Asset Selection with Machine Learning and Metaheuristic Portfolio Optimization

This work proposes a hybrid pipeline for asset selection and portfolio optimization in the stock market, combining machine learning and metaheuristics. Random Forest is used to estimate the probability of a stock achieving a minimum return of 1% over three trading days, and the selected assets are allocated using three strategies: an equal-weight portfolio, a Genetic Algorithm (GA), and GRASP. The experiments use point-in-time S&P 500 data obtained from Norgate Data, including listed and delisted assets. The eligible universe is formed monthly by the 100 most liquid assets, selected based on the liquidity observed in the previous month. The methodology separates feature selection, predictive model tuning, optimizer tuning, operational calibration, and out-of-time validation. All choices were made during the development period, from 2014 to 2018, and kept fixed in the subsequent evaluation. In the effective out-of-time period, from April 2019 to December 2025, the equal-weight portfolio based on Random Forest signals accumulated a return of 33.31%, while GA and GRASP achieved cumulative returns of 166.57% and 166.11%, respectively. Over the same period, the S&P 500 accumulated a return of 141.52%. The results show that the metaheuristic strategies outperformed both the uniform allocation of the same selected assets and the market benchmark in terms of cumulative return. In addition, the similar performance of GA and GRASP indicates that GRASP constitutes a competitive alternative to the Genetic Algorithm for allocating weights across assets in a portfolio.