Short‐term forecasting of the US unemployment rate

• Published date: 01 April 2020
• DOI: http://doi.org/10.1002/for.2630
• Date: 01 April 2020
• Author: Benedikt Maas

Received: 17 April 2019 Revised: 6 August 2019 Accepted: 28 October 2019

DOI: 10.1002/for.2630

RESEARCH ARTICLE

Short-term forecasting of the US unemployment rate

Benedikt Maas

Department of Economics, University of

Hamburg, Hamburg, Germany

Correspondence

Benedikt Maas, Department of Economics,

University of Hamburg, Von-Melle-Park5,

20146 Hamburg, Germany.

Email: Benedikt.Maas@uni-hamburg.de

Abstract

This paper aims to assess whether Google search data are useful when predicting

the US unemployment rate among other more traditional predictor variables.

A weekly Google index is derived from the keyword “unemployment” and is

used in diffusion index variants along with the weekly number of initial claims

and monthly estimated latent factors. The unemployment rate forecasts are

generated using MIDAS regression models that takeinto account the actual fre-

quencies of the predictor variables. The forecasts are made in real time, and the

forecasts of the best forecasting models exceed, for the most part, the root mean

squared forecast error of two benchmarks. However, as the forecasting hori-

zon increases, the forecasting performance of the best diffusion index variants

decreases over time, which suggests that the forecasting methods proposed in

this paper are most useful in the short term.

KEYWORDS

MIDAS, Google data, forecast comparison, US unemployment

1INTRODUCTION

In general, traditional labor statistics are available with

at least a 1-month lag. However, a more timely estimate

of the unemployment rate is desirable for investors and

policymakers, especially in times of economic uncertainty.

An accurate prediction of the US unemployment rate has

become even more important after the 2008–09 recession,

especially since the Federal Reserveannounced in Decem-

ber 2012 a shift of its monetary policy to a specific unem-

ployment rate threshold. The so-called “Evans rule” stated

that “the Committee decided to keep the target range for

the federal funds rate at 0 to 1/4 percent and currently

anticipates that this exceptionally low range for the fed-

eral funds rate will be appropriate at least as long as the

unemployment rate remains above 6-1/2 percent.”1

1For the official statement of the Federal Reserve's Open Market Com-

mittee see https://www.federalreserve.gov/newsevents/pressreleases/

monetary20121212a.htm.

This paper investigates whether or not the information

given in Google searches is useful to predict the US unem-

ployment rate. The idea behind using search engine data

is that, if an increase in searches is observed in connection

with unemployment, then this could give an early indi-

cation of an increasing unemployment rate. The potential

predictive power of Google search is used alongside other

more traditional predictors. One of these is the number

of IC. The IC is widely used in the literature as a predic-

tor variable in unemployment rate forecasts.2The current

state of the economy is also considered as a predictor in the

forecasts. The state of the economy has a major impact on

the unemployment rate: During a recession, an increase

in the unemployment rate is expected, whereas, during an

upswing and prosperity phase, a decrease in the unem-

ployment rate is expected. Totake into account the current

2See, for example, Montgomery,Zarnowitz, Tsay, and Tiao (1998)

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivsLicense, which permits use and distribution in any medium,

provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.

© 2019 The Authors. Journal of Forecasting published by John Wiley & Sons, Ltd.

Journal of Forecasting. 2020;39:394–411.wileyonlinelibrary.com/journal/for394

state of the economy,unobserved latent factors are derived

from a macroeconomic database by principal components,

as suggested in Stock and Watson (2002). These factorsare

intended to establish a link between the economic situa-

tion and the unemployment rate in the forecasting exercise

in this paper.

Given that the Google and IC data are available on a

weekly frequency, this paper uses weekly data to forecast

the monthly US unemployment rate with three DI variants

after Stock and Watson (2002) based on the MIDASregres-

sion model introduced by Ghysels, Santa-Clara, and Valka-

nov (2006) and Ghysels, Sinko, and Valkanov (2007). In

addition, factor-augmented versions, where the monthly

unobserved latent factors and the weekly data are com-

bined, are also estimated. In general, the MIDAS frame-

work allows us to combine variables of mixed frequencies

in a regression model. Related studies have used monthly

averages of the Google and IC data (D'Amuri & Marcucci,

2017), but—as empirically shown in Smith (2016), who

applies the MIDAS approach to forecasting the unemploy-

ment rate in the UK—there is no need to adjust frequency

to the target variable and thus lose valuable information.

