Walk-forward testing when holding time is minutes
Posted: Sat Oct 31, 2026 8:21 am
Scalping has one advantage in walk-forward testing: trade count. A system firing 30 trades a day can be optimised on six weeks and then tested on the following two, and each out-of-sample slice still holds several hundred trades. Roll the window forward by two weeks and repeat.
Only the stitched out-of-sample pieces matter. That combined curve is the honest one, because every trade in it was taken with settings chosen before the data existed. Compare its net result with the in-sample figure. If out-of-sample profit is under half of in-sample, the optimiser was mostly chasing noise.
Two traps. First, window length against regime: scalpers are sensitive to volatility, so an in-sample window shorter than a typical regime just fits the last mood. Second, schedule honesty. If nobody would actually re-tune every two weeks in live trading, do not test that schedule. The walk-forward should mimic the real maintenance routine, warts and all.
Also watch for settings that jump wildly between windows. Parameters that move from 8 to 40 and back are not discovering anything stable.
What window lengths have worked for you in walk-forward tests on systems that hold for under five minutes?
Only the stitched out-of-sample pieces matter. That combined curve is the honest one, because every trade in it was taken with settings chosen before the data existed. Compare its net result with the in-sample figure. If out-of-sample profit is under half of in-sample, the optimiser was mostly chasing noise.
Two traps. First, window length against regime: scalpers are sensitive to volatility, so an in-sample window shorter than a typical regime just fits the last mood. Second, schedule honesty. If nobody would actually re-tune every two weeks in live trading, do not test that schedule. The walk-forward should mimic the real maintenance routine, warts and all.
Also watch for settings that jump wildly between windows. Parameters that move from 8 to 40 and back are not discovering anything stable.
What window lengths have worked for you in walk-forward tests on systems that hold for under five minutes?