Harnessing neural networks for predicting next actions of investors

Abstract

The goal of this study is to find out whether machine learning (ML) can be harnessed to predict next actions of active household investors. More precisely to compare the performance of three machine learning model structures. The three models observed are multilayer perceptron (MLP), recurrent neural network (RNN) based on gated recurrent units (GRU) and convolutional neural network (CNN). To gain insight of the benefits of using machine learning for the problem, the performance of the ML models was compared to a benchmark model which uses simple heuristics to predict the next action of an investor. The paper gives a brief introduction to financial markets and the differences between classical finance theory and behavioral finance theory. The purpose of this is to help the reader understand the behavior of investors on financial markets and the factors which make this research demanding. The binary classification task for the networks is to predict whether the next action of an investor is to buy or to sell the stock. The performance of the networks was assessed based on the accuracy and the F1-score the network achieves on a test data set it has not seen before. To gain additional insight on the performances of the networks, a confusion matrix was plotted and analyzed for the results of each network on said data set. The data for the research was provided by Euroclear. It contains transaction data from Nasdaq Helsinki stock exchange. For the purpose of this research, the data was narrowed down to 100 most active traders of Nokia stock between years 2006 and 2009. The best performing ML model was the RNN with a slight margin. The performances of all of the network models were relatively similar and none of them offered any performance increases compared to the benchmark model. The conclusion chapter provides insight and considerations for possible future research.