If Your Pattern Has a 70% Win Rate, Something Is Probably Wrong

Everyone who starts backtesting goes through the same phase. You find a pattern with a 73% win rate over two years of data and feel like you’ve cracked it. You start imagining what your account will look like in six months. Then you run it live and it wins 48% of the time.

This is not bad luck. It is almost always overfitting, and a suspiciously high win rate is usually the first sign.

What overfitting actually means

Overfitting happens when your pattern has learned the specific quirks of the historical data you tested it on rather than a genuine repeating behaviour in the market.

Think of it this way. Markets are generated by millions of participants making decisions. Some of those decisions create real, repeatable patterns — institutions accumulating positions at key levels, liquidity grabs before reversals, momentum continuation after breakouts. These are real because the human and institutional behaviours that create them are real and consistent.

But within any historical dataset, there is also noise. Random clusters of events that happened to follow each other during that specific period. A spike in gold on a Tuesday in March 2022 because of a specific macro event. A strange RSI divergence during a holiday week. These things are in your data, but they will not repeat in the same form.

When a pattern has a 70%+ win rate on historical data, it usually means it has found a combination of conditions that happened to precede wins during that specific period of time, including those noise clusters. The more conditions you combine, the more specific your pattern becomes, and the more likely it is that you’ve accidentally described the noise rather than the signal.

Why high win rates are mathematically suspicious

With a well-constructed pattern and a realistic reward-to-risk ratio (say 2:1), you do not need a 70% win rate to make money. At 2:1 RR, breakeven is around 34%. A genuine edge might show 52-60% in backtesting — enough to be meaningfully above breakeven, but not so high that it strains credibility.

Think about what a 70%+ win rate would mean in practice. It would mean the pattern correctly predicts direction nearly three times more often than it’s wrong. In a market with professional participants on the other side of every trade, that is an extraordinary claim. Markets are not that readable. If they were, every quant fund in the world would have already arbitraged that edge away.

When you see a number that extraordinary, the right reaction is suspicion, not excitement.

The train/test split reveals it immediately

The most direct way to diagnose overfitting is to split your data and test the pattern on data it has never seen.

Split your historical bars into 60% for discovery and 40% for validation. Run the pattern on the first 60% and note the win rate. Then run the exact same pattern, unchanged, on the remaining 40%.

What happens to overfit patterns at this point is consistent: the win rate collapses. A pattern with 71% on training data might show 49% on validation. Or it triggers so rarely on the new data that you cannot even measure it — it only fired because of specific conditions that existed in the training period and not after.

A pattern with a genuine edge will show lower performance on validation data (this is normal — the training set was used to find it, so it will always look better there) but it will still show a meaningful positive result. If it was showing 58% in training and drops to 52% in validation, that’s a pattern worth keeping. If it drops from 71% to 49%, you have found noise.

This is exactly why the validation step is the only result that matters when deciding whether to run something live. Training results are for discovery. Validation results are for decisions.

The overfitting traps that are easy to fall into

Too many conditions. The more conditions a pattern requires, the more specific it becomes. A two-condition pattern might fire 80 times on your data. Add a third condition and it fires 30 times. Add a fourth and it fires 9 times. With 9 trades, your win rate is almost meaningless — 6 wins out of 9 is 67%, but that difference from 50% is well within random chance. Patterns need enough trades to be statistically meaningful. Under 30 trades is shaky. Under 20 is basically noise.

Fitting to a specific market regime. If you trained on 2020-2021 data, you trained on one of the most consistently trending, high-volatility periods in recent memory. Patterns that worked in that environment were shaped by that regime. The market in 2023 or 2024 looks completely different. A pattern that needed strong trends to work will fail in ranging conditions.

Not accounting for trading costs. A pattern that shows 71% win rate might flip to losing once you subtract spread, commission, and slippage. Real backtesting includes realistic costs. If your pattern only works on paper-perfect fills at the exact candle close price, it will not work live.

Cherry-picking your indicators. If you test 50 indicator combinations and pick the one with the highest win rate, you have not found an edge — you have found the one that fit the data best out of 50 tries. This is the core of overfitting. The more combinations you test, the higher your false discovery rate.

What a healthy pattern looks like

A properly validated pattern tends to have these characteristics:

• Win rate between 50% and 62% on unseen validation data
• At least 30 trades triggered in the validation period
• Performance that degrades gracefully from training to validation (not collapses)
• A positive result even after accounting for realistic spread and slippage
• Behaviour that makes intuitive sense — you can explain why this pattern should work, not just that it did

That last point matters more than people think. If you cannot articulate the market logic behind a pattern — why would price do this thing after these conditions? — then you have probably found noise. Real edges have mechanisms. They work because of something specific about how participants behave.

An RSI oversold + lower Bollinger Band touch pattern makes sense: price reached a statistical extreme, momentum is stretched to the downside, and institutions with liquidity are likely to step in. You can explain it. A pattern like “RSI below 32.7 AND it’s a Wednesday AND the previous candle was bearish” might have a great backtest, but it has no mechanism. It found noise.

The 70% rule of thumb

A simple heuristic: if your backtested win rate is above 65% on the training data, don’t deploy it live before validation. If it’s above 70%, be actively skeptical. Run the validation. Check how many trades it triggered. Check whether the win rate makes sense for the RR ratio you’re using. Check if you can explain the mechanism.

Most of the time, the validation test will do the work for you. Overfit patterns reveal themselves immediately when they meet data they haven’t seen. The ones that survive deserve your attention. The ones that don’t saved you real money by failing on historical data instead of live.