Forecasting the United States Unemployment Rate by Using Recurrent Neural Networks with Google Trends Data

Abstract

We study the problem of obtaining an accurate forecast of the unemployment claims using online search data. The motivation for this study arises from the fact that there is a need for nowcasting or providing a reliable short-term estimate of the unemployment rate. The data regarding initial jobless claims are published by the US Department of labor weekly. To tackle the problem of getting an accurate forecast, we propose the use of the novel Long Short-Term Memory (LSTM) architecture of Recurrent Neural Networks, to predict the unemployment claims (initial jobless claims) using the Google Trends query share for certain keywords. We begin by analyzing the correlation of a large number of keywords belonging to different aspects of the economy with the US initial jobless claims data. We take 15-year weekly data from January 2004 to January 2019 and create two different models for analysis: a Vector Autoregressive Model (VAR) model combining the official unemployment claims series with the search trends for the keyword ‘job offers’ taken from Google Trends and an LSTM model with only the Google trends time series data for the complete set of identified keywords. Our analysis reveals that the LSTM model outperforms the VAR model by a significant margin in predicting the unemployment claims across different forecast horizons.