Walk-forward optimization

The problem it solves

Optimising a strategy's parameters on a block of history and reporting the best result is not analysis, it is a search for the settings that best fit that block's noise. The resulting number has no predictive value, and it is invariably impressive.

A single in-sample/out-of-sample split improves on that: fit on the first 70%, score on the last 30%. But it produces one out-of-sample number from one period, and that period had its own character. A strategy that happened to suit those particular months looks validated when it is not.

How walk-forward works

Instead of one split, roll a pair of windows through the history. Fit on months 1-6, score on month 7. Move forward: fit on months 2-7, score on month 8. Repeat to the end.

You end up with a series of out-of-sample results, each produced by parameters chosen using only data that preceded it — which is exactly what live trading does. Stitched together, those results form an equity curve the strategy could actually have achieved without knowing the future.

The key property is consistency. One good out-of-sample window is luck. Eight in a row, with parameters that stay in the same neighbourhood as the window rolls, is evidence of something structural.

Anchored versus rolling windows

Neither is correct in general. Rolling suits markets whose behaviour genuinely changes; anchored suits strategies that should not be chasing recent conditions. Running both and comparing is more informative than choosing one.

• Rolling — the fit window is a fixed length that moves forward, so old data drops out. Adapts faster to a changing market and uses less history per fit.
• Anchored — the fit window starts at the same point and grows, so all history is always included. More stable parameters, slower to respond to genuine change.

Reading the result

Walk-forward efficiency is the ratio of out-of-sample performance to in-sample performance. Close to 1 means the strategy performs live roughly as well as it did on the data it was fitted to. Well below 1 means the fit was largely noise.

There is no universal threshold, and any article quoting one is inventing it. What matters more is whether the out-of-sample windows are consistently positive rather than carried by one exceptional period, and whether the chosen parameters stay in a similar region as the window rolls. Parameters that jump wildly between windows are a strategy with no stable optimum, which means there was nothing to find.

What walk-forward still cannot catch

This is the part usually left out, and we have a specific example of it.

A configuration adopted here passed a proper out-of-sample test with a profit factor of 1.474 over 40 trades on data it had never seen. Its live profit factor since was 0.518. The out-of-sample window was genuinely unseen — the process was not cheating.

The gap is that a search over hundreds of combinations will find something that fits two stretches of the same noise. Splitting the data does not prevent that; it only makes it harder. The additional check that catches it is looking at the parameter values around the winner: a real edge sits on a plateau where nearby settings also work, and an artefact is a lone spike surrounded by losses.

So walk-forward is necessary and not sufficient. Run it, and then check the neighbourhood, and then paper trade — each stage rules out something the previous one cannot see.

How many windows

Enough that each out-of-sample window contains a meaningful number of trades, and enough windows that the pattern is not one period's story. Those two demands pull against each other on a fixed amount of history, and the resolution is usually more history rather than smaller windows.

If a window produces fifteen trades, its profit factor is arithmetic rather than evidence. Widen the window or accept that the test cannot answer the question with the data available — which is itself a useful finding.

Common questions