Walk-forward validation: preserving the arrow of time

Prerequisite: Backtesting without hindsight shortcuts

On a small screen, scroll the figure sideways or open it full size.

Learning objective

Separate development from later evaluation and recognise the limits of repeated optimisation.

Walk-forward research repeatedly develops or selects a rule using one past window and evaluates the frozen rule on a subsequent window. The essential requirement is chronological information control. Prices, labels, parameter choices and event classifications from the evaluation window must not influence the preceding development decision.

Choose window lengths, update schedule and selection metric before comparing results. Each selected rule applies only to its next unseen segment. Combine the segment outcomes with consistent costs and capital treatment; do not retrospectively replace the weaker segments with whichever alternative later performed best.

Report sample size, turnover, drawdown, cost sensitivity and parameter stability rather than a single return number. Repeated experiments can contaminate supposedly unseen data. Preserve a final untouched holdout where feasible and label all simulations clearly. Good simulated results do not guarantee future execution or market behaviour.

Common mistakes

Do not optimise on the entire timeline and call the same history out of sample. Avoid overlapping evaluation outcomes that count identical trades twice or hiding unsuccessful parameter searches.

Practice

Draw three chronological development/evaluation folds for invented monthly data. Mark when each parameter becomes fixed and list two ways future information could accidentally leak into development.

Sources & further reading

• TradingView — Strategies: broker emulator, costs and test biases