Strategic Central Bank Communications:

Discourse and Game-Theoretic Analyses of the Bank of Japan’s

Monthly Reports

Kohei Kawamura∗, Yohei Kobashi†, Masato Shizume‡, and Kozo Ueda§

Imcomplete and preliminary

Abstract

We conduct a discourse analysis of the Bank of Japan’s monthly reports and examine

their characteristics in relation to business cycles. We find that the difference between the

numbers of positive and negative expressions in the reports leads the leading index of the

economy by approximately three months, which suggests that the central bank’s reports

have some superior information about the state of the economy. Moreover, ambigious words

tend to appear more frequently with negative expressions. Using a simple persuasion game

model, we argue that the use of ambiguity in communication by the central bank can be

seen as strategic information revelation when the central bank has an incentive to bias the

reports (and hence beliefs in the market) upwards.

Keywords: monetary policy; transparency; natural language processing; modality;

latent Dirichlet allocation (LDA); verifiable disclosure model

JEL classification: D78, D82, E58, E61

∗The University of Edinburgh†Waseda University‡Waseda University§Waseda University (E-mail: kozo.ueda@waseda.jp). The authors appreciate comments from seminar partic-

ipants at the Bank of Japan and Waseda University. All errors are our own.

1

1 Introduction

Central banks not only implement monetary policy but also provide a significant amount of

information for the market (Blinder [2004], Eijffinger and Geraats [2006]). Indeed, most of

publications by central banks are not solely about monetary policy but about data and analyses

on the current economic state. It has been widely recognized that central banks use various

communication channels to influence the expectations in the market so as to enhance the

effectiveness of their monetary policy. Meanwhile, it is not readily obvious whether central

banks reveal all information they have as it is. In particular, central banks can communicate

strategically and be selective about the types of information they disclose, if not telling lies

due to accountability and fiduciary requirements. This concern would be especially important

when central banks’ objectives (e.g. keeping inflation/deflation under control, stimulating

the economy) may not be aligned completely with those of market participants and possibly

governments.

In this paper, we study a central bank’s communication strategy by looking at how ex-

pressions used in published reports are related to the state of the economy. Specifically, we

conduct discourse analyses using the Bank of Japan’s Monthly Report of Recent Economic and

Financial Developments (the monthly reports, hereafter) from January 1998 to March 2015.

Employing a natural language processing method, we first classify expressions in the monthly

reports according to polarity (whether an expression is positive, negative, or neutral) and

modality (whether an expression is clear-cut, ambiguous, or subjective), among others. We

find that the difference between the numbers of positive and negative expressions in the re-

ports leads the leading index of the economy by approximately three months, which suggests

that the central bank and the reports have some superior information about the state of the

economy. Moreover, ambiguous expressions are more likely to be used when the economy is in a

recession. For example, when the leading index of the economy is low, monthly reports contain

a larger number of expressions with negative tones (e.g. “fall”) and the modality expressions

that indicate likelihood (e.g., “seem” and “should”) rather than certitude. We further confirm

using a latent Dirichlet allocation (LDA) model that ambiguous expressions are indeed more

likely to be used in sentences that contain negative expressions. This suggests the possibil-

ity that the central bank deliberately introduces ambiguity into sentences conveying negative

information about the economy.

Second, we develop a simple persuasion game model to understand the empirical obser-

vations as a consequence of strategic communication. In the model the central bank always

obtains coarse information about the state of the economy but may or may not find precise

information. Moreover, the central bank has an upward bias: it wants the market to hold

2

optimistic beliefs when the ongoing inflation rate is lower than the target, which has been

actually the case in the data period of our empirical analysis. We demonstrate that when the

state of the economy is bad, the central bank only discloses coarse information. In other words,

negative reports become ambiguous in order to avoid the market to be very pessimistic. On

the other hand, if the central bank receives the precise signal that the economy is good, the

signal is disclosed as it makes the market belief more optimistic than when it is withheld. We

argue that the central bank’s equilibrium reporting strategy is consistent with what we observe

in the monthly reports.

Our study contributes to existing literature in illustrating and explaining the strategic

aspect of central banks’ communication. As the first strand of relevant papers, there are a

rapidly increasing number of studies on central banks’ communications using discourse analyses.

For example, Boukus and Rosenberg (2006), Hendry and Madeley (2010), Hendry (2012), Apel

and Grimaldi (2012) analyze central banks’ communications in the Federal Reserve, the Bank

of Canada, and Riksbank, and their relevance to the real economy and/or effects on financial

markets. Born, Ehrmann, and Fratzscher (2014) find that financial stability reports released by

central banks have positive effects on financial markets when their views are optimistic but no

effect when they are pessimistic. Our study complements these previous studies by emphasizing

the usage of ambiguity and explaining its rationale using a game theory. In particular, the above

finding by Born, Ehrmann, and Fratzscher (2014) as well as our empirical results turns out

to be consistent with the predictions from our game theoretic model. It is also worth noting

that, to the best of our knowledge, there is no academic study regarding the Bank of Japan’s

communication. Furthermore, our study may be valuable to financial market participants who

are eager to know the intents of central banks and forecast the future course of monetary policy.

The LDA is employed in economics, particularly for central banks’ communications, previously

by Hansen, McMahon, and Prat (2014).1

In view of natural language processing, our analysis contributes to the field of semantic

analyses in pointing out the importance of modality in drawing the sender’s assessments and

intentions in the actual policy front. A hegemony view is often taken as granted, that is,

communications by the authority are regarded as perfectly credible (Gramsci [1971]), although

there exists critical discourse analyses by Fairclough (1989) and van Dijk (2008). Our study

challenges the hegemony view by considering a possibility that a receiver does not necessarily

believe what a sender says.

This paper is structured as follows. Section 2 explains our data and results from discourse

analyses. Section 3 proposes a game theoretic model to explain the Bank of Japan’s communi-

1They investigate how discussions and decision making proceed using the Federal Reserve’s transcripts.

3

cation strategy. Section 4 concludes.

2 Data and Results from Discourse Analyses

2.1 Data

Our data are the Bank of Japan’s monthly reports in a Japanese version from January 1998 to

March 2015 (207 in total). The Bank of Japan started to publish monthly reports in January

1998 when it strengthened independence by a law amendment. Each report is released on the

next day of a monetary policy meeting. The Bank’s view that is a summary of the report and

authorized by committee members is released on the same day of the monetary policy meeting.

The full body of the report is written by the staff of the Bank of Japan, not by committee

members.

For our study, the Bank of Japan’s monthly reports have mainly three advantages over

other types of communication methods. First, their release is monthly, which is highly frequent.

Most central banks in advanced economies hold monetary policy meetings around eight times

a year. Consequently, their policy statements and reports are released less frequently than

monthly. True, other types of communication methods entail richer information. For example,

minutes and transcripts of monetary policy meetings convey what policy committee members

discussed and Q&A’s by Governor may convey non-verbal information. However, to analyze

relationships between communications and business cycles in a timely manner, in particular,

their lead-lag relationships, it is ideal to have data of a high frequency. Second, the Bank of

Japan’s monthly reports are rich enough for us to evaluate its assessments on the economy.

Every monthly report has typically around 150 sentences. Some central banks such as the

Federal Reserve and the Bank of England release statements immediately after meetings but

their length is much shorter.2 Moreover, because the Bank of Japan have been published since

1998, data are ample enough to obtain statistically significant results. Third, the format of

monthly reports is clear and stylized. Their sections are always written in the order of summary,

economic developments, prices, and financial developments. The summary is further divided

into two: current situations and forecasts. Moreover, each section makes routine explanations.

