First Impressions Matter: Signalling as a Source of Policy Dynamics

Transcription

1 First Impressions Matter: Signalling as a Source of Policy Dynamics Stephen Hansen 1 and Michael McMahon 2 1 Universitat Pompeu Fabra and GSE 2 University of Warwick and CEP 13 January 2012 MPC Dynamics 13 January / 1

2 Introduction Policymakers often act repeatedly over several years repeatedly make similar technical policy decisions. We study the behaviour of unelected policymaking experts on the Bank of England MPC: 1. may receive additional information as they accumulate experience 2. incentives might change over time We study the MPC to understand the extent to which either of these two forces plays a role in policy dynamics: crucial for designing the institutional environment. MPC Dynamics 13 January / 1

3 Key Facts Details The MPC has met monthly since June 1997 we stop sample at March 2009 There is a clear objective function (hit 2% target) Little emphasis on consensus MPC formed of five internal members (Governor, two Deputy Governors, two Executive Directors) and four external members (anybody) 12 internals and 14 externals in our sample Prior to voting, members receive extensive staff reports and discuss among themselves No prescribed voting order, and unattributed minutes MPC Dynamics 13 January / 1

4 Dynamic MPC Voting I vote it member i s desired change in interest rate in period t if member i votes for an increase of 25bps vote it = 0.25 We define a dummy variable to indicate when a member has completed 18 meetings: { 0 if member i has served in 18 or less meetings D(Experienced) it = 1 if member i has served in more than 18 meetings Reduced form dynamic voting behavior: vote it = α i + γd(experienced) it + δ t + ɛ it MPC Dynamics 13 January / 1

5 Dynamic MPC Voting II (1) (2) (3) Vote Vote Vote D(Experienced) * [0.067] D(Experienced - 12M) * [0.093] D(Experienced - 24M) * [0.092] Constant 0.24*** 0.24*** 0.24*** [0.000] [0.000] [0.000] R-squared Number of members Model Panel OLS Panel OLS Panel OLS Member effects? FE FE FE Time effects? YES YES YES MPC Dynamics 13 January / 1

6 So what drives this behaviour? Is it obvious? MPC Dynamics 13 January / 1

7 So what drives this behaviour? Is it obvious? We examine a potential explanations: 1. Is it learning-by-doing? 2. Do MPC members learn about the structure of the macroeconomy? 3. Are members trying to signal hawkishness to anchor the public s inflation expectations? 4. Is it career concerns? 5. Are new members just trying to fit in? 6. Could regulatory capture be driving the behaviour? MPC Dynamics 13 January / 1

8 So what drives this behaviour? Is it obvious? We examine a potential explanations: 1. Is it learning-by-doing? 2. Do MPC members learn about the structure of the macroeconomy? 3. Are members trying to signal hawkishness to anchor the public s inflation expectations? 4. Is it career concerns? 5. Are new members just trying to fit in? 6. Could regulatory capture be driving the behaviour? MPC Dynamics 13 January / 1

9 Related Literature by Other People Macroeconomic learning in Monetary Policy??,?,?,?,?,? and? Policy-motivated signalling in monetary policy?,?,?,?,? and?? and? Career concerns?,? and? Regulatory Capture? and? MPC Dynamics 13 January / 1

10 Member i in time t solves Model of Individual Decision Making ] max E[ (π t π ) 2 v it {0,1} where E[ π t ] = ω t + θ v it and V (π t ) <. v it is a vote ω t {0, 1} represents economic conditions - member i belief is ŵ it θ represents structural sources of inflation - member i belief is θ i Member i chooses high rate if and only if ) ω it 1 ( θi π θ i. Higher θ i means less evidence needed of high shock to vote high preferences or a structural parameter. MPC Dynamics 13 January / 1

