Do Media Data Help to Predict German Industrial Production?

• Author: Konstantin A. Kholodilin,Tobias Thomas,Dirk Ulbricht
• Date: 01 August 2017
• Published date: 01 August 2017
• DOI: http://doi.org/10.1002/for.2449

Journal of Forecasting,J. Forecast. 36, 483–496 (2017)

Published online 3 November 2016 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/for.2449

Do Media Data Help to Predict German Industrial Production?

DIRK ULBRICHT,1KONSTANTIN A. KHOLODILIN2AND TOBIAS THOMAS3

1

IFF Hamburg, Germany

2

DIW Berlin, Germany

3

DICE, Düsseldorf, Germany

ABSTRACT

In an uncertain world, decisions by market participants are based on expectations. Therefore, sentiment indicators

reflecting expectations have a proven track record at predicting economic variables. However, survey respondents

largely perceive the world through media reports. Here, we want to make use of that. We employ a rich dataset

provided by Media Tenor International, based on sentiment analysis of opinion-leading media in Germany from 2001

to 2014, transformed into several monthly indices. German industrial production is predicted in a real-time out-of-

sample forecasting experiment and media indices are compared to a huge set of alternative indicators. Media data

turn out to be valuable for 10- to 12-month horizon forecasts, which is in line with the lag between monetary policy

announcements and their effect on industrial production. This holds in the period during and after the Great Recession

when many models fail. Copyright © 2016 John Wiley & Sons, Ltd.

KEY WORDS media data; German industrial production; forecast breakdown; real-time experiment; model

confidence set

INTRODUCTION

Typically, data on gross domestic product (GDP) are available on a quarterly basis. In addition, they are published

half a quarter following the end of the reference quarter. Therefore, in order to gain quick insight into the current

economic situation, a monthly series of industrial production is used. It is considered to be a key monthly indicator

for business activity. This is especially true in the case of Germany. Although the share of industrial production has

been shrinking since the 1980s, it remains relatively high when compared to other OECD and, especially, other EU

member countries.1Furthermore, the European Commission plans to raise the contribution of industry to GDP to as

much as 20% by 2020 (European Commission, 2014) in order to increase the competitiveness of the EU. Moreover,

industrial production contributes substantially to the business cycle dynamics.

Consequently, there have been many attempts to improve the forecast accuracy of this variable.2Most of these

studies employ ‘hard economic’ indicators, such as interest rates or manufacturing orders. There are also several

studies using ‘soft data’, such as business surveys, including the ifo and ZEW indicators (see, for example, Abberger

and Wohlrabe, 2006, or Hüfner and Schröder,2002). It is demonstrated that, owing to their forward-looking nature as

well as their timely availability,the soft data are well suited for forecasting industrial production. The underlying idea

of this approach is to employ a measure of the intentions or the expectations of the managers or analysts, respectively.

The main advantages of these indicators are their high frequency, timeliness and the fact that they are subject only to

minor revisions,3unlike many other statistical indicators.

Alternatively, media data could be used to improve the forecast accuracy. The use of media data for such an

analysis in an uncertain environment is rather straightforward. While in classical economics the Homo oeconomi-

cus is omniscient and decides independently, with his decisions leading to efficient outcomes at the market level,

Keynes (1937) underlines the role of uncertainty concerning decisions and behavior as well as the related (subop-

timal) outcomes at the macro level, just as von Hayek (1989) points to the pretense of knowledge. Similarly, both

Simon (1957) and Kahneman and Tversky (1979) show that actual human behavior clearly deviatesfrom the behavior

predicted by standard economic models. Due to their limited information processing capacity, individuals use sub-

jective models for the perception of reality. If these models are shared because of common cultural background and

Correspondence to: Dirk Ulbricht, IFF Hamburg, Germany E-mail: dirk.ulbricht@iff-hamburg.de

1According to the OECD Factbook 2011: Economic, Environmental and Social Statistics, in 2010, the percentage of total value added in industry

(including energy) was 24% in Germany,19% in the EU and 21% in the OECD countries.

2See, for example, Kholodilin and Siliverstovs (2006), Robinzonov and Wohlrabe (2010) or Drechsel and Scheufele (2012).

3As pointed out by an anonymous referee, surveys undergo revisions but of much smaller magnitude than, for example, the indicators of the

national accounts. As a rule, these small revisions are related to the fact that not all firms manage to respond in timely manner to the questionnaires

sent to them. Thus incorporation of their responses leads to revisions in survey-based indicators. Another source of revisions is an updating of the

seasonal factors, in case a survey indicator needs to be seasonally adjusted.

Copyright © 2016 John Wiley & Sons, Ltd