Deep Signature models for Financial Equity Time Series prediction

Abstract

We explore in this paper the use of deep signature models to predict equity financial time series returns. First, we use signature transformations to model the underlying shape of the input equity returns; further assuming the underlying shape remains the same, we predict future values based on that shape. Finally, different neural networks are used to process the output from signature transformation to predict equity returns: Long Short Term Memory Networks, Signet Model, and Deep Signature Model. Feeding signature transformations to a neural network brings significant improvement in prediction. Using signature transformation and Long Short Term Memory Networks proves to be the best performing model in accuracy and precision. In contrast, on RMSE terms, all three models offer very comparable performance.

Keywords: Deep Learning, Signatures, Deep Signatures, Machine Learning, Time Series

JEL Classification: C00, C10, C45, C50, G00, G11