11 Information We adopt a Bayesian approach to deriving the belief ŵ it Prior belief q t = Pr [ ω t = 1 ] that reflects public information Private signal s it N(ω t, σ 2 i ) σ i = means ŵ it = q t ; σ i = 0 means ŵ it = s it Since normal distribution satisfies MLRP, member i votes high if and only if s it > s it (q t, θ i, σ i ) MPC Dynamics 13 January / 1

12 Key Result from Earlier Paper Proposition When θ i increases, the probability that member i chooses v it = 1 increases for all q t ; when σi 2 increases, the probability that member i chooses v it = 1 increases if and only if q t > 1 θ i. MPC Dynamics 13 January / 1

13 Differences in Preferences 1 Theoretical Model of Voting Behaviour: Effect Of Higher θ Pr(v it = 1) θ = 0.55, σ = θ = 0.41, σ = q t (Prior) MPC Dynamics 13 January / 1

14 Differences in Information 1 Theoretical Model of Voting Behaviour: Effect Of Higher σ Pr(v it = 1) θ = 0.55, σ = θ = 0.55, σ = q t (Prior) MPC Dynamics 13 January / 1

15 Estimating the parameters Details Our model generates the likelihood function L it that member i votes for the high rates in time t as ( )) ( ( )) s q t (1 Φ it 1 s σ L it = i + (1 q t )Φ 1 it σ i if v it = 1 ( ) ( ) s q t Φ it 1 s σ i + (1 q t )Φ it σ i if v it = 0 We use q t and v it to structurally estimate our model (assuming θ and σ are constant among internal and external members). The Value of Information in the Court: Get it Right, Keep it Tight by Matias Iaryczower and Matthew Shum (AER, forthcoming) MPC Dynamics 13 January / 1

16 Potential Explanations for the Dynamic Behaviour 1. Is it learning-by-doing? 2. Do MPC members learn about the structure of the macroeconomy? 3. Are members trying to signal hawkishness to anchor the public s inflation expectations? 4. Is it career concerns? 5. Are new members just trying to fit in? 6. Could regulatory capture be driving the behaviour? MPC Dynamics 13 January / 1

17 Which Parameter Drives the Change Which parameter evolves with time? An information story σ Learning by doing which increases individual s expertise makes them more likely to vote against the public prior A revealed hawkishness story if θ changes θ to get more dovish MPC Dynamics 13 January / 1

18 Reduced form regression Prob(v=1) New Experienced Prior (q) MPC Dynamics 13 January / 1

19 Structural estimates Table: Structural estimates of the effect of experience on MPC voting (1) (2) θ σ New 0.57*** 0.97*** [ ] [ ] Experienced 0.48*** 1.04*** [ ] [ ] Difference *** [ ] [ ] Clustered, by member, confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 MPC Dynamics 13 January / 1

20 Why θ changes? The Idea of Macro-learning θ is an uncertain structural parameter of the macroeconomy the average belief about it changes over time Imagine member i votes in two periods t = 1 and t = 2 θ1 i be her prior belief of θ in period 1. Some new information arrives after period 1 decision She updates her belief to some new value θ2 i Beliefs formed using Bayes Rule are martingales MPC Dynamics 13 January / 1

21 Why θ changes? The Idea of Macro-learning θ is an uncertain structural parameter of the macroeconomy the average belief about it changes over time Imagine member i votes in two periods t = 1 and t = 2 θ1 i be her prior belief of θ in period 1. Some new information arrives after period 1 decision She updates her belief to some new value θ2 i Beliefs formed using Bayes Rule are martingales but...average θ2 i across members could fall from average θi 1 if: the initial average belief on θ were biased upwards, θ itself were a concave function of some uncertain parameter about which learning occurred - beliefs about θ would form a supermartingale. MPC Dynamics 13 January / 1

