Validation

AV-WF

A variant of WF. Five components are each clipped to the range 0 to 1, then added with the design weights.

Why

WF asks whether the exit, rolled forward into the later exam windows, still makes money. This variant combines return retention, the share of profitable exam windows, drawdown, parameter perturbation, and cost stress into one number.

Formula

The combination

AV-WF = 100 × (0.30 R + 0.25 P + 0.20 D + 0.15 S + 0.10 C)

The five components are each clipped to [0, 1], then added. The weights 0.30, 0.25, 0.20, 0.15, and 0.10 are uncalibrated design weights. The name is design_weights_uncalibrated. The windows, Sharpes, and net amounts all come from the walk on the discovery sample. The sealed sample is not used as the exam window here.

The five components

R = clip(SR_exam / SR_train, 0, 1)

P = clip(number of exam windows with positive net cash / number of exam windows, 0, 1)

D = clip(1 − MDD / MDD_limit, 0, 1)

S = clip(share of the perturbation grid whose exam net amount is still positive, 0, 1)

C = clip(share of the stress-cost table whose exam net amount is still positive, 0, 1)

clip(x, 0, 1) brings x into the range from 0 to 1. SR_exam and SR_train are the Sharpes of the exam window and the training window in the same walk, both on the discovery sample. MDD is the maximum drawdown. MDD_limit is the maximum-drawdown limit.

The ratio beside it

WFE = clip(SR_exam / SR_train, 0, 1)

R and WFE are the same ratio. If the in-sample Sharpe is below a floor locked in advance, or the exam-window net cash is not positive, the status of this ratio is insufficient, and it is not reported as a high score. Those two floors are not locked yet. After the ratio is clipped to 0 or 1, it cannot be treated as a high score that has already been measured.

When it is empty

If MDD_limit, the perturbation grid, or the stress-cost table is missing: AV-WF is empty
If R is empty, or P is empty: AV-WF is empty
Empty is not filled in as 1, and it is not written as 0. No star is added

A missing component is not replaced with a high default. The weights are not calibrated, and this is not the pass line of the WF star. After calibration passes and the regression test passes, this item replaces the WF star. The replacement still reads only the discovery-sample user book.