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 Nov 2015 KKSU Strategic Communications Nov 2015 1 / 24
Introductions Central banks (CB) can communicate strategically and be selective about the types of information they disclose, if not telling lies. How are expressions used in published reports related to the state of the economy? KKSU Strategic Communications Nov 2015 2 / 24
BOJ's Communications BOJ's Communications through MPM decision releases (14 times a year) Monthly reports of recent economic and nancial developments (monthly reports) Q&As MPM minutes Outlook for economic activity and prices (April and October) MPM transcripts after 10 years Speech, testimony at parliament, and so on We conduct discourse analyses using the monthly reports from January 1998 to March 2015. KKSU Strategic Communications Nov 2015 3 / 24
What We Do 1 Employ natural language processing Classify expressions in the monthly reports according to polarity (whether an expression is positive, negative, or neutral) modality (whether an expression is clear-cut, ambiguous, or subjective), and so on Correlations with the state of the economy Latent Dirichlet allocation (LDA) model 2 Construct a simple persuasion game model to understand the empirical observations as a consequence of strategic communication KKSU Strategic Communications Nov 2015 4 / 24
What We Find The dierence between the numbers of positive and negative expressions in the reports leads the leading index of the economy by approximately three months. Ambiguous 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. KKSU Strategic Communications Nov 2015 5 / 24
Data KKSU Strategic Communications Nov 2015 6 / 24
Monthly Reports Jan 1998 to Mar 2015 (# of obs: 207) In Japanese (English with some delay) Advantages Monthly Rich Released on the next day of the MPMs 150 sentences Stylized and clear Japanese Divided into 4 parts: Summary (2-3 pages); economic developments; prices; nancial developments Summary is further divided into 2 parts: current situations; forecasts Modal logic (ambiguity) is closely linked to grammar. KKSU Strategic Communications Nov 2015 7 / 24
Basic Statistics Sentences Morphemes Mor/sen mean (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) KKSU Strategic Communications Nov 2015 8 / 24
60 50 40 Sentences (left axis) Morphemes/sentences (left axis) Morphemes (right axis) 2000 1500 30 1000 20 10 500 0 199801 199809 199905 200001 200009 200105 200201 200209 200305 200401 200409 200505 200601 200609 200705 200801 200809 200905 201001 201009 201105 201201 201209 201305 201401 201409 0 KKSU Strategic Communications Nov 2015 9 / 24
Top 5 Polarity Expressions Positive Negative experience # evaluation # experience # evaluation # demand 1062 good 205 fall 522 excess 147 improve 733 good/reason 92 decline 397 weak 131 fund 626 ease 85 price 335 minus 108 recovery 501 ample 62 worsen 181 sluggish 88 economy 478 grow 60 cost 150 weak 45 Even # invest 1166 increase 1121 environment 891 modest 802 produce 541 KKSU Strategic Communications Nov 2015 10 / 24
Modality Modality (or modal logic) is a grammatical concept in linguistics and typically known as auxiliary verbs such as must, would, and may in English. Highly indicative of ambiguity. Closely tied with grammar in Japanese, rather than each individual word. High probability Low probability Unreal Examples (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. KKSU Strategic Communications Nov 2015 11 / 24
Correlations with Business Cycles Count the number of interested morphemes each month Normalized by the total number of morphemes (except for morphemes/sentences) Compute correlations with economics indicators 3 composite index; ination rate; monetary policy dummy (1: tighten, -1:eased) Real-timed A correlation is signicantly dierent from zero at a 1 and 5 percent level, if its absolute value exceeds 0.179 and 0.137, respectively for the sample size of 207. KKSU Strategic Communications Nov 2015 12 / 24
Correlations (current economy) leading coincident lagging ination mdummy mbigdummy mor/sen -0.73-0.61-0.46-0.14-0.05-0.06 high prob -0.37-0.33-0.15 0.14 0.04-0.03 low prob unreal -0.61-0.44-0.31-0.08-0.09-0.09 pos neg 0.57 0.38-0.03-0.14 0.20 0.15 increase 0.65 0.67 0.69 0.21 0.28 0.24 decrease -0.41-0.40-0.10-0.29-0.06-0.01 etc -0.55-0.39-0.24-0.10-0.08-0.04 wa 0.67 0.60 0.56 0.04 0.18 0.13 KKSU Strategic Communications Nov 2015 13 / 24
Correlations (forecasts) leading coincident lagging ination mdummy mbigdummy mor/sen -0.54-0.40-0.35-0.06-0.09-0.06 high prob 0.60 0.42 0.33 0.10 0.09 0.08 low prob -0.39-0.28-0.19-0.22-0.05-0.03 unreal -0.31-0.17-0.09 0.08 0.01-0.01 pos neg 0.14 0.16 0.05 0.19 0.21 0.06 increase 0.79 0.67 0.54 0.12 0.24 0.20 etc -0.01-0.01-0.12 0.02 0.15 0.07 wa 0.35 0.19 0.02-0.02-0.05-0.04 KKSU Strategic Communications Nov 2015 14 / 24