For example, the summary begins with a sentence that overviews current economic situations,

followed by detailed assessments on the overseas economies, exports, fixed business investment,

and so forth in the almost same order. Such a clear format enhances the accuracy of our natural

language processing.3

2The Bank of England enriched their communications in August 2015.3To our great disappointment, however, after we initiated this research, the Bank of Japan in June 2015

decided to discontinue the publication of monthly reports from January 2016. The number of monetary policy

4

We use monthly reports in Japanese although an English version exists. This is because

the Bank of Japan’s decision-making and communication with the public are made in Japanese

and because an English version is based on a Japanese version and translated to English with

some delay (a few days).4 We provide translation tables in Appendix.

2.2 Basic Statistics

Table 1 shows basic statistics of the Bank of Japan’s monthly reports. Monthly reports have

the average sentences of 32, 79, 20, and 24, respectively, for the sections of summary, economic

developments, prices, and financial developments. As for a morpheme that represents the

smallest grammatical unit, the number of morphemes per sentence is roughly the same among

the four sections, which is around 30.

From now on, we focus only on the section of summary. Although it reduces the sample

size, it turns out enough to obtain statistically significant results. Time series developments in

the number of sentences and morphemes are displayed in Figure 1. It shows a clear level shift

in October 2003, although the number of morphemes per sentence stays almost constant. This

is when the Bank of Japan decided to enhance monetary policy transparency. To mitigate the

effect of this policy shift, in analyses below, we normalize the number of expressions by the

number of morphemes except that we do the number of morphemes by the number of sentences.

Table 1: Basic Statistics

Sentences Morphemes Mor/senmean (s.d.) mean (s.d.) mean (s.d.)

Summary 31.69 (8.24) 940.14 (338.82) 29.25 (4.47)- present 22.61 (5.92) 629.11 (224.66) 27.44 (4.07)- forecast 9.08 (2.84) 311.03 (129.84) 33.89 (6.79)Economics 79.12 (27.88) 2848.83 (1110.44) 36.12 (4.19)Prices 19.55 (8.07) 704.60 (300.63) 36.80 (6.24)Financial 24.35 (3.58) 753.33 (211.22) 30.52 (4.46)

meetings is reduced from 14 times a year to 8 and the publication of the Outlook for Economic Activity andPrices increases from semi-annually to quarterly. The discontinuation of monthly reports may decrease theinterests of financial market participants who are eager to know the intent of the Bank of Japan from ourstudy. Nevertheless, we believe that our study has still a very important contribution to them in illustrating thecharacteristics of the Bank of Japan’s communication and intrinsic motives behind, which is likely to be passedon to other types of communication methods.

4Obviously, Japanese is different from English. One particularly important property is that Japanese isregarded as rich language with regard to expressing subjectivity in Japanese linguistics as we explain in Section2.3.2. The reason and influence of this difference are to be explored in the future.

5

0

500

1000

1500

2000

0

10

20

30

40

50

60

1998

0119

9809

1999

0520

0001

2000

0920

0105

2002

0120

0209

2003

0520

0401

2004

0920

0505

2006

0120

0609

2007

0520

0801

2008

0920

0905

2010

0120

1009

2011

0520

1201

2012

0920

1305

2014

0120

1409

Sentences (left axis)Morphemes/sentences (left axis)Morphemes (right axis)

Figure 1: Time Series Developments in the Number of Sentences and Morphemes

2.3 Classifications

Next we classify the morphemes of the monthly reports by various criteria, namely, polarity,

modality, a part of speech, and other miscellaneous criteria.

2.3.1 Polarity

The first is the polarity that indicates positive, negative, or even tones in morphemes. To this

end, we adopt the widely-accepted polarity criterion developed by Inui and Okazaki (XXX).5

Both positive and negative tones are further divided into two types: objective experiences and

subjective evaluations.

For each category, Table 2 illustrates the five most frequent morphemes where a number in

the right of morphemes indicates their counts. As the table shows, classifications are not perfect.

True, it seems acceptable that “increase” is categorized to even because it is positive when it

refers to an increase in demand but negative when it does to an increase in debts. However,

maybe different from the common sense of economists, “demand,” “fund,” and “economy“

are all categorized to positive. Japanese “tame” is used for two usages: “is good for” or “is

because of (reason).” Of the two, only the former contains a positive tone. Despite these

5See http://www.cl.ecei.tohoku.ac.jp/index.php?Open%20Resources%2FJapanese%20Sentiment%20Polarity%20Dictionary(in Japanese).

6

caveats, the reason we use this criterion is to maintain objectivity. Moreover, results drawn

from this criterion turns out to be indicative about business cycles.

Table 3 shows the basic statistics of polarity expressions. Evaluation morphemes, both

positive and negative, are less frequently used in monthly reports than experience ones. Com-

paring expressions in current situations and those in forecasts, we find that the latter tends to

have more positive (less negative) morphemes than the former. Indeed, the differene between

positive and negative morphemes, which is shown in the last column, is significantly larger for

forecasts than that for current situations at a one percent level.

Table 2: Top Five Polarity Expressions

Positive Negative Evenexperience # evaluation # experience # evaluation # #

demand 1062 good 205 fall 522 excess 147 invest 1166improve 733 good/reason 92 decline 397 weak 131 increase 1121

fund 626 ease 85 price 335 minus 108 environment 891recovery 501 ample 62 worsen 181 sluggish 88 modest 802economy 478 grow 60 cost 150 weak 45 produce 541

Table 3: Basic Statistics of Polarity Expressions

Positive Negative Even Positive-negativeexperience evaluation experience evaluation

Summary 0.0390 0.0040 0.0154 0.0030 0.0964 0.0246(0.0065) (0.0019) (0.0079) (0.0018) (0.0109) (0.0101)

- present 0.0347 0.0043 0.0148 0.0022 0.0954 0.0219(0.0076) (0.0019) (0.0084) (0.0014) (0.0116) (0.0120)

- forecast 0.0469 0.0036 0.0156 0.0045 0.0985 0.0307(0.0110) (0.0040) (0.0094) (0.0041) (0.0139) (0.0138)

Note: Ratio to total morphemes. Figures in parentheses represent standard deviations.

2.3.2 Modality

The second classification is modality. Specifically, we focus on the modality of truth judgement

and categorize it into three types: high probability, low probability, and unreal, opposed to

certitude. Table 4 shows expressions in each category and their basic statistics. High probability

expressions such as “seem” and “appear” suggest that a possibility that an event in the sentence

occurs is high, but not 100%. An example of low probability expressions is “may.” Unreal

expressions include “should” and “it is important to.” When they are used, a possibility that

an event occurs is nearly zero.

7

Modality is considered to be highly indicative of ambiguity. Modality is drawn from a

verb, which is located almost always at the very end of each sentence in Japanese. It has been

suggested in Japanese linguistics that Japanese syntactic structure is strongly correlated with

semantic hierarchical structure defined by the modality theory (Narrog [2009]). This structural

correlation implies that modality plays a more important role in conveying a writer’s perspective

in Japanese than that in English does.6

It is worth noting that we select the criterion of modality according to our human coding.

