Central Bank Communication about Climate Change

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Central Bank Communication about Climate Change David M. Arseneau, Alejandro Drexler, Mitsuhiro Osada 2022-031 Please cite this paper as: Arseneau, David M., Alejandro Drexler, and Mitsuhiro Osada (2022). “Central Bank Communication about Climate Change,” Finance and Economics Discussion Series 2022-031. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2022.031. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Central Bank Communication about Climate Change * DavidM.Arseneau AlejandroDrexler MitsuhiroOsada FederalReserveBoard FederalReserveBank BankofJapan ofChicago FederalReserveBoard April15,2022 Abstract This paper applies natural language processing to a large corpus of central bank speeches to identifythoserelatedtoclimatechange. Weanalyzethesespeechestobetterunderstandhowcentralbanks communicateaboutclimatechange. Byallaccounts,communicationaboutclimatechangehasacceleratedsharplyinrecentyears. Thebreadthoftopicscoverediswide,rangingfromtheimpactofclimate change on the economy to financial innovation, sustainable finance, monetary policy, and the central bankmandate. Financialstabilityconcernsaretouchedupon,butmacroprudentialpolicyisrarelymentioned. Directcentralbankactionlargelyrevolvesaroundidentifyingandmonitoringpotentialrisksto thefinancialsystem. Finally,wefindthatcentralbankstendtousespeculativelanguagemorefrequently whentalkingaboutclimatechangerelativetoothertopics. Keywords: Financialstability; Transparency; Centralbankmandate; Greenfinance; Naturallanguage processing;Centralbankspeeches JELClassifications: E58;E61;Q54 *WethankAdeleMorris, RichardRosen, RicardoCorrea, andKojiNakamuraforvaluablecommentsandsuggestions. The views expressed in this paper are those of the authors and do not necessarily represent those of the Federal Reserve Board of GovernorsoranyoneintheFederalReserveSystem,nordotheyrepresenttheviewsoftheBankofJapan. 1

1 Introduction Climatechangeisanimportantpriorityforcentralbanksaroundtheworld. Asevidence,membershipinthe NetworkforGreeningtheFinancialSystem(NGFS)hasincreasedfromeightcentralbanksandsupervisors at its inception in late-2017 to over one hundred as of early-2022. In order to join the NGFS, members arerequiredtopubliclyacknowledgethatclimate-relatedrisksareapotentialsourceoffinancialinstability. As such, the number of central banks that have acknowledged that climate-related risks fall within their supervisoryandfinancialstabilitymandateshasexplodedduringashortfiveyearperiod. Atthesametime,climatechangeisacomplextopicthatisoutsidethetraditionalareasofexpertisefor centralbanks. Theinteractionbetweenclimateandeconomicactivityisnotwellunderstoodandoccursover a longer time horizon than central banks typically consider. Little is known about propagation of climaterelated risks through the financial system. More fundamentally, there are significant data gaps which pose challenges for risk management for micro- and macroprudential supervisors alike in addition to financial market participants more generally. Taken together, these complications put central banks in a difficult position. Howcantheyestablishcredibilityfor beingresponsivetoclimatechangewhenrelativelylittleis knownabouthowitinteractswiththerealandfinancialeconomy? One approach is to build credibility through communication. Central bank communication is increasingly important in achieving monetary policy objectives.1 Moreover, recent research has shown that how centralbanksapproachcommunicationmattersgreatly,notonlyforachievingdesiredpolicyobjectivesbut alsoformanagingpublicperceptionabouttheinstitutionitself. Forexample,HaldaneandMcMahon(2018) andKryvtsovandPetersen(2021)showthatcommunicationismoreeffectivewhencentralbanksusesimple, relatable language. Bholat, et al. (2019) shows that simple, relatable communication helps to build public trust and improves people’s perception of the central bank. This research suggests that how central banks approach talking about climate change is crucially important as they learn more about its effects on financial markets and the broader economy. At this point, however, we know very little about how central banks approach communication about climate change. This paper takes a first step toward filling this gap in the literature. We apply natural language processing techniques to a corpus of over 17,000 central bank speeches to identify the subset that are climate-related. Using these climate-related speeches, we analyze not only whatcentral bank actually talk about whenthey talk about climate change, but alsohow they talk aboutthetopic. Wepayparticularattentiontomacroprudentialtoolsinthecontextofthefinancialstability implicationsofclimatechange,butanumberofothertopicsarecoveredaswell. Ouranalysiscanhelppromotemoreeffectivecommunicationaboutrisksassociatedwithclimatechangeandiscriticalforclarifying theroleofcentralbanksinaddressingtheissue. There are three main conclusions. First, by all accounts central bank communication about climate change has expanded sharply in recent years. This is true at the extensive margin (the number of speeches focusing on climate change) as well as the intensive margin (the extent to which climate-related topics are coveredwithinagivenspeech). Thesharpgrowthinclimate-relatedcommunicationappliestocentralbanks 1See,forexample,Blinder,etal. (2008)foranearlysurveyoftheliterature. Morerecently,anumberofpapershavestudied theeffectofcentralbankcommunicationonmonetarypolicytransmission,includingGuthrieandWright(2000),Gürkaynak,etal. (2005),NakamuraandSteinson(2018). 2