0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0-0.01-0.02 199801 199809 199905 200001 200009 200105 200201 200209 200305 200401 200409 200505 200601 200609 200705 200801 200809 200905 201001 201009 201105 201201 201209 201305 201401 201409 120 100 80 60 40 20 0 pos-neg (current) pos-neg (forecast) leading KKSU Strategic Communications Nov 2015 14 / 24
0.6 Correlations bw positive-negative and leading index(+i) 0.4 0.2 0-0.2-6 -5-4 -3-2 -1 0 1 2 3 4 5 6 Current Forecast KKSU Strategic Communications Nov 2015 14 / 24
0.05 0.045 0.04 0.035 0.03 0.025 0.02 0.015 0.01 0.005 0 high prob (current) high prob (forecast) leading 199801 199809 199905 200001 200009 200105 200201 200209 200305 200401 200409 200505 200601 200609 200705 200801 200809 200905 201001 201009 201105 201201 201209 201305 201401 201409 120 100 80 60 40 20 0 KKSU Strategic Communications Nov 2015 15 / 24
LDA Not today Result Modality is used in tandem with negative expressions when the leading index is low. KKSU Strategic Communications Nov 2015 15 / 24
Model KKSU Strategic Communications Nov 2015 16 / 24
Model Want to explain In good (bad) economic circumstances, BOJ makes clear (ambiguous) statements. Model of CB's incentive (not today) When the economy is bad, the CB has an upward bias (or inward bias). This is aimed for preventing self-fullling deation. In Japan, the ination rate has been below 2% target. Game theoretic models to explain communication strategy persuasion game (veriable disclosure) model (Milgrom (1981), Shin (2003)) KKSU Strategic Communications Nov 2015 17 / 24
Persuasion Game Model Sender (CB) and receivers (P) Key assumptions Upward bias of CB CB's signals are veriable. CB's lie is too costly. Actual macroeconomic data come out later. Transcripts after 10 years Three states of macroeconomy y { 1,0,1} each with strictly positive probability partially or completely known to CB, but unknown to market Market reaction: V = E[y m] m is CB's report about macroeconomy market's maximization problem is not explicitly modelled here: e.g. it may have quadratic loss function (y V ) 2 CB's utility: V (upward bias) KKSU Strategic Communications Nov 2015 18 / 24
CB's Private Signals Two signals: S {S L, S H } and s { 1,0,1} CB receives ambiguous signal S with probability 1 if S = S L, then y { 1,0} y may be low if S = S H, then y {0,1} y may be high In addition, CB receives clear signal s with probability θ (0, 1) if s = x, then y = x θ represents how well CB is informed (common knowledge) KKSU Strategic Communications Nov 2015 19 / 24
CB's signals are veriable. CB's lie is costly. CB can report signals truthfully or withhold them. When CB doesn't report s, market can NOT tell whether CB does not know s; OR CB knows s but is hiding it CB's reporting strategy S (ambiguous) or s (clear) rule out reporting both signals, wlog KKSU Strategic Communications Nov 2015 20 / 24
Equilibrium Proposition In the unique perfect Bayesian equilibrium i) if clear signal s is not observed, then CB's report is ambiguous ii) if s is observed, then CB m = s only when s = 1; or when s = 0 and S = S L Remark: CB never reports s = 1 if s = 1, then CB hides it and reports ambiguous message m = S L instead negative messages are always ambiguous Remark: If taken literally, CB's reports are upward biased: s = 1 S L and s = 0 S H (when S H is observed) But market is never deceived (rational expectation) KKSU Strategic Communications Nov 2015 21 / 24
Intuition Suppose CB has received s = 1 Also, if CB reports s, then V = 1 if CB hides s and reports S L (note that by construction S = S L since y = 1), then V = E[y m = S L ] > 1, since market cannot tell whether CB hasn't received s { 1,0} or has received sbut hidden it thus if s = 1 is observed, CB is better o hiding it E[y m = S L ] < 0, so if S = S L and s = 0, then CB reports s = 0 E[y m = S H ] > 0, so if S = S H and s = 0, then CB hides s and reports S = S H E[y m = S H ] < 1, so if S = S H and s = 1, then CB reports s = 1 KKSU Strategic Communications Nov 2015 22 / 24
Comparison: Model of CB's Cheap Talk Crawford and Sobel (Etrica), Stein (AER) No direct eect of communication on payo No cost; no commitment Fully revealing perfect Bayesian equilibrium (PBE) does not exist. Fully pooling PBE exists. Partially pooling PBE may exist. CB partially reveals its information. Ambiguous communication, whose degree depends on state. Provided an upward bias, the cheap talk model predicts Ambiguity decreases (increases) in a bad (good) state!!! To make CB's announcement credible. KKSU Strategic Communications Nov 2015 23 / 24
Direction of Future Research Robustness Other forms of publications. In the English version? Publications by the Federal Reserve or other central banks KKSU Strategic Communications Nov 2015 24 / 24