GHENet: Attention-based Hurst exponents for the forecasting of stock market indexes

Abstract

Financial forecasting is a challenging and important task, with several different approaches being explored, including deep learning methods. However, most existing deep learning approaches focus on price data and traditional technical indicators. The highly complex nature of financial time series suggests potential benefits from non-linear dynamics tools. Based on that, here we propose GHENet, a model that injects non-linear dynamics information, via generalized Hurst exponents, into a deep learning predictor. To leverage the power of the Hurst features, we process them by a self-attention module, which allows the model to attend the most relevant features. The performance of our method is investigated in the forecasting of several world-wide stock market indexes and in a trading simulation. GHENet outperforms other state-of-the-art approaches, including complex deep learning models and methods that inject exogenous variables into the data. Our proposal also demonstrates to be tolerant to hyperparameter tuning, which facilitates its use “out-of-the-box”.