Walk-forward analysis: testing a strategy the way you would actually trade it

The mechanic

Cut history into consecutive blocks. Train the strategy on block one, then run it forward on block two without changing anything. Slide the whole thing one block ahead: train on block two, test on block three. Repeat until you run out of data.

Every test segment is genuinely unseen at the moment the parameters were chosen, because the parameters were chosen before that segment existed. That is the entire point, and it is the only property that makes the result worth anything.

What it looks like when a strategy is fragile

A robust strategy produces a row of out-of-sample segments that are mostly positive, with a couple of flat or mildly negative ones. The magnitudes vary — they always do — but the sign is consistent enough that you would not have abandoned it in month two.

A fragile strategy looks different in a way that is easy to spot once you know to look for it: the in-sample segment is excellent and the out-of-sample segment that follows is poor, over and over. The strategy is not failing randomly. It is failing specifically on data it has not seen, which is the signature of memorisation.

How many windows is enough

The honest answer is: enough that the result is not dominated by one segment. Three windows means one bad segment is a third of your evidence. Five to ten is a more reasonable starting point for a strategy with a few hundred trades overall.

There is a real tension here. More windows means shorter test segments, and short segments mean fewer trades, and few trades mean the segment's result is mostly noise. So you are trading one kind of uncertainty for another. The way out is to look at the trade count per segment and ignore any window that does not contain enough trades to be interpretable — usually anything under about thirty.

The trap: re-optimising every window

Walk-forward analysis is often implemented as 're-optimise the parameters on each training block, then apply them to the next block'. Done well, that is a genuine test of a process. Done carelessly, it is a machine for producing good-looking results, because you are still searching — just searching in a way that is harder to see.

If your walk-forward requires re-optimising to work at all, ask what that means. It means the strategy has no stable parameters, and that you are relying on the optimiser to keep finding the current regime. That can work, but it is a different claim from 'this strategy is robust', and it should be stated that way.

Reading the output without fooling yourself

• Look at the sign pattern across windows before you look at any aggregate number.
• Check the trade count in each test window; discard windows too small to mean anything.
• Compare the average out-of-sample result to the average in-sample result. A large drop is not automatically fatal, but it is the number to explain.
• Check whether the strategy survives if you remove its single best window. If one window carries everything, you have one observation, not ten.
• Repeat the whole exercise on a different instrument. Real behaviour leaves a trace somewhere else.

Why we report it separately from the headline backtest

A walk-forward result and a full-sample backtest answer different questions, and averaging them into one number loses exactly the information that was worth having. That is why our strategy pages keep the checks separate and let you require the ones you care about.

Neither number is a forecast. Both are measurements of the past under different assumptions about what you knew when. The walk-forward one is simply closer to the situation you will actually be in.