Walk-forward testing in plain language
Posted: Fri Oct 30, 2026 6:40 am
The idea is simple. Optimise the parameters on one chunk of data, then test them on the next chunk, which the optimisation never saw. Then slide the window forward and repeat.
Why bother? Because parameters tuned on all available data will always look good on that data. The real question is whether they work on data they've never met. Walk-forward results approximate that.
A typical setup: optimise on six months, test on the following month, shift forward one month, and repeat across several years. The combined out-of-sample results show how the process behaves in practice, including its bad stretches.
If the out-of-sample performance is far weaker than in-sample, the strategy is probably overfitted. If it holds up reasonably, confidence is justified, though never guaranteed.
A caution: the windows need enough trades to mean something. A one-month test slice with twenty trades is mostly noise, so the slices should be sized to contain at least a hundred trades where the strategy's frequency allows it.
Has any strategy been through a walk-forward process, and how big was the gap between in-sample and out-of-sample results?
Why bother? Because parameters tuned on all available data will always look good on that data. The real question is whether they work on data they've never met. Walk-forward results approximate that.
A typical setup: optimise on six months, test on the following month, shift forward one month, and repeat across several years. The combined out-of-sample results show how the process behaves in practice, including its bad stretches.
If the out-of-sample performance is far weaker than in-sample, the strategy is probably overfitted. If it holds up reasonably, confidence is justified, though never guaranteed.
A caution: the windows need enough trades to mean something. A one-month test slice with twenty trades is mostly noise, so the slices should be sized to contain at least a hundred trades where the strategy's frequency allows it.
Has any strategy been through a walk-forward process, and how big was the gap between in-sample and out-of-sample results?