Stock Market · Trading Systems

Curve-fitting

Curve-fitting is the most dangerous mistake in systematic trading. You tweak your strategy's parameters until it shows beautiful historical returns, then go live and watch it lose money. The strategy was never real. It was fitted to past noise.

What curve-fitting looks like

Start: Simple moving average crossover, 50-day vs 200-day. Backtest shows 8% CAGR.

Tweak: Change to 47-day vs 197-day. Backtest shows 11% CAGR.

Tweak: Add RSI filter > 55. Backtest shows 14% CAGR.

Tweak: Only trade between 10 AM and 2 PM. Backtest shows 16% CAGR.

...

Final 'optimised' strategy: 22% CAGR in backtest. Goes live → loses money.

> The more parameters you tune, the more you're fitting past noise rather than capturing genuine market behaviour.

Why it happens. Overfitting in plain English

Past market data has random patterns. The more rules you add, the more closely you can match those random patterns. But random patterns DON'T REPEAT, so your future returns are random.

Signs you're curve-fitting

• You've tested 50+ parameter combinations
• Your strategy has 8+ rules
• Tiny parameter changes drastically change results (15-day MA = profitable, 16-day MA = unprofitable)
• Strategy looks too good (40%+ CAGR, no drawdowns)

Defenses against curve-fitting

1. Keep rules MINIMAL (3-5 max)

2. Test parameters on 'in-sample' data (2018-2022)

3. Validate on 'out-of-sample' data (2023-2024)

4. If out-of-sample performance is dramatically worse than in-sample, you've curve-fitted

5. Robust strategies show similar performance across small parameter changes (10-day and 12-day MA both work)

Few rulesrobust

Many rulesfragile

Looks too good to be trueis too good to be true

Takeaway. Curve-fitting is tuning strategy parameters until backtest looks great. Capturing past noise instead of real edge. Keep rules minimal (3-5), validate on out-of-sample data. Strategies looking "too good" almost always are.