We choose all the end-of-sentence expressions that appear at least ten times in monthly re-

ports. Then, referring to previous studies (XXX), we classify them into high probability, low

probability, unreal, and certitude.

The table suggests that modality is more frequently used in forecasts than in current situa-

tions. This is natural because forecasts are uncertain, requiring the use of “seem” or “appear.”

Low probability expressions are never used in current situations.

Table 4: Modality Expressions

High probability Low probability UnrealExamples (seems, appear, expected, (may, warrant (should, it is

considered, forecasted, likely) careful monitoring) important to)

Summary 0.0090 0.0001 0.0007(0.0028) (0.0003) (0.0010)

- present 0.0006 0.0000 0.0004(0.0010) (0.0000) (0.0007)

- forecast 0.0261 0.0003 0.0014(0.0090) (0.0009) (0.0020)

Note: Ratio to total morphemes. Figures in parentheses represent standard deviations.

2.3.3 A Part of Speech

Third, we classify morphemes by a part of speech. A part of speech is a rough but reasonable

indication of contents. In Japanese, it consists of nouns, verb-type nouns, verbs, adjectives,

adjectival nouns, adverbs, particles, conjunctions, and suffixes. PLEASE EXPLAIN THEM.

Lexical units are the sum of nouns, verb-type nouns, verbs, adjectives, adjectival nouns, and

adverbs, while functional (grammatical) units are the sum of other parts of speech. A lexical

density is defined by a fraction of lexical units in total, which indicates the density of contents

according to Halliday (1985). Written languages generally have a higher lexical density than

spoken languages. Table 5 shows the basic statistics of a part of speech. In the order of

6One caution is that ambiguous expressions do not necessarily come from a writer’s subjective perspective.Expressions may become ambiguous because uncertainty surrounding the writer is large. In this empirical partof the study, we do not specify a fundamental reason for ambiguity.

8

appearance, particles, nouns, conjunctions, and verb-type nouns are frequently used, while

adjectives are the least.

Table 5: A Part of Speech

Nouns Verb-type nouns Verbs Adjectives Adjectival nouns Adverbs

Summary 0.2038 0.1067 0.0972 0.0081 0.0164 0.0374(0.0107) (0.0066) (0.0067) (0.0040) (0.0051) (0.0062)

- present 0.2143 0.1017 0.0935 0.0067 0.0130 0.0373(0.0129) (0.0077) (0.0065) (0.0045) (0.0055) (0.0067)

- forecast 0.1829 0.1170 0.1047 0.0106 0.0234 0.0377(0.0193) (0.0141) (0.0109) (0.0054) (0.0066) (0.0102)

Particles Conjunctions SuffixesSummary 0.5303 0.1094 0.0709

(0.0089) (0.0075) (0.0079)- present 0.5335 0.1113 0.0722

(0.0099) (0.0084) (0.0083)- forecast 0.5238 0.1056 0.0679

(0.0182) (0.0138) (0.0117)

Note: Ratio to total morphemes. Figures in parentheses represent standard deviations.

2.3.4 Other Miscellaneous Criteria

We also collect the expressions other than the above that are considered to be associated

with either polarity or ambiguity. They are (1) “increase,” (2) “decrease,” (3) conjunctions

with negative tones (e.g., “although,” “but”), (4) adverbs with clear tones (e.g., “clear”), (5)

adverbs with ambiguous tones (e.g., “slight,” “a bit,” “in general,” “almost”), (6) nouns with

ambiguous tones (e.g., “trend”), (7) “etc,” and (8) Japanese “wa.” We pick up (1) and (2)

specifically because they are particularly indicative of polarity. Expressions (5) to (7) entail

ambiguous tones. As for (8), Japanese “wa” is one of the most routinely used particle that is

attached to a subject. For example, “demand decreases” is translated to “juyou (demand) wa

genshou-shiteiru (decreases).”

2.4 Correlations with Business Cycles

Using expressions classified in the aforementioned method, we compare them with macroeco-

nomic data.

2.4.1 Macroeconomic Data

We use macroeconomic data that indicate Japan’s business cycles and/or relate to monetary

policy. Their frequency is all monthly. First, we use three kinds of composite indexes complied

9

by the Cabinet Office: the leading, coincident, and lagging indexes. The leading index, on which

we focus most, is compiled by combining 11 variables such as new orders for machinery, housing

construction started, the commodity price index, and the stock price. It leads the composite

index by a quarter. The composite indexes are published about 40 days after the month

concerned ends (e.g;, around March 10 for the index of January). Thus, to align timing with

monthly reports, we use two-month lagged composite indexes as the real-time data. The second

data is the year-on-year inflation rate based on the consumer price index (CPI). Again, we make

it real-timed by using two-month lagged series considering 30 days delay in its publication.

Moreover, we take account of revisions in the CPI which occurs every five years. We also

eliminate effects of consumption tax increases in 1997 and 2014. Third, we construct a monetary

policy change dummy from actual monetary policy changes. The variable takes one when policy

is tightened and minus one when policy is eased. Otherwise, it is zero. Some policy changes

may have been anticipated before monetary policy meetings, so it does not necessarily reflect

a monetary policy shock. Because there have been a number of small monetary easing in our

sample period, we construct the alternative dummy variable, which we call the big change

dummy, by choosing rather big policy changes. The dummy data are explained in Appendix.

2.4.2 Correlations

To compare expressions in monthly reports with macroeconomic data, we take a simple ap-

proach to look at their correlations. This is a basic and clear way to our aim. Although

correlations are silent about causality, a causality is likely to go from business cycles to the

Bank of Japan’s communications at least in the monthly time horizon. In other words, macroe-

conomic data on business cycles are considered to be exogenous, although monetary policy is

surely to influence the macroeconomy with a lag of several months to a couple of years. Spuri-

ous correlations tend to arise when data are non-stationary, but our data are stationary. That

said, we admit that our method is not immune to several caveats, so we check robustness in

the next subsection. A correlation is significantly different from zero at a one and five percent

level, if its absolute value exceeds 0.179 and 0.137, respectively for the sample size of 207.

Table 6 shows correlations between the expressions related to current situations and the

macroeconomic data. Several findings are worth noting. First, polarity expressions with posi-

tive (negative) tones are positively (negatively) correlated with the leading index, that is, the

future economy. This supports the usefulness of the polarity criterion developed by Inui and

Okazaki (XXX), although it is not designed to be applied to economics. In particular, dif-

ferences of positive and negative expressions (denoted by “pos neg” in the table) are highly

informative. Moreover we find that the words of “increase” and “decrease” are correlated with

10

the future economy positively and negatively, respectively. Conjunctions with negative tones

are also negatively correlated.

Second, expressions associated with ambiguity are negatively correlated with the future

economy. This is illustrated by a number of results. First, the ratio of morphemes to sentences

is negatively correlated with the leading index. In short, the future economy is worse, as

words per sentence are longer. Long explanations may suggest, in part, detailed information

revelation about the economy, but at the same time, increase unnecessary excuses and impair

clarity. Second, modality is negatively correlated with the leading index. When monthly reports

use expressions with high probability or unreal tones in the section of the current situation,

the future economy tends to deteriorate. Third, while nouns and verbs are indicative of an

increase in the leading index, adjectives are indicative of its decrease. These parts of speech

are all lexical, but nouns and verbs are considered to be professional and objective and they

are pro-cyclical. By contrast, adjectives are subjective and ambiguous and they are counter-

cyclical. Fourth, the use of “etc” is negatively correlated with the leading index. Clearly, this

word increases ambiguity.

