Incorporating the Beige Book into a Quantitative Index of Economic Activity

• Published date: 01 August 2017
• Date: 01 August 2017
• DOI: http://doi.org/10.1002/for.2450
• Author: Ren Zhang,Michael Fulmer,Nathan S. Balke

Journal of Forecasting,J. Forecast. 36, 497–514 (2017)

Published online 21 November 2016 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/for.2450

Incorporating the Beige Book into a Quantitative Index of

Economic Activity

NATHAN S. BALKE,1,2MICHAEL FULMER1AND REN ZHANG1

1

Department of Economics, Southern Methodist University, Dallas, TX, USA

2

Federal Reserve Bank of Dallas, TX, USA

ABSTRACT

We apply customized text analytics to the written description of economic activity contained in the Beige Book (BB)

in order to obtain a quantitative measure of current economic conditions. This quantitative BB measure is then incor-

porated into a dynamic factor index model that also contains other commonly used quantitativeeconomic data. We find

that at the time the BB is released it has information about current economic activity not contained in other quantitative

data. This informational advantage is relatively short lived. By 3 weeks after their release date, ‘old’ BBs contain lit-

tle additional information about economic activity not already contained in other quantitative data. Copyright © 2016

John Wiley & Sons, Ltd.

KEY WORDS Beige Book; text analysis; dynamic factor model

INTRODUCTION

The Beige Book (BB) is a written description of economic conditions in each of the 12 district banks of the Federal

Reserve System. It is released eight times a year, roughly 2 weeks before the FOMC meeting. Its release is typically

greeted with interest in the financial press. Is this attention warranted?1Does the BB contain information about current

economic conditions? If so, can one use this information in a systematic manner? Can the information in the BB be

combined with other quantitative information to provide a clearer picture of current economic conditions? In order to

answer these questions, one must convert the qualitative information in the BB into quantitative information.

In this paper, we apply customized text analytics developed by Fulmer (2014) to extract quantitative information

about current economic conditions from the BB. These quantitative BB indices are strongly correlated with current-

quarter real GDP. We then incorporate a quantitative BB measure into a revised version of the Aruoba et al. (2009)

(ADS) index model of daily economic activity. The ADS index is based on a dynamic common factor model that

extracts a daily economic activity factor from six economic indicators: weekly jobless claims, monthly industrial

production, personal income less transfers, employment, manufacturing trade growth and quarterly real GDP growth.

We find that when using the full sample of final release data the incorporation of the BB index into the ADS model

has little effect on the estimated daily index of economic activity. We find this somewhat comforting. Our priors are

that, once the full sample of quantitative information is available, written information in the BB is unlikely to cause

one to rewrite economic history. However, when conducting a ‘real-time’ analysis of the last three recessions, we find

that the BB provides information about current economic activity not included in the other quantitative indicators

present in the ADS index. Specifically, on the day that the BB is released, including the BB typically moves the ADS

index towards its full-sample estimate (‘the truth’). The informational contribution of the BB is typically largerduring

recessionary periods than in normal times. By about 3 weeks after the BB’s release date, however, the ADS indices

with or without the BB are very similar; the other information in the ADS index dominates whatever information the

‘old’ BB contains about current economic activity.

The paper is organized as follows. The next section contains a brief description of the BB and a brief review of

the previous literature on quantifying the BB. The third section describes the customized text analytics applied to the

BB. The fourth section describes how we incorporate the quantified BB into the dynamic factor model underlying

the ADS index. The fifth section compares the properties of the ADS index with and without the BB. The sixth

section concludes.

BACKGROUND

The BB is largely based on information that each of the 12 individual Federal Reserve District Banks gather from

‘contacts’ in their district. These ‘contacts’ are survey respondents, members of bank board of directors, news reports,

Correspondence to: Nathan S. Balke, Department of Economics, Southern Methodist University,Dallas, TX, USA. E-mail: nbalke@smu.edu

1There is some skepticism about the usefulness of the BB. Alan Blinder (1997) somewhat disparagingly refers to the Fed’s use of anecdotal

information as ‘ask your uncle’ means of gathering information.