As stated generally in Andreou, Ghysels, and Kourtellos

(2010), there is no reason to ignore the fact that variables

involved in empirical models are generated from processes

of mixed frequencies and are used to estimate econometric

models based on an aggregation scheme of equal weights,

because an equal weighting scheme can lead to informa-

tion losses and thus to inefficient or biased estimates.

The forecasts in this paper are conducted in real time and

almost all exceed an autoregressive benchmark for each

forecast horizon. However, the results show a mixed pic-

ture, in which a combination of predictor variables is most

favorable because the best empirical results change from

horizon to horizon. Comparing the MIDAS short-term

forecasts with the forecasts of D'Amuri and Marcucci

(2017), which are based on monthly averages of an alter-

native Google index, the models presented here obtain a

lower RMSFE for the shortest forecast horizons compared

to this benchmark.

The rest of this paper is organized as follows. Section

2 gives a compact overview of the related literature deal-

ing with the use of Internet data to forecast economic

variables. Moreover, it focuses on potential pitfalls when

using Internet data and the choice of keyword to obtain the

Google index. Section 3 explains the econometric frame-

work. Section 4 describes the data, the forecasting models,

and the real-time forecasting design. Section 5 states the

empirical results, and Section 6 concludes.

2THE USE OF INTERNET SEARCH

DATA IN FORECASTING

2.1 Related literature

Internet search data have been used in a number of dif-

ferent research topics. In economics, Choi and Varian

(2012) show that Google Trends data can help to forecast

near-term values of economic indicators, such as auto-

mobile sales, travel destinations, consumer confidence

and initial claims for unemployment benefits. Their paper

inspired many economists to use Google Trends data to

predict a variable that can be linked to the behavior of

households. For example, Vosen and Schmidt (2011) fore-

cast consumption of goods, whereas Bangwayo-Skeeteand

Skeete (2015) and Yang, Pan, Evans, and Lv (2015) use

Google data to predict future tourism demand. Wu et al.

(2015) predict US housing prices and sales. Using a Markov

switching framework, Chen et al. (2015) use Google search

data to improve the timeliness of business cycle turn-

ing point identification, and they successfully nowcast the

peak date within a month that the turning point occurred.

In their analysis, they use the three keywords “reces-

sion,” “foreclosure help,” and “layoff,” which represent

the aggregated economy, the credit market, and the labor

market, respectively. Liu, Xu, and Fan (2018) also use

Internet search behavior to forecast Chinese GDP. How-

ever,because Google is not prevalent in China, the authors

use data from its Chinese counterpart: Baidu.

Considering inflation expectations, Guzmán (2011) pro-

poses a real-time measure using search queries obtained

from Google. She demonstrates that higher frequency

measures tend to outperform standard lower frequency

measures such as the SPF in tests of accuracy, predictive

power and out-of-sample forecasts.

Dergiades, Milas, and Panagiotidis (2014) analyze

whether Google and social media data influence European

financial markets. They find that the data provide signif-

icant short-run information for the Greek–German and

Irish–German government bond yield differential.

Considering the unemployment rate, McLaren and

Shanbhogue (2011) analyze the labor market in the

UK and compare standard autoregressive (AR) mod-

els with those augmented with Internet data, finding

that the augmented models outperform the autore-

gressive benchmarks. Askitas and Zimmermann (2009)

demonstrate strong correlations between Google key-

word searches and unemployment rates for Germany,

and Fondeur and Karamé (2013) find that including

Google data improves youth unemployment predictions

in France. Vicente, López-Menéndez, and Pérez (2015)

investigate the unemployment rate in Spain using autore-

gressive integrated moving average (ARIMA) models with

an included explanatory variable that is derived from

the Google search term “job offers.” They find that sig-

MAAS 395