inadvancedandemergingmarketeconomiesalike,pointingtotheglobalnatureoftheproblem. Second, the breadth of climate-related topics addressed by central banks is wide, ranging from the impactofclimatechangeontheeconomytofinancialinnovation,sustainablefinance,monetarypolicy,andthe central bank mandate among other topics. Financial stability concerns are touched upon, but the speeches rarelydiscussmacroprudentialpolicy. Atthispoint,publiccommunicationregardingdirectcentralbankactiontoaddressclimate-relatedfinancialstabilityriskscentersonidentifyingandmonitoringpotentialrisks notonlyfrommacroprudentialperspectivebutalsoformicroprudentialsupervision. Climatescenarioanalysisorstresstestingisonesuchtooldiscussedinthiscontext,butcentralbanksalsotouchonotheractions includingsettingsupervisoryexpectationsonfinancialinstitutions’climateriskmanagementandenforcing mandatory disclosure of exposure to climate-related risks. A handful of central banks discuss actions relatedtotheirowncarbonfootprint—eitherthecompositionofthebalancesheetorthroughresponsibilities associatedwithmanagementofdomesticpensionfunds—–butnotinthecontextofmacroprudentialpolicy. Finally, a third conclusion relates to the nature of the language used by central banks to communicate about climate. We present statistical evidence indicating that central banks tend to use “speculative language”,orlanguagethatindicatesuncertainty,morefrequentlyinclimate-relatedspeechesrelativetoother speeches. Given the complexity of the topic and the fact that climate change is not an area of traditional expertiseforcentralbanks,somedegreeofambiguityaboutthetopicinpubliccommunicationseemsappropriate. Nevertheless, previous studies of central bank communication have shown that clarity of language influences the credibility of the message.2 Going forward, as central banks deepen their knowledge about climate-related risks, the language used in public communication should evolve with the goal of providingsharpercommunicationwithmorepreciselanguage. Developingclimate-relatedcommunicationinthis directioncanhelpestablishcredibilitythatcentralbanksarebeingresponsivetoclimatechange. Thispapercontributestoagrowingliteraturethatconsiderstherolethatcentralbanksinrespondingto climate change. McKibbin, et al. (2021) argues that climate change considerations should be incorporated intomonetarypolicydecisions. Campiglio,etal. (2018)makesthecasethatthefinancialstabilityimplicationsofclimatechangeandthetransitiontoalow-carboneconomywarrantconsiderationinsettingfinancial regulationandcentralbankpolicies,moregenerally. Chenet,etal. (2021)goesonestepfurther. Buildingon Weitzman(2009),theseauthorsarguethattheseverityoffutureshocksduetoclimatechangearesopotentiallydevastatingthatstrongexantemacroprudentialpolicywhichprecludesinvestmentincarbon-intensive assetsiswarranted. Incontrast, Hansen(2022)contendsthatinordertomaintaincredibility, centralbanks needtobeclearabouttheirambitionsandexecutionandcanbestdosobyadheringnarrowlytotheirmandated roles.3 The author makes the case that because most effective climate policy tools are in the fiscal, rather then the monetary, toolkit. As such, if central banks pursue policies to combat climate change they riskunderminingtheirowncredibility. Finally,Bolton,etal. (2020)arguesthatcentralbanksneedtowork proactively with other policymakers and institutions to better understand and address the increasing uncer- 2InadditiontoHaldaneandMcMahon(2018),KryvtsovandPetersen(2021)andBholat,etal. (2019)discussedabove,see alsoMontesandNicolay(2017). 3Dikau and Volz (2021) highlight significant differences in how climate objectives do and do not fit within central banks mandatesacrossdifferentcountries. Forsomecountriesactivelyengaginginclimate-relatedpoliciesiswellwithinthemandate whileforothersthecaseisnotsoclear. 3

