Portfolio selection using artificial intelligence techniques in the mexican stock exchange

Abstract

Integrating artificial intelligence (AI) techniques in the portfolio selection problem has shown high potential for performance improvement. Even a slight improvement in the portfolio selection problem can produce significant profit for investors. This thesis project proposes an evaluation scheme using a combination of AI techniques to create portfolios for the Mexican Stock Exchange. The portfolio selection problem is divided into two stages: stock selection and portfolio optimization. In the stock selection stage, we use three labeling strategies, one of them proposed by us, to train supervised machine learning classifiers, including Support Vector Machines (SVM), Random Forests (RF), Neural Networks (NN), and PBC4cip. The classifiers perform a binary classification to filter stocks with a high potential to achieve positive returns for the next month. In the optimization stage, we use and adapt to the context a state-of-the-art multi-objective genetic algorithm (NSGA-III) to find solutions that set the participation weights for each stock. From the efficient frontier (Pareto front), we consider three portfolios for evaluation performance: maximum return (GA max return), minimum risk (GA min risk ), and average risk (GA avg risk ). For comparison purposes, we also made portfolios using all stocks and baseline optimizers; equally weight distribution (1/N), maximum Sharpe ratio (max Sharpe), and minimum risk (min risk) using analytical methods. The portfolios are created using national and international stocks that trade on the Mexican Stock Exchange. We implement an evaluation scheme that finds the combination of a set of AI techniques that produce the portfolios with better performance according to three different investor profiles: aggressive, conservative, and moderate. The portfolios are evaluated monthly from December 2019 to December 2020 using the return, risk, and Sharpe ratio as objective goals. The proposed historical and t+1 labeling strategy produced higher average returns using every classifier and optimization technique. Portfolios with the highest returns were produced by SVM and NN classifiers and GA max return optimizer, while PBC4cip and GA min risk created portfolios with the lowest risk. The results achieved have higher returns than the risk-free interest rate in Mexico (CETES) and S&P 500 and S&P/BMV IPC indexes.