Neural networks in household investor behavior prediction

Abstract

Machine learning (ML) has been widely applied to various fields and areas, including the financial market. In this study, the behavior of active Finnish household investors is being predicted using three different neural network types: the multi-layer perceptron (MLP), long short-term memory (LSTM) and convolutional neural network (CNN). The target of the networks is to predict, whether an investor net-buys, net-sells or keeps their position on the next market day. In essence, the problem at hand is a three-class classification problem. The aim of this thesis is to find out which of the aforementioned networks receives the best F1 prediction scores and to get some insight to which inputs are of most importance. For the latter aim, attention mechanism is used. The data at hand, received from EuroClear, contains detailed information on all trades executed in Helsinki stock exchange. Focus is narrowed down to the 100 most active individual household traders of Nokia stock between the years 2006 and 2009. First, a short review of the functionality of the stock exchange is made, and most central behavioral finance findings presented. The review shows, that investors do not act rationally in the markets. Next, the basic working principles of neural networks is explained, after which the used data and data preparation processes are brought out. From the data, 13 inputs are created, which are used to predict future household investor behavior. Finally, the experiment setting is explained and results presented. Out of the networks experimented on in this study, the LSTM received the best prediction scores. However, none of the models falls significantly behind. Insight by the attention mechanism suggests, that the most important feature is the previous action of the investors. In the end, suggestions for receiving more satisfying results in future research are discussed. More sophisticated and complex models could give more satisfactory results.