Mitigating Overfitting on Financial Datasets with Generative Adversarial Networks

Abstract

Overfitting is an inevitable phenomenon when applying deep learning techniques to financial data, given the relative scarcity of available historical data and the ever-changing nature of financial series. This seemingly unavoidable pitfall has heavily impaired the application of many machine learning techniques, such as deep or reinforcement learning, to financial settings. Data augmentation can be an option to circumvent this, specifically generative adversarial networks (GANs). GANs, are a type of neural network architecture that focuses on sample generation. Through adversarial training, the GAN can implicitly learn the underlying structure inherent to the dynamics of financial series and acquire the capacity to generate scenarios that share many similarities to those seen in the historic time series. In this article, the authors propose a data augmentation technique using the Wasserstein GAN with gradient penalty and show how training deep models on synthetic data mitigates overfitting, improving their performance on test data when compared to models trained solely on real data.

TOPICS: Statistical methods, simulations, big data/machine learning

Key Findings

• • Generative adversarial networks can capture the complex dynamics that govern many financial assets and produce realistic synthetic scenarios based on historic data.

• • Synthetic financial scenarios can be used to enlarge training datasets in order to improve the accuracy and robustness of other deep learning models.

• • Synthetic financial series may also be applied in other tasks such as model backtesting, risk analysis, or option pricing.