Choose your words wisely: Analysing RBI s monetary policy communication Aakriti Mathur (IHEID, Geneva) Rajeswari Sengupta (IGIDR, Mumbai) Preliminary. Please do not cite. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 1 / 47
Contents Motivation and Objective Questions Data and methodology Preliminary results Next steps Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 2 / 47
Part I Motivation Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 3 / 47
Motivation An algorithm to track Australia s central bank (Bloomberg, Aug 29, 2017) The latest index for Australian market watchers do not track stocks, bonds or currencies. It uses an algorithm to track the tone of central bank Governor Philip Lowe s words. LSE and BOE released a handbook of textmining for central banks in June 2015. In the US, minutes of FOMC meetings are closely watched and analysed by researchers. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 4 / 47
Motivation Central bank communication: Why is it important? The aim of monetary policy conducted by a central bank is to influence agents inflation expectations. The central bank controls the short term nominal interest rate. A change in the short term rate by itself does not affect economic decisions of agents in a significant manner. Agents form their views about future economic developments and responses of the central bank to these developments. They look at CB communication: information made publicly available by the CB about its current and future policy objectives, current economic outlook and likely path of future monetary policy decisions. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 5 / 47
Motivation Central bank communication: Why is it important? Information content of monetary policy changes i t = φf (Γ t, Γ e t+j t ) + ɛ t (1) The CB reveals i t, the nominal rate at time t when it announces its monetary policy decision. i t is based on φ reaction function to the current state (Γ t ) and future expectations (Γ e t ) of economic inputs. The reaction function φ itself is quantitatively unobserved. Γ t are revealed in several ways, mostly through monetary policy statements or press conference held by Governor immediately after the decision. Forecasts of economic inputs (Γ e t ) are also revealed through the same fora. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 6 / 47
Motivation Central bank communication: Why is it important? Information content of monetary policy changes CB informs the agents inflation expectations through the following: 1 Its decision of i t 2 Its belief about current and future economic outlook (Γ t and Γ e t ) 3 Forward guidance: Its expected deviations from the average rule (ɛ t) due to risks to various forecasts. Our objective is to analyse the last two aspects of RBI s communication using NLP tools. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 7 / 47
Motivation Objective 1 To quantify the content of RBI s MP communication using techniques from computational linguistics. 2 To quantify the impact of RBI s MP communication on financial market variables. Advantage: automated tools enable scaling up across a large database of texts in a consistent manner. Challenge: converting raw communication, in words, into meaningful quantities which can be analysed. Importance: broader applications beyond monetary policy analysis; brings economics close to the world of big data. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 8 / 47
Monetary Policy in India Brief background Motivation From 1998 to 2015, RBI looked at multiple indicators to decide the interest rate (such as exchange rate, trade balance, inflation, output, unemployment etc). Governors Bimal Jalan, YV Reddy, Subbarao and Rajan belonged to this era. In 2015 February RBI shifted to an inflation targeting framework under Rajan. The formal operating procedure of IT was not in place for a year. The monetary policy committee was formed and met for the first time in October 2016. April 1998-October 2016: RBI communication was through Governor s statements. October 2016 onward: RBI communication has been through MPC s statements. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 9 / 47
Part II Questions Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 10 / 47
Hypotheses Quantification of content: Question I Testing the move to IT/ MPC 1 Is the move to inflation targeting reflected in the RBI s communications? Has there been a semantic shift in the statements after inflation targeting was adopted and MPC started functioning in October 2016? Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 11 / 47
