Financial time series forecasting : from machine learning to deep learning

Abstract

In recent years, technological advances have had an immense effect on the way we conduct ourselves in a vast number of industries. The financial industry is one such industry which is known to have been affected in such a manner, such that almost all financial activity, be it accounting, investing, auditing, or financial modelling, are channelled through some form of technological medium. A vast range of systems have been developed, using an array of fi nancial standards and expert input, enabling users with a financial background to perform their day-to-day tasks in an orderly and efficient manner. Whilst these systems are necessary, there exists another spectrum in the technology industry where, rather than developing systems with in-built specifi c rules, the machine itself learns from past experiences and takes actions accordingly. This branch of science is known as arti cial intelligence, more speci cally, machine learning. In this thesis, we focus on the use of machine learning to forecast financial time series, in particular relating to the stock market. The field of financial time series forecasting is one that has been exploited through various statistical and machine learning techniques, some of which achieved promising results. There exists an assumption that those who achieve the best results do not publish findings in order to be ahead of other traders, which makes this field of study all the more challenging. A number of industry leading classi ers and regressors are implemented, after which we approach this task by using a novel branch of neural network based algorithms known as deep learning. Deep Learning is a new branch of Machine Learning, which has been introduced with the objective of moving Machine Learning closer to one of its original goals: Arti ficial Intelligence. These techniques are known to excel in tasks such as image and text recognition, but have not been exploited as much in the field of finance. Through experimentation, we achieve a number of notable results, the best of which is an accuracy of 81% for long-term trend direction forecasting and 0.012 RMSE for next day price forecast, using so called, traditional Machine Learning methods. The Deep Learning methods fail to reach the levels of accuracy achieved by Logistic Regression and Support Vector Machines. We consider this drop in performance to be mainly due to the complexity of the deep architecture setup, wherein the task at hand may favour a more simple model. Whilst one may think that a complex problem such as stock market prediction should favour a complex model, in reality the almost random nature of the fluctuations may in fact favour a more generalised model with less layers and complexities.