Walk-forward testing on tick data without fooling yourself

[Amanhasnoname]

Walk-forward analysis divides history into segments. Optimise on one, test on the next, then roll forward. It simulates how a trader would adapt in real time.

For scalping on tick data, the setup needs care. Segments should be long enough to contain a meaningful number of trades, and also span different conditions. A month might be enough for a high-frequency system, while slower setups need several months.

The key discipline is to never look at the out-of-sample results while adjusting the process. Every time the test period influences a decision, it becomes part of the training data.

Reporting the stitched-together out-of-sample results gives a more honest picture than the best in-sample number. It's usually lower, sometimes dramatically so.

If the walk-forward result is much worse than the optimised one, the optimisation was fitting noise. If it's similar, there's a better chance the edge is real.

Computation time is the main cost, since tick tests are heavy. Planning the segments carefully saves a lot of waiting.

Do you use walk-forward testing, and how large are your segments?