The pitfalls that invalidate backtests

Most impressive backtests are wrong. The reasons are well known, and avoiding them is most of the skill.

Look-ahead bias

Using information that would not have been available at decision time. The classic case: deciding to buy on today's close using today's High — which you only know after the session ends. Even subtler: fitting a scaler or indicator on the whole dataset before splitting. If any future data leaks into a past decision, the backtest is fiction.

Survivorship bias

Testing only on instruments that still exist today. Backtesting a stock strategy on the current index members ignores every company that went bankrupt or was delisted — exactly the losers your strategy would have held. Returns look far better than reality.

Overfitting (curve-fitting)

Tuning parameters until the strategy fits the historical noise perfectly. With enough parameters you can fit any past — and predict no future. A strategy with twelve optimized parameters that looks flawless in-sample is the single most common trap in retail algo trading.

Costs and slippage

Ignoring the real cost of trading. Every trade pays the spread, possibly a commission, and suffers slippage (filling at a worse price than expected, especially on stops and in fast markets). A strategy that trades often can be wildly profitable before costs and a guaranteed loser after them. Always model costs, and model them pessimistically.

The defence

Reserve out-of-sample data the model never sees during development, prefer fewer parameters, test across multiple instruments and periods, and assume your real-world results will be worse than the backtest. If they are not, be suspicious.