A backtest that says your strategy works is a hypothesis, not a proof. Forward testing is where you find out whether it's real.
Backtesting is the practice of running a trading strategy against historical price data to see how it would have performed. It's an invaluable first step — it can quickly eliminate strategies that simply don't have positive expectancy over large samples, saving you months of live experimentation. But backtesting has a flaw its output numbers rarely admit: the data you tested on shaped the rules you're now testing.
Curve fitting: the silent failure mode
When you repeatedly tweak a strategy's parameters to improve its historical performance, you're fitting the rules to the noise in the data, not to an underlying edge. The result is a strategy that looks great on past data and struggles on new data — because the "edge" was pattern-matching to anomalies that don't repeat. Classic signs of curve fitting: too many parameters, perfect performance on a specific period followed by poor performance on adjacent periods, and results significantly better than any sensible benchmark would suggest.
Walk-forward testing
The cleanest solution is walk-forward analysis: divide your historical data into in-sample and out-of-sample periods. Optimise on the in-sample period, then test the result unchanged on the out-of-sample period. Repeat across multiple windows. If the strategy holds up out-of-sample across most windows, you have real evidence of an edge. If it degrades sharply, the backtest was overfit.
A strategy that survives out-of-sample is a candidate. One that doesn't is a lesson.
Forward testing: the real answer
Forward testing — running the strategy in real time, with real or simulated money, on data that didn't exist when you built the strategy — is the only definitive test. Paper trade for a statistically meaningful sample (minimum 100 trades for a day-trading strategy). Only if edge persists forward does it deserve real capital. Backtesting finds candidates; forward testing confirms them.
The discipline required: once you commit to forward testing, you cannot change the rules mid-test. Any parameter change resets the clock. This constraint is uncomfortable, but it's what separates data from rationalisation.
For what happens to the drawdown figure in particular, see why your backtest lies about drawdown.