Unemployment Rate Forecasting using Machine Learning (Student Presentation, Group 3)

This is a STAT 451 class project presentation

by Susan Jiao, Yuanhang Wang, and Yi Xiao

This presentation is shared with the students' permission.

Abstract:

Building accurate forecasting models for economic indicators is a research area that many policy researchers work on. Traditional time series forecasting methods such as autoregressive moving average (ARMA) models often lead to unsatisfactory results. In this project, we exploit machine learning and deep learning techniques to forecast unemployment rate. We use three months of data to predict the following one month. Using linear regression as a baseline model, we compare results from random forests, XGBoost, and long short-term memory. There are two variants in all models: one uses 11 relevant economic indicators as input features, while another uses unemployment rate as the only feature. Both mean squared error (MSE) and mean absolute error (MAE) are used as evaluation metrics. We exclude year 2020 to control for noise from the COVID-19 pandemic. Among models that utilize all 11 features, XGBoost gives the best performance with MSE of 0.055. Among mod- els that use unemployment rate as the only feature, baseline linear regression performs the best. This could be due to the single-step forecasting structure in our models which is a relatively simple task compared to multiple-step forecast- ing. In addition, we test our models on year 2020 to see their performance during unusual time, we find XGBoost and LSTM to perform the best in the variant that uses all 11 features.