose from 0.87 at 2:1 to 1.12 at 3.5:1 — the target, not the entry, decided profitability.
We took one strategy — a liquidity sweep fade on XAUUSD — and changed exactly one variable: how far away the take-profit sat, expressed as a multiple of the stop distance. Same entries, same stops, same 60 000 bars of M15 data spanning 2.54 years. Every run was charged 30 points round-trip spread subtracted from every trade.
Holding everything else constant is the point. Most "risk-reward" advice is asserted rather than measured, and it is impossible to know whether a change helped if you alter the entry at the same time.
Net of costs, across 173-odd trades per run:
| Reward:Risk | Trades | Win rate | Profit factor | Expectancy/trade | Total P&L |
|---|---|---|---|---|---|
| 2:1 | 174 | 32.2% | 0.87 | $-0.35 | $-61.18 |
| 2.5:1 | 174 | 28.7% | 0.96 | $-0.11 | $-19.96 |
| 3:1 | 173 | 26.6% | 1.02 | $0.07 | $12.3 |
| 3.5:1 | 173 | 25.4% | 1.12 | $0.37 | $63.53 |
| 4:1 | 173 | 23.1% | 1.12 | $0.36 | $62.58 |
| 5:1 | 172 | 19.8% | 1.13 | $0.41 | $70.83 |
At 2:1 the strategy was right 32.2% of the time — the best accuracy in the table — and finished $61.18 down. At 5:1 it was right just 19.8% of the time and finished up $70.83.
This is the single most counter-intuitive result in retail trading, and it is pure arithmetic. Every reward-to-risk ratio implies a minimum win rate you must clear to break even:
| Reward:Risk | Break-even win rate needed |
|---|---|
| 1:1 | 50% |
| 1.5:1 | 40% |
| 2:1 | 33.3% |
| 3:1 | 25% |
| 3.5:1 | 22.2% |
| 4:1 | 20% |
At 2:1 you need 33.3% to break even and the strategy delivered 32.2% — a fraction under, and once spread is charged, clearly under. At 3.5:1 you need about 22% and it delivered 25.4% — comfortably over.
The entries never changed. What changed was whether the winners were large enough to pay for the losers. This is why we treat "what's your win rate?" as close to a meaningless question and ask for profit factor instead.
5:1 produced the best total, but we run 3.5:1. The results from 3.25 to 4.0 formed a broad plateau of similar profit factors, and picking the centre of a plateau is materially more robust than picking a single peak.
A lone peak in an optimisation is the classic signature of curve fitting: it works on this sample and nowhere else. A plateau says the strategy is insensitive to the exact value, which is what you want out-of-sample.
There is a practical reason too. At 5:1 the win rate falls to 19.8% — four losses for every win — and long strings of losses become more frequent and harder to sit through.
Wide targets mean fewer winners, longer holds, and more trades that go most of the way before reversing. Psychologically this is far harder than a 70%-win-rate scalping approach that quietly loses money.
Our worst losing streak at 3.5:1 was 11 trades. If your position size makes eleven consecutive losses intolerable, you will not execute the strategy that the backtest describes — you will execute a different, worse one. Sizing is covered in risk, properly.
Do not adopt 3.5:1 because we use it. The correct ratio depends on your entry's actual hit rate — the point is the method: hold the entry constant, vary only the target, measure net of costs, and choose from a plateau rather than a peak.
If your strategy cannot clear its break-even win rate at 3:1 or better, the honest conclusion is usually that the entry has no edge, not that you need a different target.
Not in our test. At 2:1 the strategy had the highest win rate (32.2%) and still lost $61.18 net of costs.
25% to break even before costs, so realistically a little over that once spread is charged.
No. Results flattened into a plateau above 3.25:1, and win rate keeps falling — at 5:1 it was only 19.8%, which is very hard to execute consistently.
Because a single optimal peak is usually curve fitting. A plateau centre generalises better to unseen data.