22 Common Learning θ i 1 θ i 2 θ i 2 E [ θ i 2 θ i 1 ] = θ i 1 θ i 2 t=1 t=2 MPC Dynamics 13 January / 1

23 Common Learning θ i,bias 1 θ i 2 θ i 2 E [ θ i 2 ] θ i,bias 1 θ i 2 t=1 t=2 MPC Dynamics 13 January / 1

24 Common Learning θ t θ New 1 θ Exp 2 θ2 New θ Exp 3 θ3 New t=1 t=2 t=3 MPC Dynamics 13 January / 1

25 Common learning I ω t is partly discovered through as it is a frequent, transitory shock not fully revealed in public data. Learning about a structural parameter takes place more slowly. common learning is a natural assumption? Test Do new members consistently have a higher θ than experienced members serving at the same time? if yes not common learning. if all members share the same belief on θ which gradually drifts down over time common learning = We use rolling regressions to test this effect MPC Dynamics 13 January / 1

26 Common learning II m1 2003m1 2004m1 2005m1 2006m1 2007m1 2008m1 2009m1 Date Upper 95% CI Experience effect Lower 95% CI Hawkishness New Hawkishness Experienced Figure: Rolling window estimates of θ for new and experienced members MPC Dynamics 13 January / 1

27 Private learning I Test 1 Individual beliefs can on average decline over time but some i with θi 2 > θi 1 A private learning model in which there were only a small probability that a member responded to new information by increasing his belief would require some unusual assumptions on the environment and/or behavior Estimate the experience effect for each individual separately Use randomly generated pseudo-committees draws MPC Dynamics 13 January / 1

28 Private Learning θ j 2 θ j 1 θ i 1 θ h 1 θ2 i j 2 θ2 h θ2 i j 2 θ2 h θ2 i θ2 h Var of θ i 2 Var of θi 1 t=1 t=2 MPC Dynamics 13 January / 1

29 Private learning II Table: Experience Effect by individual member Member θ New θ Exp Experience effect p-value Allsop Barker Bean Bell Clementi George Gieve Goodhart Julius King Large Lomax Nickell Plenderleith Sentence Tucker Wadhwani MPC Dynamics 13 January / 1

30 Private learning III Kernel Density Estimated experience effect in different random samples of MPC members Experience effect Figure: Kernel Density plots of estimated experience effect using random combinations of members MPC Dynamics 13 January / 1

31 Private learning IV Test 2 Variance of θ2 i should be higher than the variance of θi 1 Conditional on θ1 i, θi 2 is a r.v. With private learning, the individual values of θ2 i are not perfectly correlated Can again use our quasi-committees MPC Dynamics 13 January / 1

32 Private learning V Kernel Density Estimated Hawkishness in different random samples of MPC members Experienced Hawkishness New Hawkishness Figure: Kernel Density plots of estimated θ for new and experienced members using random combinations of members MPC Dynamics 13 January / 1

33 2-period Model of Signalling Reputation I θ i is a fixed preference parameter known only to member i new MPC members want to signal that they are hawkish in order to anchor inflation expectations the evaluator places probability r on θ i = θ H θ L as a Dove and type θ H as a Hawk Let be the net reputational reward to voting high in 1st period Let β measure the strength of the signalling incentive in 1st period we assume reputational concerns in the 1st period but not the 2nd as no future after 2nd period in which reputation pays off Derive similar cut-off voting rules: ω i1 1 θ i β ω i2 1 θ i use Perfect Bayesian Equilibrium - responsive equilibrium MPC Dynamics 13 January / 1

34 2-period Model of Signalling Reputation II When we average over all new members and experienced members, we get, respectively, the average θ is: rθ H + (1 r)θ L + β rθ H + (1 r)θ L 1. At any given point in time, new members would have a higher estimated θ than experienced ones 2. All individual experience effects would be negative 3. Average distance between new and experienced members is the same no variance increase 4. If do for a hawk or a dove, the difference should be the same - β contrast with give-in story and learning convergence MPC Dynamics 13 January / 1

35 2-period Model of Signalling Reputation III Tests of Signalling 1. Test Hawk-Dove differences Identify hawks and doves in the data Calculate and compare the experience effect using both overall structural estimates and pseudo-committee approach 2. Consider multi-period signalling prediction of declining effect of reputation over time MPC Dynamics 13 January / 1