tainlyassociatedwithclimatechangetoensurelongrunfinancialandpricestability. Thispapercontributes tothisliteraturebyfocusingonhowcentralbankscommunicateaboutclimatechangeregardlessofwhether or not they are actively considering policies to address it. In keeping with the conclusions of both Hansen (2022)andBolton,etal. (2020),havingabetterunderstandingofhowcentralbanksapproachcommunication about climate will help to improve the quality of that communication and, in doing so, makes central banksengagementmorecredibleandtransparent. Thispaperalsorelatestoabodyofliterature,surveyedinBholat,etal. (2015),thatusesnaturallanguage processingtounderstandcentralbankcommunication. Anumberofstudiesfocusonextractinginformation fromspeeches,publications,orpolicystatementstoseeifthereareimplicationsforfinancialmarketactivity. Forexample,Born,etal. (2014)andCorrea,etal. (2021)bothanalyzehowinformationconveyedthrough Financial Stability Reports affects financial markets. Shapiro and Wilson (2021) use sentiment analysis to estimatecentralbankpreferencesusingFOMCstatements. Kawamura,etal. (2019)findthatambiguityin centralbankcommunicationisrelatedtofinancialmarketvolatility. Thispaperdiffersfromthesecontributions in that we study a specific topic (climate change) and the use of language in communication focused narrowlyonthattopic. Withinthetextanalysisliterature,ourpaperisanapplicationoftopicdetectiontechniques. There are a handful of studies that also apply topic modeling to central bank communication, but topic modeling is used in these papers either for dimension reduction (Hansen, et al., 2018, Hansen, et al., 2019, and Hansson, 2021)or foridentifyingthe mainormost frequenttopicsaddressed incommunication (forexample,MoschellaandPinto,2018,Armelius,etal.,2020,andBenchimol,etal,2021). Wecontribute to this literature by developing a methodology to identify central banks communications about a narrowly defined topic. Although we are particularly interested in climate change, the methodology developed here couldbeappliedtomorebroadly. The remainder of this paper is organized as follows. The next section offers details on the full set of speeches used in our analysis. In Section 3 we describe a novel methodology for identifying climaterelated speeches. Section 4 analyzes these climate-related speeches with particular attention on the role of macroprudential policy in addressing climate-related financial stability risks. Section 5 analyzes some qualitativeaspectsofcentralbankcommunicationaboutclimatechange. Finally,Section6concludes 2 Central Bank Speech Dataset OurstartingpointisthesetofallcentralbankspeechesavailablefromtheBankofInternationalSettlements (BIS)overthetwenty-fiveyearperiodfromJanuary1997toDecember2021.4 Table 1 presents some summary statistics. Since 1997, a total of 17,405 speeches were delivered by representatives from 108 different central banks. Roughly 70 percent were given by central banks in advanced economies with an average of about 17 speeches per advanced economy central bank in any given year.5 That said, there is considerable dispersion in frequency. For example, larger central banks such as 4TheBIScollectstranscriptsofspeechesinEnglishthatareavailablefromcentralbanks’websites. Forsomecentralbanks, thedatabaseonlycoversasubsetofthetotalnumberofspeechesavailableontheirwebsitesduetoselectedreportingfromtheir mediaoffices.Thedatabaseisupdatedonadailybasisandisavailableathttps://www.bis.org/cbspeeches/. 5Our split between advanced and emerging market economies is informed by IMF classification for their World Economic 4