Copyright © 2016 John Wiley & Sons, Ltd

498 N. S. Balke, M. Fulmer and R. Zhang

Figure 1. Real GDP growth and Balke–Petersen BB index (average of national sectors). [Colour figure can be viewed at

wileyonlinelibrary.com]

and even officially released economic data. The information from these contacts is then summarized in the form of a

written document. In addition to a written summary provided by each of the 12 districts, there is a national summary

section in the BB. This is written by one of the district banks and is ostensibly based on the 12 district write-ups (the

task of writing this summary is rotated among the 12 district banks). The information in the BB is compiled during

the period up to the ‘closing date’, which is roughly 10 days prior to the official release date. For example, the BB

that was released on 3 September 2014 had a closing date of 22 August 2014, and the BB that was released on 16

July 2014 had a closing date of 7 July 2014. The BB has been publicly released 2 weeks before the FOMC meeting

since June 1983. Before that it was known as the ‘Red Book’ and was not publicly available. When we conduct our

empirical analysis, we use both the Beige and Red Books.

Literature

Balke and Petersen (1998, 2002) were the first to attempt to quantify the information in the BB. They ‘manually’

read and scored each of the BBs from 1983 to 1997. Various features of the BBs were scored on a 2 to 2 scale by

0.5 increments.2The quantitative scores were then compared to actual real GDP growth. Figure 1 displays one of

the Balke–Petersen BB indices and real GDP growth over their sample. Clearly, from Figure 1, the Balke–Petersen

index tracks real GDP growth pretty well. In general, Balke and Petersen found that the BB had predictive content

for current-quarter real GDP above and beyond other quantitative information available to analysts at the time the

BB was released. On the other hand, Balke and Petersen found that the predictive content of the BB for next-quarter

real GDP growth was substantially lower than for current-quarter real GDP growth. Follow-up studies that have used

human readers/scorers (see Fettig et al., 1999; Balke and Yucel, 2000; Ginther and Zavodny, 2001; Zavodny and

Ginther, 2005) have also generally found that the BB contains information about current economic conditions both at

the national and district level, but that the BB loses its informational advantage as other more recent data are released.

While the studies based on human readers of the BB suggest that there is information in it that could be useful for

‘nowcasting’ current economic activity, little effort has been given to updating these original studies of the quantified

BB. There are several problems with employing human readers and scorers of the BB. First, the time it takes to read

and score the BB makes it costly to update on a continual basis. Second, it is difficult to get a consistent reading of

the BB over longer time periods—readers may vary over time and even the same reader’s interpretation can change

over time. These two problems make it difficult to replicate and systematically update the human reader BB indices.

Given the difficulties with ‘manually’ reading and scoring the Beige Book, there have been a fewattempts to apply

simple textual analysis to it. Payne (2001) developed a list of key wordcombinations and gave each a numerical score

(1 to 1 by 0.5 increments). He then counted the number of times these word combinations occurred in each dis-

trict summary sentence. He found that the BB had strong predictive content for current quarter real GDP growth and

for the Coincident and Leading Economic Indicators. Armesto et al. (2009) applied off-the-shelf text analytics soft-

ware, Diction, to the BB. Using the word dictionary in Diction, Armesto et al. conducted a frequency count of words

associated with ‘optimism increasing’ and ‘optimism decreasing’ to construct an optimism index (‘optimism increas-

ing’) and a pessimism index (‘optimism decreasing’). They used mixed frequency estimation (MIDAS) to account for

2They scored the National Summary, the four major sectors discussed in the National Summary,and the 12 district write-ups.

Copyright © 2016 John Wiley & Sons, Ltd J. Forecast. 36, 497–514 (2017)