What is Walk-forward testing?

What does Walk-forward testing mean in practice?

An ordinary backtest uses the same data to choose parameters and to judge them, which flatters the result. Walk-forward testing separates the two: optimise on a window, then evaluate on the data immediately after it that the optimiser never saw, and repeat.

If a strategy holds up on each unseen slice, its edge is less likely to be an accident of one period. If it works only on the data it was tuned on, the tuning was fitting noise. It is one of the closest things a backtest has to an honest out-of-sample question.

A worked example

Illustrative only: tune on months 1 to 12, test on month 13. Then tune on months 2 to 13, test on month 14. A strategy that earns in the tuned window and loses in every test window fails the check.

How to read or test it

Decide the window lengths before you look at any results, for example optimise on twelve months and test on the following one, and then stick to them. Count how many of the test windows were profitable and how the losing ones compare with the winning ones. A strategy that is positive in most unseen windows and mildly negative in the rest is behaving differently from one that is excellent in two windows and ruinous in the others.

How AKQX handles Walk-forward testing

AKQX's scorecard includes a walk-forward score. A result is auto-rejected if that score is below 5, alongside the Monte Carlo and overfitting checks.

See Backtesting methodology for the feature page.

A common mistake

Running walk-forward and then re-tuning until the test windows look good, which turns them back into training data.

Related terms