Genetic optimization of relative-volatility crossover rules

A three-layer contract

This article records a historical combination workflow and then, as editorial commentary, reads it as a three-layer contract. The relative-volatility-index is only a forecast-style oscillator. The trigger line is only a falsifiable moving-average-crossover. System-optimization may add sibling rules only when the fitness-product and the date evolution must stop are written down before the last holdout-window is opened.

The oscillator is a forecast layer

On each bar the relative-volatility-index is the close minus the open, divided by the high minus the low. A four-weight finite-impulse smoother is applied separately to the numerator and the denominator so shared filter lag cancels in the ratio. Cycle-length is the lookback that sets how many filtered bars enter the oscillator average. Editorial reading: this object stays a forecast-style oscillator. It is not, by itself, an entry or exit rule.

The trigger is a crossover hypothesis

The working signal is a moving-average-crossover. The position is long when a one-bar lagged trigger crosses above the live index, and short when it crosses below. Editorial reading: the trigger line is only a falsifiable crossover hypothesis. Crossing the live index does not turn the oscillator into a second search procedure.

Sibling rules under system-optimization

After informal trials settled on a cycle-length of 15, the rule set was expanded to crossovers of stochastic and adaptive versions of the same index. System-optimization breeds candidate rule-and-length combinations and scores each by the fitness-product, which is return on account multiplied by equity-curve correlation. For the stochastic variant, admissible lengths run from 2 through 16, and mutation can change those lengths so the population does not collapse to one lookback.

Early-stop and the holdout-window

Rules were generated on 2 January 2000 through 31 December 2002 and judged on 2003. Early-stop halted further evolution so the 2003 window would not be fully fitted. The single-rule crossover and the multi-rule optimized set were both applied afterward to the holdout-window of 2 January 2004 through 23 March 2005. Editorial reading: sibling rules are admissible only after that fitness-product and that early-stop date are written down, and only while the holdout-window stays closed.