36 Identifying Hawks and Doves Average theta across random groups Blanchflower Wadhwani Julius Allsop Bell Nickell Lambert Lomax Barker Bean King Tucker George Plenderleith Clementi Gieve Large Besley Sentence Goodhart Hawks Doves Individual theta Average theta across random groups Blanchflower Julius Wadhwani Bell Allsop King Besley Sentence Goodhart Tucker Vickers Budd Buiter Plenderleith George Clementi Dale Gieve Nickell Lomax Lambert Barker Bean Large Walton Hawks Doves % of votes that are high Davies (a) Relationship between two measures of individual hawkishness (b) Relationship between hawkishness and % of votes that are high Figure: Ranking individual MPC members based on individual hawkishness MPC Dynamics 13 January / 1

37 Experience Effect for Hawks and Doves I Table: The experience effect for Hawks and Doves (1) (2) (3) Dove Hawk Difference New 0.44*** 0.68*** -0.24*** [ ] [ ] (0.000) Experienced 0.37*** 0.56*** -0.19*** [ ] [ ] (0.000) Difference *** [ ] [ ] (0.351) Clustered, by member, 95% confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 MPC Dynamics 13 January / 1

38 Experience Effect for Hawks and Doves II Kernel Density Experience effect estimated in different random samples of MPC members Hawks Experience effect Doves Experience effect Figure: Kernel density plots of experience effect on random pseudo-committees MPC Dynamics 13 January / 1

39 Timing of the Experience Effect Table: Effect of each year of experience on θ (1) (2) θ θ First 12 months 0.61*** First 12 months 0.62*** [ ] [ ] 12 to 24 Months 0.51*** 12 to 24 Months 0.53*** [ ] [ ] 24+ Months 0.48*** 24 to 36 Months 0.49*** [ ] [ ] 36+ Months 0.50*** [ ] Experience Effects Experience Effects Year 2 - Year ** Year 2 - Year ** [ ] [ ] Year 3 - Year *** Year 3 - Year ** [ ] [ ] Year 4 - Year *** [ ] Clustered, by member, confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 MPC Dynamics 13 January / 1

40 Other Signalling Could be another, non-policy, motivation for signalling hawkishness. We examine a few potential stories: 1. career concerns - signal to advance their careers 2. signal to other members - justify your position 3. regulatory capture MPC Dynamics 13 January / 1

41 Career Concerns I Table: Internal and external members θ Estimates (1) (2) (3) D(External) i = 0 D(External) i = 1 Difference New 0.60*** 0.54*** [ ] [ ] (0.496) Experienced 0.54*** 0.34*** 0.19** [ ] [ ] (0.016) Difference *** -0.20** 0.14 [ ] [ ] (0.155) Robust 95% confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 Column (1)-(3), H0: Estimate = 0 MPC Dynamics 13 January / 1

42 Career Concerns II Table: Academic members θ Estimates (1) (2) (3) D(Academic) i = 0 D(Academic) i = 1 Difference New 0.57*** 0.60*** [ ] [ ] (0.779) Experienced 0.51*** 0.39*** 0.12 [ ] [ ] (0.115) Difference ** -0.22* 0.16 [ ] [ ] (0.181) Robust 95% confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 Column (1)-(3), H0: Estimate = 0 MPC Dynamics 13 January / 1

43 Career Concerns III Table: Private Sector θ Estimates (1) (2) (3) D(Private Sector) i = 0 D(Private Sector) i = 1 Difference New 0.62*** 0.49*** 0.13* [ ] [ ] (0.052) Experienced 0.51*** 0.42*** [ ] [ ] (0.256) Difference -0.11*** [ ] [ ] (0.582) Robust 95% confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 Column (1)-(3), H0: Estimate = 0 MPC Dynamics 13 January / 1