Hypotheses Quantification of content: Question II Testing Governor fixed effects 1 How has the tone of communication of RBI s monetary policy strategy changed across regimes? How much has been the forward looking element of the communication of respective governors and the MPC vis-a-vis the emphasis on current economic outlook? Is it possible to discern the respective hawkish vs. dovish Governors and MPC members from the statements? Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 12 / 47
Hypotheses Quantification of impact: Question III Testing effects of communication on asset prices 1 Are there significant and/or persistent quantitative effects of RBI s statements on asset prices? Effect on yields in the bond market, returns and volatility in the equity market. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 13 / 47
Part III Data and methodology Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 14 / 47
No. of regimes Data and methodology Data sources and cleaning Governor/ Regime Term Dr. D. Subbarao 2008-09 to 2013-09 Dr. Raghuram Rajan 2013-09 to 2016-09 Monetary Policy Committee 2016-10 to Present Table: Governors of the RBI, 2008-2017 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 15 / 47
Data and methodology Data sources and cleaning Data sources: Subbarao Subbarao: Sep 2008 - Sep 2013 Quarterly statements July (1st qtr), October (2nd qtr), January (3rd qtr) Annual policy statement April/May Mid quarter press releases Started Sep 2010; June, September, December, March. Press conference transcripts 2009, 2010 (April: annual policy statements); July 2010 onward for every quarterly MP review. Teleconference with researchers Started Jan 2010 for every quarterly review; April 2010 unedited and edited versions. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 16 / 47
Data and methodology Data sources and cleaning Data sources: Rajan Rajan: Sep 2013 - Sep 2016 Only press release and transcripts Sep 2013 only Quarterly statements Oct 2013 - Jan 2014 Bi-monthly statements Started April 2014; April, June, Aug, Sep, Dec, Feb Mid quarter press releases Dec 2013 only Press conference transcripts Sep 2013 (missing in June 2014) Teleconference with researchers Sep 2013 onward (missing in June 2014) Governor s statement on MP Jan and March 2015 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 17 / 47
Data and methodology Data sources and cleaning Data sources: MPC MPC: Oct 2016 - Present Resolution of MPC Started Oct 2016 Press conference transcripts Started Oct 2016 Minutes of meetings Started Oct 2016 (released after 2 weeks) Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 18 / 47
Data and methodology Data sources and cleaning Data sources We use: All monetary policy statements by Subbarao and Rajan: October 2008 to August 2016. All statements and minutes of meetings of the MPC: October 2016 to August 2017. We do not use: Mid quarter press releases (Subbarao), press conferences (Subbarao, Rajan, MPC), teleconference with analysts (Subbarao, Rajan). Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 19 / 47
No. of regimes Data and methodology Data sources and cleaning Governor/ Regime Duration of MP reviews Statements minutes Dr. D. Subbarao 2008-10-24 to 2013-07-30 20 - Dr. Raghuram Rajan 2013-10-29 to 2016-08-09 17 - Monetary Policy Committee 2016-10-04 to 2017-08-02 6 5 Table: Governors of the RBI, 2008-2017 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 20 / 47
Data and methodology Data sources and cleaning Data sources and cleaning We read the PDF files as text files. We create a corpus for each set of statements. We eliminate: puctuations (eg.!,.) stop words (eg. the, that, this ) numbers (eg. 0-9) other non-informative words (eg. another, bank, per cent ) Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 21 / 47
Data and methodology Methodology for quantifying content of communication Descriptives Present analysis: Complexity No. of words (raw counts) Trends Word clouds Associations Pairwise correlations across commonly occurring pairs of words Going forward: Visualisation Graphs with dates of domestic and external shocks Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 22 / 47
Part IV Descriptives Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 23 / 47
Descriptives Complexity No. of words in Subbarao s statements Complexity I (post-cleaning) 20000 15000 Oct 2008 Jan 2009 Apr 2009 Jul 2009 Oct 2009 Jan 2010 Apr 2010 Jul 2010 Nov 2010 Jan 2011 May 2011 Jul 2011 Oct 2011 Jan 2012 Apr 2012 Jul 2012 Oct 2012 Jan 2013 Jul 2013 Number 10000 5000 0 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 24 / 47
Descriptives Complexity No. of words in Rajan s statements Complexity II (post-cleaning) 4000 3000 Oct 2013 Jan 2014 Apr 2014 Jun 2014 Aug 2014 Sep 2014 Dec 2014 Feb 2015 Apr 2015 Jun 2015 Aug 2015 Sep 2015 Dec 2015 Feb 2016 Apr 2016 Jun 2016 Aug 2016 Number 2000 1000 0 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 25 / 47
