Walk-forward validation, or: how to tell a backtest is lying

Here is how most published backtests are made: try strategies on ten years of history until one shows a beautiful curve, then publish that curve. The result isn't evidence — it's survivorship. With enough attempts, random noise will hand you something spectacular, and it will fall apart the day real money touches it.

Walk-forward validation breaks the trick. The strategy's rules are fixed using one period of history, then tested on a later period it has never seen — walked forward through time, out-of-sample at every step. Performance that survives this is evidence of an edge; performance that only exists in-sample is curve-fitting wearing a costume.

Two more tells separate honest tests from theater. First, costs: every simulated trade should pay realistic spread and slippage, and fill at the next bar's open rather than the signal price — frictionless backtests overstate results by amounts that routinely flip profitable into unprofitable. Second, coverage: a strategy tested as one portfolio, with its real position limits (for us, one open trade per instrument), behaves nothing like the same signals tested in isolation.

This is the process behind our instrument selection: only markets with positive expectancy after costs on both a multi-year daily test and a live-timeframe test are traded — which cut 100+ candidates down to the traded few. The validation figures on the track record are labeled hypothetical, because that's what any backtest is. The live record accumulating beside them is the part no simulation can fake.

Educational content, not financial advice. Trading involves substantial risk of loss. Figures referenced from the live track record change as the record grows — that's what makes them worth referencing.