44 Signal to each other? Table: Experience Effect by individual member Member θ New θ Exp Experience effect p-value Goodhart King MPC Dynamics 13 January / 1

45 Regulatory capture I Not signalling but direct influence by financial insitutions - regulatory capture. Financial industry apply pressure to pursue cheap money? After 18 months on the MPC, members eventually give in to the lobbying effort of the financial industry Test of Regulatory Capture We use the Times MPC data - replicate the decision each month No actual impact of decision. If regulatory capture, Times MPC votes systematically higher 5 MPC members have served on the Times MPC Budd, Goodhart, Lambert, Sentence, and Wadhwani MPC Dynamics 13 January / 1

46 Regulatory capture II Table: Voting on the MPC and the Times MPC: Fraction of High Votes (%) and count (n) of total votes on different committees Member Times 1st MPC New MPC Exp Times 2nd (%) (n) (%) (n) (%) (n) (%) (n) Budd Goodhart Lambert Sentence Wadhwani MPC Dynamics 13 January / 1

47 Regulatory capture II Table: Structural estimates of θ on the MPC and the Times MPC New 0.69*** [ ] Experienced 0.58*** [ ] Times 0.52*** [ ] Difference [ ] Clustered, by member, 95% confidence interval in brackets *** p<0.01, ** p<0.05, * p<0.1 θ MPC Dynamics 13 January / 1

48 Conclusion We examine a number of potential explanations for dynamic MPC voting behaviour: 1. Is it learning-by-doing? 2. Do MPC members learn about the structure of the macroeconomy? 3. Are members trying to signal hawkishness to anchor the public s inflation expectations? 4. Is it career concerns? 5. Are new members just trying to fit in? 6. Could regulatory capture be driving the behaviour? MPC Dynamics 13 January / 1

49 Conclusion We examine a number of potential explanations for dynamic MPC voting behaviour: 1. Is it learning-by-doing? 2. Do MPC members learn about the structure of the macroeconomy? 3. Are members trying to signal hawkishness to anchor the public s inflation expectations? 4. Is it career concerns? 5. Are new members just trying to fit in? 6. Could regulatory capture be driving the behaviour? Future work: Examine the effect of early member voting on inflation expectations Pursue the study of policy-motivated signalling in other policy settings MPC Dynamics 13 January / 1

50 THE END MPC Dynamics 13 January / 1

51 The Monetary Policy Committee (MPC) The Bank of England MPC has been meeting monthly since June 1997 to decide UK interest rates The Bank of England Act (1998) mandated an inflation target of 2.0% (2.5% until Jan 2004) The committee is independent in the sense that each member is free to put forth his or her own view: Members do not represent individual groups or areas. They are independent. Each member of the Committee has a vote to set interest rates at the level they believe is consistent with meeting the inflation target. The MPC s decision is made on the basis of one-person, one vote...it reflects the votes of each individual member of the Committee. MPC Dynamics 13 January / 1

52 9 members: 1 Governor 2 Deputy Governors 2 x Executive Directors 4 external members Personnel Governor and Deputy Governors appointed by Chancellor and HMT to serve 5 year terms Chief Economist and Director for Financial Markets appointed by Governor and Court of Directors (subject to Parliamentary approval) for 3 year term. External members appointed by Chancellor and HMT to serve 3 year terms (discuss reappointment later) Through March 2009, 14 external members have served on the committee and 12 internal members MPC Dynamics 13 January / 1

53 Meetings The MPC meets on the first Wednesday and Thursday of every month. 1. The Friday before the meeting the members receive a briefing from Bank of England staff about economic conditions 2. On Wednesday the members gather to discuss their views, with the discussion resuming on Thursday morning 3. Each member summarizes his or her view to the rest of the committee and suggests what vote they favor. 4. A formal vote is taken. 5. Votes and unattributed minutes released after 2 weeks. Return MPC Dynamics 13 January / 1