Descriptives Complexity No. of words in MPC s statements Complexity III (post-cleaning) 2500 2000 Number 1500 1000 Oct 2016 Dec 2016 Feb 2017 Apr 2017 Jun 2017 Aug 2017 500 0 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 26 / 47
Descriptives Complexity No. of words in MPC s minutes Complexity IV (post-cleaning) 6000 5000 4000 Oct 2016 Dec 2016 Feb 2017 Apr 2017 Jun 2017 Number 3000 2000 1000 0 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 27 / 47
Descriptives No. of words across regimes Complexity Regime Mean words Dr. D. Subbarao 9251 Dr. Raghuram Rajan 2373 Monetary Policy Committee Statements 2142 Minutes 4332 Table: Summary statistics of document lengths, 2008-2017 - I Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 28 / 47
Descriptives Trends Most important words: Subbarao Trends I (using raw frequencies) conditions commercial risks prices global review rates <b7> repo food capital current account market economy markets government monetary credit financial price inflation risk growth rate india liquidity sector domestic demand guidelines committee central interest foreign system wpi measures expected net Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 29 / 47
Descriptives Trends Most important words: Rajan Trends II (using raw frequencies) demand conditions market guidelines frameworkmonetary prices repo food sector repos risks global india current account securities measures system rate liquidity inflation growth financial investment ndtl billion credit foreign activity issued government services markets exchange domestic net capital monsoon production Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 30 / 47
Descriptives Trends Most important words: MPC s statements Trends III (using raw frequencies) expectations momentum conditions activity markets production risksmpcliquidity services cpi food output prices inflation expected imports global index price repo marketgrowthrate domestic <ad> headline sector demand outlookemes months trillion monetary trade rates industrialfinancial investment crude Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 31 / 47
Descriptives Trends Most important words: MPC s minutes Trends IV (using raw frequencies) financial demonetisation impact likely production investment liquidity activity output outlookresolution expected repo fuel target cpi economy committee monetary rate food inflation growth prices headline high demand global mpc risks services industrial india ratesmeetingmonths expectations sector <ad> conditions Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 32 / 47
Descriptives Associations Associations with inflation How central is the word to the document? No. of words 0 100 200 300 400 500 600 Subbarao Rajan MPC statements MPC minutes 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Correlation Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 33 / 47
Part V Measuring tone Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 34 / 47
Sentiment analysis Sentiment analysis Very preliminary. Three in-built dictionaries in tidyr package in R 1 AFINN from Finn Arup Nielsen 2 bing from Bing Liu and collaborators 3 nrc from Saif Mohammad and Peter Turney We currently use bing as it is binary and each word is defined as positive or negative This reduces the number of words being analysed (on average to 1/10 th ) We obtain counts of each type of word and define sentiment of document d at time t as: sentiment d,t = n.positive d,t n.negative d,t (2) We then plot these against various series reflecting the economic outlook Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 35 / 47
Sentiment analysis Subbarao I Document Negative Positive Sentiment 1 2008Oct_Subbarao.pdf 501.00 491.00-10.00 2 2009Jan_Subbarao.pdf 209.00 210.00 1.00 3 2009April_Subbarao.pdf 392.00 656.00 264.00 4 2009July_Subbarao.pdf 166.00 188.00 22.00 5 2009Oct_Subbarao.pdf 301.00 438.00 137.00 6 2010Jan_Subbarao.pdf 79.00 122.00 43.00 7 2010April_Subbarao.pdf 180.00 365.00 185.00 8 2010July_Subbarao.pdf 69.00 142.00 73.00 9 2010Nov_Subbarao.pdf 156.00 304.00 148.00 10 2011Jan_Subbarao.pdf 93.00 132.00 39.00 11 2011May_Subbarao.pdf 199.00 306.00 107.00 12 2011July_Subbarao.pdf 108.00 92.00-16.00 13 2011Oct_Subbarao.pdf 186.00 263.00 77.00 14 2012Jan_Subbarao.pdf 127.00 109.00-18.00 15 2012April_Subbarao.pdf 188.00 288.00 100.00 16 2012July_Subbarao.pdf 113.00 67.00-46.00 17 2012Oct_Subbarao.pdf 187.00 281.00 94.00 18 2013Jan_Subbarao.pdf 102.00 95.00-7.00 19 2013July_Subbarao.pdf 88.00 90.00 2.00 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 36 / 47