Third, a result specific to Japanese is that the use of “wa” is highly correlated with the

leading index. As was previously explained, “wa” is one of the most routinely used particle that

is attached to a subject. This result implies that the economy is strong when the fraction of

routine expressions is high. By contrast, under unprecedented circumstances, the central bank

needs to make longer and unorthodox explanations, which reduce the usage of “wa.” Moreover,

a peculiar property of Japanese is that a sentence does not always require a subject. Japanese

allows to omit a subject and leaves the judgment as to who is the subject to readers or listeners.

This functions to increase ambiguity. Therefore, a less frequent use of “wa” implies greater

ambiguity.

Fourth, of the three composite indexes, the leading index tends to be most highly correlated

with the expressions such as polarity and modality. The coincident index is the second and the

lagging index is the least. In this respect, the Bank of Japan in monthly reports discusses the

future economy of about a quarter ahead, despite in the section of the current situation. This

is a natural thing for monetary policy because it needs to be forward looking.

Fifth, correlations with the inflation rate are lower in their absolute size than those with

the leading index. Polarity expressions have of little value in connecting with inflation develop-

ments. However, the words of “increase” and “decrease” are significantly correlated with the

inflation rate positively and negatively, respectively.

Finally, correlations with the monetary policy change dummy are examined. Differences

of positive and negative expressions are positively correlated with the dummy. That is, when

11

positive expressions are more used than negative expressions, the Bank of Japan tends to

tighten monetary policy.

Table 6: Correlations (Current Situations)

leading coincident lagging inflation mdummy mbigdummy

mor/sen -0.73 -0.61 -0.46 -0.14 -0.05 -0.06high prob -0.37 -0.33 -0.15 0.14 0.04 -0.03low probunreal -0.61 -0.44 -0.31 -0.08 -0.09 -0.09

pos neg 0.57 0.38 -0.03 -0.14 0.20 0.15pos exp -0.11 -0.27 -0.54 -0.28 0.03 -0.05pos eval 0.45 0.48 0.35 0.18 0.18 0.20neg exp -0.82 -0.67 -0.35 0.02 -0.18 -0.17neg eval -0.18 -0.02 -0.06 0.01 -0.05 -0.12increase 0.65 0.67 0.69 0.21 0.28 0.24decrease -0.41 -0.40 -0.10 -0.29 -0.06 -0.01

noun 0.41 0.34 0.31 0.09 -0.05 0.04verb noun -0.37 -0.23 0.04 -0.09 0.04 0.04

verb 0.28 0.14 -0.05 0.02 0.05 0.03adj -0.72 -0.79 -0.62 -0.41 -0.18 -0.13

adj noun 0.12 0.34 0.46 0.39 0.14 0.11adv -0.16 -0.16 -0.22 -0.11 -0.04 0.00

func word -0.06 -0.07 -0.23 -0.02 0.03 -0.10par 0.18 0.03 -0.02 -0.27 -0.01 0.00conj -0.06 -0.10 0.12 -0.06 0.09 -0.04suf 0.10 0.17 0.08 0.20 0.02 0.04

conj neg -0.44 -0.55 -0.46 -0.45 -0.04 -0.12adv clear -0.15 -0.34 -0.36 -0.46 -0.10 0.01adv amb 0.18 0.16 -0.07 -0.16 -0.11 -0.02

etc -0.55 -0.39 -0.24 -0.10 -0.08 -0.04noun amb 0.38 0.39 0.21 0.09 0.06 0.06

wa 0.67 0.60 0.56 0.04 0.18 0.13

Next, we turn to the statements on forecasts. Table 6 shows correlations between the

expressions related to forecasts and the macroeconomic data. This suggests that almost all the

previous results hold true except for some noteworthy differences. The first difference is a role of

modality. For the section of forecasts, modality expressions with high probability are positively

correlated with the leading index. This is natural because high probability expressions such

as “seem” and “likely” are tightly linked to forecasts. On the other hand, expressions with

low probability and unreal are negatively correlated with the leading index. In other words,

ambiguous expressions using “may” or “should” are counter-cyclical. Second, polarity remains

to be a useful indicator for the future economy, but the size of correlations is lower in forecasts

than in current situations. This is surprising to us because the section on forecasts should be

more concerned about the future economy than that on current situations. Readers may think

that this is because the Bank of Japan talks about farther future economy in the sections of

forecasts, but this is not the case as we will see soon below.

12

Table 7: Correlations (Forecasts)

leading coincident lagging inflation mdummy mbigdummy

mor/sen -0.54 -0.40 -0.35 -0.06 -0.09 -0.06high prob 0.60 0.42 0.33 0.10 0.09 0.08low prob -0.39 -0.28 -0.19 -0.22 -0.05 -0.03unreal -0.31 -0.17 -0.09 0.08 0.01 -0.01

pos neg 0.14 0.16 0.05 0.19 0.21 0.06pos exp -0.21 -0.13 0.06 0.01 0.19 0.07pos eval -0.33 -0.25 -0.26 0.15 -0.07 -0.13neg exp -0.60 -0.45 -0.09 0.01 -0.06 -0.04neg eval 0.02 -0.05 -0.01 -0.28 0.01 -0.01increase 0.79 0.67 0.54 0.12 0.24 0.20decrease -0.01 0.11 0.31 0.04 -0.08 0.01

noun 0.07 0.25 0.51 0.43 0.09 0.06verb noun 0.25 0.22 0.03 0.10 0.08 0.12

verb 0.11 -0.03 -0.27 -0.08 -0.18 -0.10adj -0.21 -0.22 -0.06 -0.13 -0.04 -0.05

adj noun 0.22 0.29 0.25 0.19 0.17 0.06adv -0.30 -0.30 -0.36 -0.26 -0.23 -0.18

func word -0.19 -0.28 -0.28 -0.36 0.03 0.00par -0.15 -0.11 0.04 0.17 0.04 -0.01conj 0.42 0.23 0.08 -0.22 0.06 0.06suf 0.12 0.15 0.20 -0.13 0.03 -0.02

conj neg -0.04 -0.06 -0.18 -0.17 -0.05 -0.03adv clear -0.08 -0.12 -0.26 -0.23 -0.12 -0.03adv amb -0.19 -0.22 -0.31 -0.21 -0.20 -0.06

etc -0.01 -0.01 -0.12 0.02 0.15 0.07noun amb 0.67 0.61 0.33 0.11 0.02 0.08

wa 0.35 0.19 0.02 -0.02 -0.05 -0.04

13

0

20

40

60

80

100

120

-0.02

-0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

19

98

01

19

98

09

19

99

05

20

00

01

20

00

09

20

01

05

20

02

01

20

02

09

20

03

05

20

04

01

20

04

09

20

05

05

20

06

01

20

06

09

20

07

05

20

08

01

20

08

09

20

09

05

20

10

01

20

10

09

20

11

05

20

12

01

20

12

09

20

13

05

20

14

01

20

14

09

pos-neg (current) pos-neg (forecast) leading

Figure 2: Time Series Developments in the Polarity Expressions (Left) and the Leading Index(Right)

Figure 2 demonstrates time series developments in the polarity expressions and the leading

index. As for the polarity expressions, we plot differences of positive and negative expressions

for both current situations and forecasts. Clearly, they are highly correlated with the leading

index.