54 Constructing v it Table: Number of Unique Votes Cast in Each Meeting Frequency Percent (%) Need to know which other interest rate was considered; 2. v it = 1 for higher of two rates; 3. Need to know which are the main two rates under consideration. MPC Dynamics 13 January / 1

55 Constructing Public Beliefs The key empirical challenge is to come up with a measure of the public prior We draw on three sources of data: Reuters survey data Times MPC Financial market data We also use the Reuters data to identify the second vote when committee decision is unanimous MPC Dynamics 13 January / 1

56 Reuters Data Table: Example of Survey Data +50bps +25bps 0-25bps -50bps UBS 15% 80% 5% Goldman Sachs 20% 75% 5% JP Morgan 45% 45% 10% AIB 15% 85% Average 23.75% 71.25% 5% Our proxy measure for the prior would be q R t = = 0.25 MPC Dynamics 13 January / 1

57 Times MPC Times MPC is group of experts (some former and future MPC members) that vote each month in the Times of London We construct q t T as the fraction of Times MPC votes for the higher of the two rates under consideration Data only exist from November 2002 MPC Dynamics 13 January / 1

58 Financial Market Data Construct q t M as the conditional probability that risk-neutral trader attaches to high rate change Interest Rate (level, pp) m1 2000m1 2003m1 2006m1 2009m1 Date Current Libor 25% to 75% 45% to 55% 5% to 95% Figure: Short-Sterling Implied PDF and 3 Month LIBOR MPC Dynamics 13 January / 1

59 Combining the Three Measures Each of these three measures has strengths and weaknesses We therefore use common factor analysis to construct the common component driving the variation in all three We find there is one common factor that drives the majority of the variation We then use this common factor as a measure of public beliefs ( q RMT based on all three measures; q RM based on all two measures) MPC Dynamics 13 January / 1

60 Key Empirical Variables Prior Probability that State is High Prior Probability that State is High m1 2000m1 2003m1 2006m1 2009m1 Date 1997m1 2000m1 2003m1 2006m1 2009m1 Date Reuters q Market q Times q Loosening Cycle Extracted Belief RMT Extracted Belief RM Loosening Cycle (a) Measures of the Prior q t (b) Extracted Beliefs MPC Dynamics 13 January / 1

61 Is the transformed principal common factor a belief? We use two preliminary checks: v it = α + ψ 1 q x t (2) Deviates from MPC it = α + ψ 1 q x t + ψ 2 ( q x t ) 2 + ε it (3) MPC Dynamics 13 January / 1

62 Checks on the Behavior of q t I Table: Estimates of Equation (??) (1) (2) (3) (4) v it v it v it v it q RMT 6.35*** 8.58*** (0.00) (0.00) q RM 6.74*** 8.56*** (0.00) (0.00) Constant -3.04*** -4.14*** -3.27*** -4.85*** (0.00) (0.00) (0.00) (0.00) Observations ,201 1,198 Estimation Panel Logit Panel Logit Panel Logit Panel Logit Member effects None FE None FE p-value in parentheses *** p<0.01, ** p<0.05, * p<0.1 MPC Dynamics 13 January / 1

63 Checks on the Behavior of q t II Table: Estimates of Equation (??) (1) (2) (3) (4) D(Deviate) D(Deviate) D(Deviate) D(Deviate) q RMT 6.25*** 6.99*** (0.00) (0.00) ( q RMT ) *** -6.98*** (0.00) (0.00) q RM 3.16* 3.30* (0.07) (0.08) ( q RM ) ** -3.29* (0.05) (0.07) Constant -3.09*** *** -1.90*** (0.00) (0.64) (0.00) (0.01) Observations ,201 1,087 Estimation Panel Logit Panel Logit Panel Logit Panel Logit Member effects None FE None FE p-value in parentheses *** p<0.01, ** p<0.05, * p<0.1 MPC Dynamics 13 January / 1

64 References MPC Dynamics 13 January / 1