Essays on forecasting time series with machine learning techniques

Abstract

This dissertation consists of three essays on forecasting time series with machine learning techniques, delving into different financial and economic domains. The first essay investigates the forecasting accuracy of the S&P 500 stock market index during the pandemic, utilizing text mining and technical analysis. It finds that the LSTM model, which uses numerical data rather than financial news, offers superior accuracy in predicting price movements, showing the effectiveness of machine learning over traditional analysis methods. The second essay tests the hypothesis that the Federal Reserve responds to data revisions when setting monetary policy. To deal with the large number of data revisions that the Federal Reserve can potentially respond to, we use four machine learning techniques, Lasso, ridge regression, elastic net, and post-Lasso. When using our preferred method, elastic net, we conclude that the Federal Reserve responds to five types of data revision. We discuss the implications of this finding for theoretical macroeconomic models in the context of the signal extraction problem. The third essay addresses the predictability of West Texas Intermediate crude oil prices, influenced by macroeconomic factors. By comparing the performance of Long Short-Term Memory (LSTM) and Random Forest (RF) models, they are superior capability in both short and long-term forecasts during significant economic shocks like the 2008 financial crisis and the COVID-19 pandemic. The inclusion of SHAP analysis further enriches the understanding of how historical prices and macroeconomic indicators like the Consumer Price Index and exchange rates play pivotal roles in forecasting.