Letting walk-forward optimization choose your stop
It gravitates to the tightest stop whenever the objective restates the parameter, so score in dollars at constant risk instead.
- Verdict
- Settled by the data
- Decided by
- Objective restates the parameter
- The number
- Selection bias toward the tightest grid value, every time
The hypothesis, as pre-registered
Walk-forward optimization selects robust parameters by ranking in-sample performance and testing out of sample.
What the test found
When the objective is average R and R is defined by the stop, tightening the stop mechanically raises average R without improving the strategy at all. The optimizer is not finding a better stop. It is finding a smaller denominator.
In practice this shows up as a walk-forward that keeps choosing the extreme edge of the parameter grid, run after run. That pattern is close to diagnostic: if your sweep always lands on the boundary, check whether your metric contains your parameter.
The same defect appears whenever a strategy is scored in units the strategy itself defines. Reporting in dollars at a fixed risk budget removes it.
What would make this work
The selection objective would have to be independent of the parameter being selected — dollars at constant risk, for example, rather than a multiple of the stop.
Every insight ends with this line: the exact condition that would turn the finding into a tradeable edge. That is what makes each entry a next experiment rather than a full stop.
Screened March 21, 2026
Run the arithmetic on your own numbers with the deflated Sharpe ratio calculator.
Testing something similar?
The screener puts your version of this idea through the same gate stack, on your market, and answers in seconds. Your first verdict is free.
Related insights
An edge that only exists in the last three years
Ninety-eight percent of a six-stream book's lifetime profit sat after 2022, which is why every result is now split by era before it ships.
SettledPicking the strategy with the best return-to-drawdown ratio
Return persists weakly out of sample and the drawdown ratio does not, so rank on return and correlation and let the portfolio shape the drawdown.