Data-driven multiobjective algorithms : applications in portfolio optimization

Abstract

Practical portfolio optimization models have been bringing challenges that computa- tional intelligence tools are helping to solve. A class of portfolio optimization problems that

have been attracting computational intelligence applications is index tracking. The index tracking problem aims to build a portfolio that replicates the performance of a market index with a subset of assets. Recent applications of deep learning in index tracking have limited application in real environments since the proposed frameworks are not flexible to include more practical constraints and objectives. A novel application of Generative Adversarial Network (GAN) which guarantees model extension flexibility is presented. The efficiency of the GAN was evaluated considering the difficulties imposed by the combinatorial nature of the index tracking problem. We also proposed and evaluated two new metaheuristics for the index tracking model with multiple scenarios. The results showed that solving the model using GAN’s market simulations produces more stable portfolios when compared to portfolios optimized with real data. Also, the models trained in a specific rebalancing strategy could perform well in other rebalancing strategies. This work also brings discussions about problems related to the application of GANs in this context. Obtaining the optimal Pareto front in a feasible time can be impractical in multiobjective portfolio optimization with practical constraints. Another unsolved problem is the extraction of preference information to find the most preferable nondominated solution. Thus, it is interesting to consider Evolutionary Multi-criterion approaches (EMO) to find good fronts within a time constraint guided by preference information. We propose a way to learn a rough approximation of the investor’s preference model to guide the EMO search for the single most preferable portfolio and to perform preference-driven portfolio updates. This model can be obtained using Interactive Multiobjective Optimization using Dominance-based Rough Sets Approach (IMO-DRSA), which is able to guide evolutionary algorithms using a rule-based model that is refined in each interaction with the investor. The problem is that there is no evidence on how to reduce the number of representative portfolios to minimize Decision-Maker (DM) cognitive effort during the interaction, taking the satisfaction of preferences in future distributions of portfolio components returns into account. The results showed that the proposed simulated IMO-DRSA can study the impact of different variables and approaches to reduce the cognitive effort in the performance of the EMO approach to achieve and maintain good preference satisfaction over time.