Walk-forward evaluation of genetic index rules

Four coupled choices

A one-day index rule search is specified as four coupled choices: how candidates are encoded, how fitness is scored, which candidates reproduce, and how mutation and crossover generate replacements. System-optimization, in that setting, is the search of a population of candidate entry, exit, and abstention rules under one stated encoding, fitness score, and variation scheme.

Locked comparisons and abstention

Candidates were limited to lists of greater-than and less-than comparisons among the prior ten sessions of high, low, and close. The search could not invent rules that subtract or divide prices. That constraint is the rule-representation: the limited language of comparisons a candidate is allowed to read from price history.

Every listed condition had to hold before a candidate issued its next-session rise forecast. On other sessions the candidate used abstention, leaving the session untraded, and its statistics were left unchanged.

Fitness, selection, and variation

Fitness is a single score that ranks candidates by combining how often they fire with the average subsequent move. In this workflow it was the product of trade count raised to 0.7 and the average subsequent index move, placing sparse large moves and frequent small moves on one scale.

Selection copied each candidate in proportion to its fitness divided by the population average, eliminating very weak candidates and duplicating very strong ones before variation. Mutation replaced one condition at random. Crossover swapped condition blocks at a split, and two split points could leave an offspring with a different number of conditions than either parent.

Windows that step forward

Rules were discovered on a 1,000-observation training window. The survivor was chosen on a separate 500-observation tuning-window, a sample unused by search and used only to pick which trained candidate proceeds to confirmation. Confirmation statistics were taken on a 10-session test window.

Those three windows then advanced by 10 sessions and the search was rerun. The 500-session and 100-session summaries were averages across 50 and 10 such shifts. Walk-forward-analysis is that retraining and reselection on windows that step forward, so confirmation never uses the observations that created the rule.

Discovery and confirmation stay apart

Discovery and confirmation used different observations because scoring a rule on the same sample that created it can overstate quality. The middle window existed to limit that overfitting before the final window was scored. Robustness-testing checks whether that same search-and-selection procedure still yields a usable signal after the windows move and after discovery and confirmation stay on separate samples.