Figure 3 shows correlations with the polarity expressions and the leading index. To study

their lead-lag relations, we compute the correlations using the leading index that differs in

timing from minus six to plus six months. The horizontal axis represents the month. For

example, +1 indicates a correlation between the polarity expressions and the leading index

with a one-month lead. This figure shows that correlations are peaked at x = 3 for both

current situations and forecasts. In other words, polarity expressions in monthly reports lead

the leading index by three months. This implies that the Bank of Japan has a forecasting power

about the three-month ahead leading index. Even if we exclude the effect of the two-month

delay in the publication of the leading index, the Bank of Japan maintains a forecasting power

by one month. Another finding is that the timing of the peak is the same between current

situations and forecasts. The section of forecasts does not necessarily discuss the economy of

a farther future horizon than that of current situations. Even worse, the correlation is lower.

Figure 4 illustrates time series developments in the modality expressions and the leading

index. The modality expressions are those of high probability for both current situations and

14

-0.2

0

0.2

0.4

0.6

-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6

Correlations bw positive-negative and leading index(+i)

Current Forecast

Figure 3: Correlations with the Polarity Expressions and the Leading Index

forecasts. This figure suggests that the modality for current situations increases when the

leading index falls. Here we give three illustrating examples. First, during the financial crisis

of 1998, the Bank of Japan used such expressions as “may be attributed to” and “appear

to.” Aaccording to its English version, the monthly report of July 1998 stated “stock prices

and yields on long-term government bonds have rebounded since mid-June 1998. This may be

attributed to a slight recovery in market sentiment, although still weak. . . ” (Italics are added

by the quoters.) It also said “growth in M2+CDs has been slowing. . . These developments

appear to strongly reflect the further decline in credit demand of private firms. . . ” Second, the

report of May 2009 uses “seem” in the aftermath of the Lehman shock as “It seems that firms’

funding costs ... have remained more or less unchanged at low levels.” The third example is

“appear to,” which is used from May 2013 to the end of our sample: “Inflation expectations

appear to be rising on the whole.” During this period, the economy was relatively in a better

shape due to the large-scale monetary easing introduced in April 2013. The stock price jumped

up and the exchange rate depreciated. However, the inflation rate and its expectations were

well below the Bank of Japan’s inflation target, two percent, although the Bank of Japan

promised to achieve this level within two years. Such a background seems to induce the Bank

of Japan to use the expression of “appear.” Moreover, the expression of “on the whole” in the

same sentence is also ambiguous.7

7We notice that the nuance in the English version differ at times. For example, the Japanese version in May

15

0

20

40

60

80

100

120

0

0.005

0.01

0.015

0.02

0.025

0.03

0.035

0.04

0.045

0.05

19

98

01

19

98

09

19

99

05

20

00

01

20

00

09

20

01

05

20

02

01

20

02

09

20

03

05

20

04

01

20

04

09

20

05

05

20

06

01

20

06

09

20

07

05

20

08

01

20

08

09

20

09

05

20

10

01

20

10

09

20

11

05

20

12

01

20

12

09

20

13

05

20

14

01

20

14

09

high prob (current) high prob (forecast) leading

Figure 4: Time Series Developments in the Modality Expressions (Left) and the Leading Index(Right)

The figure also implies a caveat of our study. Although high probability expressions are

used for forecasts every month, they are not used for current situations every month. As this

example illustrates, some expressions are sparse, which may yield a bias in looking at simple

correlations.

2.4.3 Robustness

Because correlations are too simple statistics to obtain robust conclusions, we check robustness

in various ways. First, we compute correlations using a time difference. Although the variables

we use are stationary, some of them tend to obey a slow moving process and embedding a

persistence. To resolve this and to check robustness, we take monthly differences for all the

variables and compute correlations. We confirm that monthly changes in polarity expressions

remain to be significantly correlated with monthly changes in the leading index. However, many

of our results regarding ambiguity disappear. In particular, neither words length per sentence

nor modality is no longer correlated with the leading index when their monthly changes are

taken.

2009 exploits two more modality expressions “considered” and “seem” in the following sentences. However, suchexpressions disappear in the English version. This may reflect a large importance of modality in Japanese tojudge the writer’s perspective than in English.

16

Second, we conduct the Granger causality test for the same reason above. Specifically, we

choose two variables: the leading index and the polarity expression defined by the difference of

positive and negative expressions for current situations. The AIC chooses the lag of six months;

the polarity expression Granger-causes leading index with a one percent significance; and the

leading index Granger-causes the polarity expression only with a 10 percent significance.

Third we confirm that our results are robust to the split of sample at October 2003. This

is when the Bank of Japan decided to enhance monetary policy transparency and reduced the

number of sentences considerably.

Fourth, although we have computed correlations between two variables, more than two

variables are likely to interact each other. In particular, some expressions in monthly reports

may be used together depending on economic circumstances. For example, both negative

polarity expressions and modality expressions may tend to be used in monthly reports when

the economy is in a recession. To tackle this issue, we employ the LDA method in the next

subsection.

2.5 Latent Dirichlet Allocation (LDA)

2.5.1 Methods

The LDA enables us to study what kind of expressions are used together and under what

circumstances. SOME EARLIER STUDIES AND/OR BRIEF MATHEMATHICAL EXPLA-

NATIONS.8 The LDA makes use of a generative probabilistic model for text corpora. The

model consists of a finite mixture over an underlying set of topics, where topics are the latent

variables that represent properties common to a number of expressions. Put intuitively, the

topics are like the principal components of expressions. The LDA is Bayesian unsupervised

learning and does not rely on the supervised classification that is based on human’s subjective

judgments. An important advantage of the LDA over other similar methods such as factor

and cluster analysis is that the LDA allows one expression to be categorized in different topics.

This appears plausible in that the same expression such as “increase” and “seem” can be used

in different contexts depending on economic circumstances and objectives. Moreover, the LDA

can resolve high-dimensional and sparse data (XXX), which is important in our data.

More concretely, we apply the LDA to our data in the following manner. We use expressions

at a low level, that is, in the unit of morphemes, not polarity, modality, or a part of speech.

Expressions are sorted in the order of those associated with modality, polarity, and a part of

speech. No double counting is allowed. For example, once “increase” is selected as a polarity

8See also Hansen, McMahon, and Prat (2014). ANY Difference???

17

expression, it is not selected as an expression of either noun or verb. The number of topics is

chosen from the AIC.

While many variants of models can be proposed, as the benchmark, we consider a rather

parsimonious model which consists of expressions associated with modality, polarity, adverbs,

and adjectives sorted in this order. The minimum appearance of five times is required for

expressions to be selected.

2.5.2 Results

First, we report a result for the section of current situations in Table 8. The AIC chooses four

topics. In the table, each column indicates a topic. The second row indicates the label of each

topic, which we name from selected modality and polarity. The third to ninth rows show the

concrete expressions that characterize the topic in the order of their selection frequency. The

rows from the tenth represent correlations with the time-series development in each topic and

macroeconomic data.

This table illustrates that Topic 2 is pro-cyclical with business cycles, while Topics 3 and

4 are counter-cyclical. Topic 2 is positively correlated with both the leading index and the

inflation rate. At the same time, Topic 2 consists of expressions with positive tones such as

“increase” and “ease” although they are categorized as even according to the polarity criterion.

