Generative adversarial networks applied to financial market’s forecast

Abstract

Artificial Neural Networks have been applied to Time Series Price forecasting for many years, with results improving as computers became more powerful and new frameworks and methodologies were introduced. The advent of Generative Models (or Forward Models), culminating with Generative Adversarial Networks (GANs), in which synthetic data is generated with the same statistical characteristics as real (historical) data, have further improved price prediction results, when compared to classic neural network frameworks. Recent research using a GAN model with a single Long Short-Term Memory (LSTM) layer in the Generator and Dense layers in the Discriminator has shown that this framework is able to outperform other GAN models and classical machine learning models when applied to stock market price forecasting. In this project, we start from the model proposed by our benchmark publication (single LSTM Generator and Dense Discriminator) and explore alternatives involving the type of data provided to the Generator during training, number and type of recurrent units in the Generator, number of layers in the Generator, and Discriminator models, aiming at improving price forecasting results to be used by a simple algorithmic trading engine, and comparing simulated return on investment with return on investment of the Buy and Hold strategy. Results show that, by providing historical context data in addition to random data to the Generator during training (while assuring a balanced GAN Model, in which the relative strength of the Generator and the Discriminator are similar), simulated profitability improves when compared to simulated profitability based on predicted data from Generators trained on random data only. Additionally, increasing the number of units and the number of recurrent layers in the Generator positively affects simulated profitability, while using simpler recurrent units (Gated Recurrent Units or Simple Recurrent Units) affects simulated profitability negatively. While modifying the type of data provided to the Generator or modifying the Generator model has a large impact on simulated profitability, alternative Discriminator models tested (Convolutional, Deep, and LSTM) have little or no effect on results. Our best model provides better price forecasts when compared to the benchmark publication. Simulated profitability when using predictions of our best model outperforms the Buy and Hold strategy profitability, showing that these modifications to our models improve overall prediction effectiveness.