Freeze the research procedure
Choose the training length, test length, allowed parameter values and selection criterion before inspecting the test results. At each boundary, use only information available then. A rolling training window drops old observations; an expanding window retains them. These are different experiments and should be labelled accordingly.
Specify how existing trades are handled when settings change. Closing them, keeping their original exits or transferring them to new rules produces different returns. Labels based on future outcomes must also end before the training boundary; otherwise a seemingly chronological split can still leak information.
Worked example: three six-month tests
Consider a fictional rolling design: train on months 1–24, test months 25–30; train on months 7–30, test months 31–36; then train on months 13–36 and test months 37–42. Suppose the three test windows return +4%, −3% and +5%, after assumed costs.
The combined evaluation return is 1.04 × 0.97 × 1.05 − 1 = 5.924%. It is neither the arithmetic sum of 6% nor an annualised return. No training-window profits belong in that figure. Include the costs of changing positions at the boundaries, and do not count a test month twice.
The procedure can also be overfitted
A walk-forward result still follows one historical path. Trying many window lengths and publishing only the winner can overfit the procedure itself. Sparse trades and overlapping holding periods reduce the amount of independent evidence. A final untouched period or prospective paper record can add evidence but cannot guarantee future success.
Keep a dated closed-bar review
For a manual closed-bar review, write the setting and decision date before examining the subsequent available daily, weekly or monthly bars. Keep that dated record outside the chart. Student Wyckoff Pro can support examination of available closed bars; the scheduling, parameter fitting and performance calculations described here are a separate research workflow.
Primary sources
Sources explain the mechanics and limitations; the editorial examples use invented numbers rather than market quotations.
Your next practical step
Explore the available closed daily, weekly and monthly bars in the analysis workspace.
Open the first-chart instructions ↗