Thesis - Open Access
Bachelor of Arts
John V. Duca
Edward F. McKelvey
Barbara J. Craig
Machine learning, Forecasting, Neural networks, Artificial intelligence, Unemployment rate
This paper examines different machine learning methods to project the U.S. unemployment rate one year ahead. The forecasts include a naive forecast equal to the current unemployment plus the change of unemployment over the last year, along with forecasts from a Lasso regression and a neural network model. The last two models, which can be quickly run using an SQL database, select data from the Federal Reserve Economic Database (FRED) and are fitted (trained) in-sample from 1970 to 2000 to forecast quarterly unemployment rates over 2001 to 2018. The training window is updated in each forecast quarter to include new data. A rolling-window and non-rolling window period are tested for the training window. This paper finds that a non-rolling neural network model forecasts bests and outperforms the Survey of Professional Forecasters (SPF) across all time periods as does our Lasso regression model, though to a lesser extent. From experiments dropping broad categories of FRED, international data were the most important in forecasting the unemployment rate, followed in order by data from the FRED categories: Population, Employment, Labor Markets; and Money, Banking, and Finance.
Kreiner, Aaron S., "Can Machine Learning on Economic Data Better Forecast the Unemployment Rate?" (2019). Honors Papers. 126.