In this topic, no modality expression is selected. On the other hand, Topics 3 and 4 are

negatively correlated with both the leading index and the inflation rate. Although Topic 3

includes positive expressions, they do not show any positive tones if we look at the expressions

closely (e.g., “demand,” “fund,” and “credit”). Rather, Topics 3 and 4 consist of negative

expressions such as “fall” and “worsen.” Moreover, Topic 3 embeds modality expressions

related to high probability and unreal. Such expressions like “seem” and “should” indicate

ambiguity and/or a lack of objectivity. Four topics also include many expressions belonging to

adverbs and adjectives. However, no noteworthy property is observed.

Next, we report a result for the section of forecasts in Table 9. Similar results can be found.

Four topics are selected from the AIC, of which Topics 1 and 2 are counter-cyclical and Topics

3 and 4 are pro-cyclical. Topics 1 and 2 consists of not only negative expressions but also

modality ones. In particular, modality expressions associated with low probability and unreal

are linked to these topics. Its examples are “attention should be paid to the possibility,” “may,”

and “should.” Again these expressions have ambiguous tones. When these tones are used, the

economy is likely to be weak with respect to the leading index and the inflation rate. By

contrast, Topics 3 and 4 are positively correlated with the macroeconomic data. These topics

consist of even polarity expressions, but they actually have positive tones (e.g. “increase”). For

18

Table 8: LDA Results (Current Situations)

Topic 1 2 3 4Label positive; even even high prob.; unreal;

positive; negativenegative

high prob. seem, appearlow prob.unreal shouldpositive fund; demand;

improve; creditdemand; good; ef-fect; money; profit;economy; income

demand; fund; creditdemand; interest;credit; activity;service; income;good/reason; resolve

negative fall; weak; weak; de-cline; weak; atten-tion; cost; excess;worsen; expenditure

price; fall; decline;worsen; subjue; slug-gish; financial posi-tions; severe; reces-sion; risk

even envirnonment;ease; level; issue;invest; increase;finance; modest;thing/maturity;under/middle

increase; environ-ment; invest; level;modest; issue; state;thing/maturity;ease; finance

others (top three) adv: previous year;adv: meanwhile;adv: generally; andso on

adv: previous year;adv: meanwhile;adv: generally; andso on

adv: recently; adv:still; adv: slight; andso on

adv: previous year;adv: still; adv: se-vere; and so on

Correlations withleading 0.068 0.514 -0.553 -0.537coincident -0.091 0.462 -0.421 -0.444lagging -0.227 0.393 -0.419 -0.343inflation -0.087 0.199 -0.249 -0.195mdummy -0.042 0.035 -0.061 -0.062mbigdummy -0.039 0.011 -0.038 -0.018

19

the section of forecasts, modality expressions associated with high probability appear together

with positive ones. As was stated, this is natural because this section discusses the future and

thus often use the expressions such as “seem” and “forecasted.”

Table 9: LDA Results (Forecasts)

Topic 1 2 3 4Label low prob; positive;

negativelow prob; unreal;negative

high prob; even high prob; even

high prob. seem, appear; con-sidered; forecasted;likely

seem, appear; fore-casted; considered;likely

low prob. attention should bepaid to the possibil-ity of; may

may

unreal should; is importantto

positive economy; demand;improve; supply anddemand conditions;information; profit;technology; income;progress; capital

negative fall; price; worsen;weak; risk; weak; ad-verse effect; uncer-tain; sluggish; uncer-tain

decline; excess;price; fall; subjue;expenditure; restruc-turing; attention;pass through; minus

even modest; increase; in-vest; trend; consume;effect; produce; con-sumer; invest; em-ployment

increase; trend; in-vest; modest; pro-duce; effect; employ-ment; consumer; ex-pand; under/middle

others (top three) adv: for the time be-ing; adv: whole; adv:still; and so on

adv: future; adv: forthe time being; adv:still; and so on

adv: for the time be-ing; adv: previousyear; adv: gradual;and so on

adv: for the time be-ing; adv: meanwhile;adv: previous year;and so on

Correlations withleading -0.495 -0.484 0.263 0.473coincident -0.433 -0.394 0.192 0.459lagging -0.364 -0.405 -0.062 0.548inflation -0.350 -0.216 0.183 0.056mdummy -0.136 -0.014 -0.178 0.221mbigdummy -0.036 -0.026 -0.119 0.160

We can construct many kinds of models by selecting a various combination of expressions.

For example the richest model is made up of all expressions used in monthly reports. However,

we find that such less parsimonious models tend to yield a larger number of topics, sometimes

more than 15, which prevents us from drawing useful implications. At the same time, we

confirm that the above results hold in many models. One interesting note is that “etc” is often

20

used in tandem with negative expressions.

2.6 Findings So Far

In this empirical part of the study, we find mainly two things. First, the Bank of Japan has a

forecasting power for the economy. Positive-negative indicators complied from monthly reports

lead the leading index by three months. This implies that the Bank of Japan has some superior

information to private agents, which serves as one of our modeling bases in the next section.

The second concerns the characteristics of the Bank of Japan’s communications. We find

that ambiguity tends to increase (decrease) when the economy is bad (good). More specifi-

cally, word length tends to be longer and modality and “etc” expressions tend to appear more

frequently when the leading index is low. The LDA suggests that modality is used in tandem

with negative expressions when the leading index is low.

3 Game-Theoretic Model

In this section we develop a simple game-thoretic model to understand the empirical findings

we have presented so far. We model communication by a central bank as a persuasion game

or verifiable disclosure model (e.g. Milgom, 1981; Shin, 2003), where the sender can choose to

disclose or withhold private information to the receiver, but cannot fabricate it. The assumption

is in contrast to that in “cheap talk” models (e.g. Stein, 1991), where the sender is allowed to

send any kind of message costlessly no matter what the private information is. The assumption

that the sender (central bank) cannot lie but can withhold information is relevant to the context

of the central bank’s periodic reports, because the data they contain may be verified later, and

because repeated interaction with the receiver/market (which is not modeled explicitly here)

means that there may be significant reputational and political costs if the central bank is found

to have fabricated information. Meanwhile it would be much more difficult for the market to

discern whether the central bank did or did not have a certain piece of information, as assumed

in our model below.

3.1 Setup

The economy consists of a central bank (CB) and a representative market participant (P).

The CB is the sender of information and the P is the receiver. There are three states of the

macroeconomy y ∈ {−1, 0, 1}. Each state arises with strictly positive probability and is either

partially or completely known to the CB but unknown to the P, as we will describe in detail

21

shortly. The feature that the CB has private information is consistent with our finding that

the Bank of Japan has forecasting power for the economy.

The P’s payoff is given by a quadratic loss function −(y−V )2. The CB’s report is denoted

by m, and the P Bayesian-updates the belief about the economy based on m and best responds

so that the P’s reaction is given by V ∗ = E[y | m].

We assume that the P’s interest and that of the CB are not completely aligned in the sense

that, conditional on the state y, the CB wants the P to take a higher action than y. In the

Appendix, we present a simple foundation of the CB’s bias and demonstrate that the CB is

upward biased when the inflation rate is lower than the target and downward biased otherwise.9

In this paper, we adopt such an assumption that the CB has an upward bias, since during the

data period Japan has seen deflation or inflation lower than the current target level of two

percent. For simplicity the CB’s payoff function is given by V .10 This implies that the CB is

better off when the market reaction is higher.

