Usando conexiones residuales en redes de alimentación hacia adelante de aprendizaje profundo por refuerzo aplicadas a mercados financieros

Abstract

In the last few years, reinforcement learning has seen an immense growth of research interest driven by the recent successes in performing tasks that were thought impossible for machines. This article explores the use of deep reinforcement learning (DRL) models for algorithmic trading of stock markets. Moreover, trying to leverage the benefits that residual connections brought to convolutional neural networks (CNNs) the researcher attempts to replicate the same advantages on feed forward neural networks (FFNNs). Historical data of the Alphabet stock prices was used to create simulations to train and test the DRL model. Through an iterative experimentation process different models were tested, and their hyperparameters were adjusted to maximize the accumulate rewards (profits) they obtained in the test simulated environment. Finally, a definitive experiment was conducted to compare the performance of bare FFNNs and FFNNs with residual connections. The results show that adding residual connections and augmenting the number of layers only worsened the performance of the DRL model. However, the experiments exhibit interesting aspects on the application of DRL to trade in financial markets.