Sentiment analysis Subbarao II Words contributing most to sentiment Contribution to sentiment 200 100 0 100 200 300 sentiment negative positive risk risks crisis inflationary decline crude debt issue gross concerns issues stress sharply negative limit adverse weak improvement positive easing maturity improved strong sharp respect appropriate available confidence support consistent well regard important effective cooperative recommendations significant advanced recovery stability term Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 37 / 47
Sentiment analysis Rajan I Document Negative Positive Sentiment 1 2013Oct_Rajan.pdf 37.00 82.00 45.00 2 2014Jan_Rajan.pdf 19.00 24.00 5.00 3 2014April_Rajan.pdf 51.00 103.00 52.00 4 2014June_Rajan.pdf 20.00 49.00 29.00 5 2014Aug_Rajan.pdf 26.00 42.00 16.00 6 2014Sep_Rajan.pdf 48.00 85.00 37.00 7 2014Dec_Rajan.pdf 50.00 63.00 13.00 8 2015Feb_Rajan.pdf 78.00 102.00 24.00 9 2015April_Rajan.pdf 70.00 126.00 56.00 10 2015June_Rajan.pdf 46.00 53.00 7.00 11 2015Aug_Rajan.pdf 43.00 61.00 18.00 12 2015Sep_Rajan.pdf 99.00 85.00-14.00 13 2015Dec_Rajan.pdf 34.00 42.00 8.00 14 2016Feb_Rajan.pdf 53.00 54.00 1.00 15 2016April_Rajan.pdf 68.00 145.00 77.00 16 2016June_Rajan.pdf 56.00 62.00 6.00 17 2016Aug_Rajan.pdf 47.00 62.00 15.00 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 38 / 47
Sentiment analysis Rajan II Words contributing most to sentiment Contribution to sentiment 40 0 40 80 risks weak crude decline risk limit marginal subdued issue debt fall downside volatility inflationary limits stress accommodative progress robust significant positive strong modest revival confidence gold supported available improve recommendations respect support appropriate sharp improving recovery eased regard improvement easing well term sentiment negative positive Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 39 / 47
Sentiment analysis MPC statement I Document Negative Positive Sentiment 1 Oct 16-17 23.00 42.00 19.00 2 Dec 16-17 44.00 45.00 1.00 3 Feb 16-17 26.00 50.00 24.00 4 Apr 17-18 46.00 62.00 16.00 5 Jun 17-18 47.00 61.00 14.00 6 Aug 17-18 46.00 55.00 9.00 Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 40 / 47
Sentiment analysis MPC statment II Words contributing most to sentiment Contribution to sentiment 20 0 20 sentiment negative positive risks crude decline risk subdued volatility falling weak fall loss marginal depressed fell gross slowed stress accommodative advanced boost durable gains optimism recovery soft support eased favour gold improved robust consistent easing significant supporting improvement positive savings sharp balanced evenly respect award supported improving strong well term Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 41 / 47
Sentiment analysis Weighting scheme Raw counts of words do not give us a sense of the importance of the word. There are words that may occur many times but are not informative: for e.g., the, is, of, that, although, along, statement, rather etc. tf-idf tells us the frequency of a term adjusted for how rarely it is used. It decreases the weight for commonly used words/ terms, and increases the weight for other words that are not very much used tf-idf weighting is represented as follows: tf idf = (1 + logf t,d ) log( D df t ) (3) f t,d : frequency of term t in document d; D: No. of documents; df t : number of documents. in which the term t appears. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 42 / 47
Sentiment analysis Pooled common terms using tf-idf Regime Total Average/ document Dr. D. Subbarao 3411 180 Dr. Raghuram Rajan 1625 96 Monetary Policy Committee Statements 709 120 Minutes 921 5 Table: Summary statistics of documents, 2008-2017 - II Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 43 / 47
Part VI Next steps Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 44 / 47
Next steps Proposed Methodology for quantifying content of communication Measuring topics Coverage Extend coverage of statements backwards to April 1998, covering 17 years and 5 Governors. Descriptives Repeat with the weighted count of words Clustering Measuring economic topics using Latent Dirichlet Allocation (LDA) Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 45 / 47
Next steps Proposed Methodology for quantifying content of communication Measuring tone Sentiment analysis Measuring tone using central banking specific dictionaries Assigning words positive or negative connotations Use existing dictionaries to discern the tone of statements: eg. Apel and Blix-Grimaldi 2012 (for CB communications) Loughran and McDonald 2011 (for financial context) General Inquirer s Harvard IV-4 Psychosocial dictionary Modify dictionaries to reflect India-specific terms Forward guidance Measuring amount and direction of guidance Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 46 / 47
Next steps Proposed Methodology for quantifying content of communication Thank you. Rajeswari Sengupta and Aakriti Mathur NLP MPC Preliminary. Please do not cite. 47 / 47