Before publishing the report m, the CB receives two types of private signals about the state

of the macroeconomy, namely S ∈ {SL, SH} and s ∈ {−1, 0, 1}. The CB receives an ambiguous

signal S with probability 1. If S = SL then y ∈ {−1, 0}, that is, y may be low. If S = SH ,

then y ∈ {0, 1}, that is, y may be high. In addition, the CB receives a more detailed, clear

signal s with probability θ ∈ (0, 1). The clear signal is perfectly informative about the state: if

s = x then y = x. The parameter θ is common knowledge and represents how well the CB is

informed. The CB’s choice in this game is which signal to disclose or withhold.

3.2 Equilibrium

Let us consider how information is revealed in a perfect Bayesian equilibrium of this game. The

first step is to see that in equilibrium the CB cannot completely withhold private information.

Suppose that the CB does not publish any information. Then the P’s reaction will be V = E[y],

where E[y] is the unconditional expectation of y. However, when S = SH , the CB reveals the

signal since it induces a higher reaction E[y | SH ] > E[y]. In turn, if the CB does not reveal

S = SH or s, then the P can infer that S = SL (recall the assumption that the CB always

receives S). The P is indifferent between publishing S = SL and not publishing any information,

and in any case S is perfectly revealed in equilibrium. Naturally, when s is not observed, the

CB only publishes S ∈ {SL, SH}.When the CB observes a clear signal s, there are four cases.

First, if s = −1, the CB withholds s = −1 and publishes only S = SL. This is because

9Alternatively, see e.g. Walsh [Chapter 7, 2010] for discussions on such inflation bias.10This particular form of the payoff function is not essential. Our results hold, for example, if the CB’s payoff

function is −(y + b− V )2 and b is large enough, where b > 0 is the CB’s upward bias.

22

we have V = E[y | m = SL] > −1, which holds since the P cannot tell whether the CB has

received s = −1 and withheld it or has not received s and the state y can be either −1 or 0.

If s = 0 and S = SL, the CB reveals s = 0, since it induces higher reaction V = 0 > E[y |m = SL].

If s = 0 and S = SH , the CB withholds s = 0, since V = E[y | m = SH ] > 0.

Finally, if s = 1, the CB reveals s = 1, since V = 1 > E[y | m = SH ].

The above arguments can be summarized in the following proposition.

Proposition 1 In the unique perfect Bayesian equilibrium,

i) if clear signal s is not observed, then the CB’s report is ambiguous;

ii) if s is observed, then the CB sends the report of m = s only when s = 1 or when s = 0

and S = SL.

The proposition has the simple intuition that, because of the upward bias, the CB hides

a clear signal whenever the corresponding ambiguous signal induces a more optimistic belief

(amd reaction). The results can be related readily to our empirical findings.

Remark 2 The negative reports are always ambiguous.

The CB never reports s = −1. If s = −1, then the CB hides it and sends an ambiguous

report m = SL instead. In the context of our empirical analysis, m = SL can be thought of as

reporting negative sentences with modalities, which makes them ambiguous and less categorial

about the state of the economy. The market cannot know for certain whether modalities are

used because the CB does not have clear information, or because the CB has clear information

but withholds it to influence the beliefs in the market. Positive reports can also be ambiguous

(m = SH) but if a clear signal has been obtained, it is revealed, which suggests that positive

expressions in the monthly reports are less likely to contain modality.

In addition, the model generates further implications beyond our empirical findings.

Remark 3 If taken literally, the CB’s reports are upward biased.

If s = −1, then m = SL. If s = 0 and SH is observed, then m = SH . Although the P rationally

Bayesian-updates and thus is never deceived, the expressions in the monthly reports should

include more positive and less negative expressions than the state of the economy indicates.

Remark 4 Let m ∈ {SL,−1} be an pessimistic report and m ∈ {SH , 1} be an optimistic report.

Then on average, optimistic reports have more impact on the market reaction than pessimistic

reports.

23

The apparent asymmetric reaction of the market is only due to the fact that m = −1 is never

revealed and thus the overall reaction is dampened when the report is negative. This can be an

explanation for the finding by Born, Ehrmann, and Fratzscher (2014) that financial stability

reports from the central banks around the world lead to significant positive stock market returns

when they are optimistic, but no such effect is found when they are pessimistic.

4 Concluding Remarks

TBA

References

[1] Apel, M. and M. Blix Grimaldi (2012) “The Information Content of Central Bank Min-

utes,.” Sveriges Riksbank Working Paper Series 261.

[2] Blinder, Alan S. (2004) “The Quiet Revolution: Central Banking Goes Modern.” New

Haven: Yale University Press.

[3] Born, Benjamin, Michael Ehrmann, and Marcel Fratzscher (2014) “Central Bank Com-

munication on Financial Stability.” Economic Journal, 124, pp. 701–734.

[4] Boukus, E. and J. V. Rosenberg (2006) “The Information Content of FOMC Minutes.”

mimeo, Federal Reserve Bank of New York.

[5] Crawford and Sobel (Etrica);

[6] Eijffinger, Sylvester and Petra Geraats (2006) “How Transparent are Central Banks?”

European Journal of Political Economy, 22, pp. 1–21.

[7] Fairclough (1989)

[8] Gramsci, Antonio (1971) “Selections from the Prison Notebooks of Antonio Gramsci.”

New York: International Publisher.

[9] Hendry, S. (2012) “Central Bank Communication or the Media’s Interpretation: What

Moves Markets?” Bank of Canada Working Papers 12-9.

[10] Hendry, S., and A. Madeley (2010) “Text Mining and the Information Content of Bank of

Canada Communications.” Bank of Canada Working Papers 10-31.

[11] Hansen, Stephen, Michael McMahon, and Andrea Prat (2014) “Transparency and Delib-

eration within the FOMC: a Computational Linguistics Approach.” mimeo.

24

[12] Inui and Okazaki (XXX)

[13] Milgrom, Paul (1981) “Good News, Bad News: Representation Theorems and Applica-

tions.” Bell Journal of Economics, 12, pp. 380–391.

[14] Narrog, Heiko (2009) “Modality in Japanese: The Layered Structure of the Clause and

Hierarchies of Functional Categories.” Amsterdam: Hohn Benjamins Publishing.

[15] Palmer, F. R. (2001) “Mood and Modality.” Second Edition. Cambridge: Cambridge

University Press.

[16] van Dijk (2008)

[17] Shin (2003)

[18] Stein (1989 AER);

[19] Walsh, Carl E. (2010) “Monetary Theory and Policy.” Cambridge: MIT Press.

[20] Woodford, Michael (2003) “Interest and Prices.” Princeton: Princeton University Press.

25

A Japanese-English Correspondence Tables

In each cell, the left expression indicates the original Japanese morpheme corresponding to the

right English expression that we translated for this paper.

