Three Essays in Finance : Machine Learning and Finance Applications

Abstract

The thesis contains three chapters. All chapters are centered around the themes of deep learning/machine learaning and finance applications.

The first chapter improve out of sample forecasting ability using an artificial neural network. Compared with traditional methods, deep learning I exploited two settings, respectively, memoryless and memory models. Memory models consistently outperform previous research and achieve both out of sample statistical and economic significance. I visualize the decision-making process of memory models and show that the model does not only pay attention to the current state during the forecasting process.

The second chapter extends on first chapter. By adding attention mechanism, I can investigate how the decision-making process of model is related to variations of economic conditions. I deploy dual attentive LSTM model to predict equity premiums and then extract the attention weights after the training process. The attention weights are regressed on economic condition variables. I find that the weights are either negatively or positively related to economic conditions and represent a cyclical pattern.

In the third chapter, I use machine learning to re-evaluate the classic, but controversial corporate governance problem------board structure and firm performance. I first show how GMM results are sensitive to various lag lengths selection. Based on the selected lag instruments using three alternative algorithms, I re-evaluate the effects of board structure on firm performance using GMM and find significant effects of board structure on firm performance under some cases. Show more