Validation

WF

Robust walk-forward

Why

If the exit is set only once for the whole sample, later prices may not carry it. Each fold re-selects the exit only inside the preceding training window. The exam window that follows accepts only the exit selected that time.

Formula

How the windows are cut

Only the discovery sample is cut. Its length is W, in bars. Entry, entry mode, and the filter stay fixed. What rolls is the exit.

Exam window H = floor(W / 6)
Training window = 3H

When H < 50, this item is not scored.

Fold k:
  The training window starts at bar k×H, length 3H
  The exam window follows immediately, length H
  Stop when the training window plus the exam window runs past the discovery sample

Choosing the exit in the training window

Inside the training window, rerun every row of the exit grid on the user book
Take the exit with the highest compounded return and at least 5 fills inside the window
If the compounded returns tie, keep the earlier row in the grid
The exam window does not take part in this choice

The book is the discovery-sample user book: taker fee 0.05, maker fee 0.02 (percentage points), slippage of 10 basis points, and funding of 0.0001 per 8 hours. Slippage must not be set to 0. If one row of the grid cannot run, this item is not scored. Do not run only the rows that can.

The exit in each window sees only that window's last bar and the bars before it. A fill counts as closed inside the window only if the stop, the target, or the holding-bar count expires inside the window. A fill still open when the window ends does not enter that window's score, but it occupies the book from entry until the window ends, so a later signal inside the window cannot open again.

How the exam window continues

The exam window is walked again from a flat position
Count only fills that open in this window and also close in this window
Join the fills from the exam windows in time into one series

Fills in that series ≥ 20
Compounded return of that series > 0
No liquidation

All three must hold for this item to pass. If the discovery sample cannot be cut into one full training window plus exam window, or the joined series has fewer than 20 fills, it is not scored. The sealed sample does not take part in choosing parameters, and it is not joined onto this curve.