Walk-forward evaluation of a five-parameter parabolic stop-and-reversal

The stop as one searchable procedure

A five-parameter parabolic stop-and-reversal variant can be specified so that starting acceleration, increment, maximum acceleration, a noise-crossover amount, and an initial-stop offset are searched together as one procedure.

The stop update uses the prior stop plus the current acceleration factor times the gap between the extreme price and that prior stop. The acceleration factor rises only when a new extreme is made, up to a stated maximum.

A noise-crossover increment can be added so a small penetration of the stop does not reverse the position. That addition addresses whipsaws from minor price noise.

Walk-forward analysis after system optimization

System optimization can select the five inputs from an in-sample combinatorial search. That selection does not establish that those inputs will hold on later data, because the search can fit both repeatable structure and one-time noise.

A single full-sample optimization is not a substitute for repeated walk-forward checks. The best searched inputs on a fixed noisy series are guaranteed to include curve-fit noise.

An evaluation design can pair many calendar-day in-sample windows with following one-week out-of-sample windows so chosen inputs are judged on unseen bars.

Robustness testing across many weeks

Averaging many out-of-sample weekly results, rather than relying on one or two windows, is presented as the way to reduce luck in the chosen in-sample metric and to estimate an expected weekly return and its dispersion.

Editorial reading

TradersWeek editorial: treat a trailing-stop rule as a testable procedure whose shape, slope, speed, and noise filters must survive many in-sample searches and out-of-sample windows. A single optimized backtest is not that check.