The short answer
Backtests fail live mostly because they were too optimistic about costs, used information that wasn't really available in the moment, were tuned too tightly to the past, or were tested in market conditions that have since changed. Each one quietly inflates the backtest relative to reality.
Costs and slippage
A backtest that ignores trading fees and assumes you always get the exact price you wanted will overstate returns - sometimes dramatically for strategies that trade often. In reality, every trade pays a fee, and the price can move between deciding and executing (slippage).
The more frequently a strategy trades, the more these small frictions compound, and the more a 'profitable' backtest can turn into a flat or losing live record.
Overfitting and look-ahead bias
If the strategy was tuned until it fit the past perfectly, it likely captured noise that won't repeat - the hallmark of overfitting. And if it accidentally used future information (look-ahead bias), its backtested entries and exits were never achievable in real time.
Both produce gorgeous backtests and disappointing live results, because the edge they showed wasn't real to begin with.
Regime change
Markets shift. A strategy that thrived in a low-volatility bull market can struggle in a choppy or falling one. A backtest run over a single favourable regime measures how the rule did in that weather - not how it does when the weather changes.
This is why testing across varied conditions - bull, bear, and flat - matters so much: it's the difference between a strategy that works and one that merely worked once.
How to close the gap
Steps that make a backtest more honest about live performance:
- Include realistic fees and slippage in the test, especially for high-frequency rules.
- Step through time strictly so no future information leaks in.
- Keep rules simple and test the neighbours, so you're not relying on overfit settings.
- Test across multiple market regimes, not one lucky stretch.
- Treat the backtest as a filter that rejects bad ideas, not a promise of future profit.