Table 10: Top Ten Polarity Expressions

Positiveexperience evaluation

juyou demand ryouko goodkaizen improve tame good/reasonshikin fund yawaragu ease

kaifuku recovery juntaku amplekeiki economy takamaru grow

shotoku income kousuijyun high levelshuueki profit meikaku clear

shikin juyou credit demand medatsu conspicuouskanousei possibility kenchou firmjyukyuu supply and sekkyoku active

demand conditions

Negative Evenexperience evaluation

geraku fall kajou excess toshi investteika decline yowai weak zoka increase

kakaku price mainasu minus kankyou environmentakka worsen teichou sluggish yuruyaka modest

kosuto cost toboshii weak seisan produceshikin guri financial positions fukakujitsu uncertain kanwa ease

donka subjue kanman lackluster naka under/middlekibishii severe futoumei uncetain suijun levelteimei weak zeijaku fragile kichou trendrisuku risk - - koyou employment

26

Table 11: Modality Expressions

high risk u seem, appearga ukagawareru seem, appearkoto ga mikomareru expectedkoto wa tenbou shinikui unlikelyte iku koto ga kitai sareru expectedte iku to mirareru seem, appearte iku to yoso sareru forecastedte iku to kangae rareru consideredte iku mono to kitai sareru expectedte iku mono to kangae rareru consideredte iku kanousei ga takai likelyte iru to mirareru seem, appearte iru mono to mirareru seem, appearte iru yo ni ukagawareru seem, appearde iku to mirareru seem, appearde iku to kitai sareru expectedto mirareru seem, appearto yosou sareru forecastedto kangae rareru consideredto mikomareru seem, appearwa izen toshite kitai shinikui jokyo ni aru still difficult to expectmo ukagawareteiru seem, appearreteiru to kangae rareru consideredwo tadoru tono mikata ga ippanteki de aru generally thoughtkousan ga ookii likelykanousei ga ookii likelyhajimeru to kangae rareru consideredtsudukete iku to mirareru seem, appear

low risk risuku niwa hikitsuzuki ryuui ga hitsuyou de aru attention should still be paid to the possibility ofkanousei ga aru maykanousei nimo ryuui ga hitsuyou de aru warrant careful monitoring

unreal ga hitsuyou de aru shouldte iku koto ga hitsuyou de aru shouldte iku koto ga jyuuyou to kangae rareru is important tote iku koto mo jyuuyou to kangae rareru is important tote iku hitsuyou ga aru shouldte mite iku hitsuyou ga aru should be observedhitsuyou ga aru should

27

Table 12: Other Miscellaneous Expressions

Negative conjunctions ga, mo, monono, nao, nagara, but, althoughshikashi, tadashi, mottomo, tada but, althoughtsutsu, ippou while

Clear adverbs hakkiri clearkiwamete extremekanari very, quite

Ambiguous adverbs yaya, ikubun, wazuka, jakkan slight, a bitoomune, soujite, zentai generaljojoni gradualhobo, zengo, aru teido almost, abouttoumen for the time beingichibu parthoka etc, and so on

Ambiguous nouns keiko, kichou tendkennnai more or less

Etc nado, tou etc, and so on

B Monetary Policy Change Dummy

28

Table 13: Monetary Policy Change Dummy: Otherwise dummies are zero.

Monetary policy Big changechange dummy dummy Notes

1998.09 -1 -1 call rate from 0.5% to 0.25%1999.02 -1 -1 call rate to 0.15%2000.08 1 1 call rate to 0.25%2001.02 -1 -1 call rate to 0.15%2001.03 -1 -1 quantitative easing (5 tril yen)2001.08 -1 0 6 tril yen2001.09 -1 0 over 6 tril yen2001.12 -1 0 10 to 15 tril yen2002.02 -1 0 increase the purchase of long-term bonds ( 0.8 to 1 tril yen/month)2002.10 -1 0 increase the purchase of long-term bonds ( 1 to 1.2 tril yen/month)2003.03 -1 0 17 to 22 tril yen2003.04 -1 0 22 to 27 tril yen2003.05 -1 0 27 to 30 tril yen2003.10 -1 -1 27 to 32 tril yen; enhance monetary policy transparency2004.01 -1 0 30 to 35 tril yen2006.03 1 1 terminate quantitative easing; understanding of price stability2006.07 1 1 call rate from 0% to 0.25%2007.02 1 1 call rate to 0.5%2008.09 -1 0 U.S. dollar funds-supplying operation2008.10 -1 -1 call rate to 0.3%2008.12 -1 -1 call rate to 0.1%; purchase or long-term bonds (1.2 to 1.4 tril yen/month)2009.03 -1 0 purchase or long-term bonds (1.4 to 1.8 tril yen/month)2009.12 -1 -1 enhance easy monetary conditions; clarify price stability2010.04 -1 0 strengthen the foundations for economic growth2010.10 -1 -1 comprehensive monetary easing; call rate 0 to 0.1%; asset purchase program2011.03 -1 0 asset purchase program to 40 tril yen2011.08 -1 0 asset purchase program to 50 tril yen2011.10 -1 0 asset purchase program to 55 tril yen2012.02 -1 0 asset purchase program to 65 tril yen2012.04 -1 0 asset purchase program to 70 tril yen2012.09 -1 0 asset purchase program to 80 tril yen2012.10 -1 0 asset purchase program to 91 tril yen2012.12 -1 0 asset purchase program to 101 tril yen2013.01 -1 -1 2% inflation target; accord with the government2013.04 -1 -1 Quantitative Qualitative Monetary Easing (QQE)2014.10 -1 -1 expand QQE

29

C Model of a Central Bank’s Incentive

In this appendix, we provide a model example in which the CB has an upward bias based on

a standard New Keynesian model.

The CB’s loss function is given by

LCB = π2 + λ(π − πe)2, (1)

where π and πe represent the inflation rate and the inflation expectations, respectively. The

first term indicates an inflation aversion as is commonly used in the New Keynesian model

such as Walsh (2010) and Woodford (2003). To this, we add the second term that indicates

an aversion to inflation surprises. Rationales for this assumption is that surprises increase

uncertainty, and in turn, dampen economic activity. Or, the P may simply dislike uncertainty,

as is assumed below, and a part of the P’s utility constitutes the CB’s utility. A parameter λ

represents a weight on inflation surprises.

The P’s loss function is expressed as

LP = π2 + γ(π − πe)2, (2)

where γ represents a weight. A divergence of preferences arises between the CB and the P

unless λ = γ.

The standard New Keynesian Phillips curve is assumed as

π = κy + βπe, (3)

where y represents an output deviation from its natural level and κ and 0 < β < 1 represent

parameters.

The model proceeds as follows. First, the CB sends information on y, y∗, after observing

true y. Second, the P receives y∗ and forms expectations for y, and in turn, π, as

πe =κ

1− βye. (4)

This is derived from the the rational expectation solution of equation (3), where πe = E(π|y∗)and ye = E(y|y∗).

Then, the CB’s loss function becomes

LCB =

{κy +

κβ

1− βye}2

+ λκ(y − ye)2 (5)

30

This equation suggests that, when y < 0, the CB has an incentive to achieve ye > y. In

other words, the CB wants the P to be optimistic about y. To see this, first consider a case

where λ is zero. In this case, the CB’s loss is minimum when ye = −1−βκβ y, which is positive

and larger than y. If the P thinks that y is positive, it raises its inflation expectation, which

in turn, contributes to the stabilization of the inflation rate around zero due to equation (3).

This helps lower the CB’s loss. Second, consider a case where λ is infinite. In this case, the

CB’s loss is the lowest when ye = y. In an intermediate region of λ, the optimal ye for the CB

lies between −1−βκβ y and y, which is below y.

In other words, when the economy is bad (y < 0), the CB has an upward bias. The CB

has an incentive to send better information than fundamentals to prevent deflation. In Japan,

the inflation rate has been below the two percent target for two decades. In this respect, the

assumption of y < 0 seems plausible.

31