A Hybrid Genetic Algorithm with Learning-to-Rank-to-Optimization for US Equity Portfolio Construction

This study develops and evaluates an equity selection pipeline that converts quarterly fundamentals into a monthly frequency, constructs profitability, leverage, liquidity, and growth characteristics, and learns a linear ranking model via a genetic algorithm (GA). The GA is trained to maximize either (i) mean monthly NDCG@30 using 12-tile relevance labels or (ii) mean monthly Spearman information coefficient (IC). The learned ranker is tested out-of-sample using monthly forward returns, benchmarked against the S&P 500, with different types of allocation weights, and further evaluated under sector concentration limits. In the last layer, the monthly-selected stock universe is used in a daily dynamic allocation which is solved by the penalized Max-Sharpe or Min-Variance optimization problems under only long positions and transaction fees. Performance is examined across Pre-COVID, COVID, Post-COVID (Train), and Final Test regimes, demonstrating how ranking objectives and diversification constraints impact performance and stability. Results show that TTM-based accounting signals, when optimized through genetic learning and disciplined allocation, yield economically meaningful stock selection and robust portfolio performance across market regimes.