Big Data and Machine Learning in Quantitative Investment

CHAPTER 13Deep Learning in Finance: Prediction of Stock Returns with Long Short‐Term Memory Networks

Miquel N. Alonso Gilberto Batres‐Estrada and Aymeric Moulin

13.1 INTRODUCTION

Recurrent neural networks are models that capture sequential order and therefore are often used for processing sequential data. RNNs are powerful models due to their ability to scale too much longer sequences than would be possible for regular neural networks. They suffer from two serious problems: the first has to do with vanishing gradients and the second with exploding gradients (Graves 2012; Hochreiter and Schmidhuber 1997; Sutskever 2013). Both of these are solved by the LSTM. In recent years LSTMs have solved many problems in speech recognition and machine translation, where the goal is often to match an input series to an output series. The LSTM network can be used to solve both classification and regression problems. There are two important things that distinguish these two domains in machine learning. The first is the type of the output, where in regression it takes values in the real numbers, whereas in classification it takes values in a discrete set. The second is the type of cost function used during training.

The chapter is ordered as follows. Section 13.2 presents related work on the subject of finance and deep learning, Section 13.3 discusses time series analysis in finance. Section 13.4 introduces deep learning in general, Section 13.5 covers RNNs, its building blocks and methods of ...