the European Central Bank (ECB) and the Federal Reserve System (Fed) tend to be more active, giving an average of close to 90 speeches per year, whereas smaller advanced economy central banks such as the National Bank of Slovakia, the Bank of Slovenia, and the Central Bank of Cyprus communicate less frequently.6 Alltold,thefivemostactivecentralbanks—theECB(2,337speechestotal),theFed(2,073),the Deutsche Bundesbank (753), the Bank of Japan (711), and the Bank of England (708)—account for over halfthespeechesgivenbyadvancedeconomycentralbanksinourdataset. Centralbanksinemergingmarketeconomiesaccounttheremaining30percentofallspeeches,withthe average emerging market economy central bank giving about 8 per year. The most active are the Reserve BankofIndia(837speechestotal),theCentralBankofMalaysia(485),theCentralBankofthePhilippines (484), the South African Reserve Bank (382), the Bank of Albania (281), and the Bank of Thailand (219). Together,thesesixcentralbanksaccountforhalfofallspeechesgivenbyemergingmarketcentralbanks. Table1: Summarystatistics,Allspeeches Avg.#of #of #of Avg.#ofSpeeches Words CentralBanks Speeches AnnuallyperBank perSpeech AdvancedEconomy 37 12,016 16.8 1,037 EmergingMarketEconomy 71 5,389 7.8 817 Total 108 17,405 12.4 969 Source:Author’scalculations. Communicationisskewedtowardadvancedeconomycentralbanksnotonlyinthenumberofspeeches (withovertwiceasmanyspeechesperbankannually),butalsointhelengthofthespeech. Thelastcolumn ofTable1showsthatadvancedeconomycentralbankspeechesaverageabout1,000wordswhilespeeches byemergingmarketcentralbanksaverage20%fewerwordsatjustunder800.7 Since2005,Figure1showsthenumberofadvancedeconomycentralbanksthathavegivenatleastone speechinagivenyearaveragesabout30(redlineinthetopleftpanel). Foremergingmarkets(greenline) the number is slightly lower despite the fact that there are nearly twice as many emerging market central banks (71 as opposed to 37 in the advanced economies). The top right panel shows the total number of speeches for a given year. For emerging markets this number has stayed roughly constant since 2005, but it has steadily increased in the advanced economies. This increase reflects a near doubling of the average number of speeches per advanced economy central bank from about 10 to over 20 per year (bottom left Outlook(seehttps://www.imf.org/external/pubs/ft/weo/2021/01/weodata/groups.htm). Acompletelistofcentralbankswithat leastonespeechinthedatasetisgiveninAppendixA. 6Smaller central banks (advanced and emerging market economies alike) may be less likely to report speeches to the BIS (translatingthespeechtoEnglishmaybecostly,smallcentralbanksmayhavefewerrepresentativesavailabletogivespeechesor maysimplychoosetomakefewerspeechesmadeavailableontheirwebsites, etc). Accordingly, somecross-bankcomparisons needtobeinterpretedwithcaution. 7The number of words are counted after the raw text is pre-processed. We follow standard practice in the text analysis literaturebydroppingextremelycommonwords(so-calledstopwords)andcreatingnoun-phrasesbyPart-of-Speech(POS)tagging techniques.SeeAppendixB.1fordetails. 5

panel) and is in keeping with a ongoing trend toward greater central bank transparency. Finally, the lower right panel shows speech length has stayed roughly constant at about 1,000 words for advanced economy central banks, while speeches given by central banks in emerging market economies have gotten shorter overtime. Figure1: Centralbankspeechesovertime,1997-2021 Source:Author’scalculations. 3 Identifying Climate-related Speeches Using this corpus of speeches, we identify those that are “climate-related”. A “climate-related speech” is a speech given by a representative of a central bank that discusses the impact of climate change on the economy and/or the financial system. Ideally, we would identify these speeches using off-the-shelf toolsfromthetextanalysisliteraturethatfocuson“classification”or“topicidentification”but,forreasons discussedinthenextsubsection,thisisnotpractical. So,wedevelopourownmethodology. 3.1 ExistingApproachesfor“TopicIdentification" Therearethreeapproachescommonlyusedinthetextanalysisliteraturetoclassifytextonatopicalbasis.8 The first is to use what is known as an unsupervised topic model. This approach is most useful when the topics contained in the corpus are largely unknown to the researcher. In this case, unsupervised topic models are effective at reducing a large quantity of text down to a more manageable set of categories in a 8GrimmerandStewart(2013)andGentzkow,KellyandTaddy(2019)providesurveysofthisliterature. 6

process known as dimension reduction. One of the most widely used topic models is the Latent Dirichlet Allocation (LDA), introduced by Blei, Ng, and Jordan (2003). Owing to its simple structure, LDA can be easilyappliedasadimensionreductiontoolandhasbeenusedwithincreasingfrequencyintheeconomics literature.9 However, LDA (and unsupervised topic models, more generally) does not easily lend itself to our task because the objective here is to identify text related to a very specific, and relatively new, topic. Unsupervisedmethodsoffernoguaranteethatclimatechangewillbeoneofthetopicalcategoriesidentified through the dimension reduction. In fact, given that climate change is a topic that central banks have only recently started to address, it is unlikely that unsupervised methods would be effective at being able to identifyclimate-relatedspeeches.10 Analternativeapproachistousesupervisedmachineleaningtechniquessuchastextregression,including linear and non-linear techniques, or the Naive Bayes Classifier.11 In contrast with unsupervised topic models,thisapproachisappropriatewhentheresearcherknowsthetopicsofinterestaheadoftimeandhas asmallsampleoftextsthataremeaningfullyrelatedtothetopicsofinterestreadilyavailable.12 Inthiscase, thesampletextscanbeusedtotraincomputationalmodelstoefficientlyidentifytheseknowntopicsinthe largercorpus. Forourpurposes,whilewedoknowthetopicofinterestaheadoftime(climatechange),we do not have an initial set of speeches that have already been identified as definitively “climate-related” to serveasatrainingset. Thethirdapproachisknownasadictionaryapproachwhichinvolvesusingapre-establisheddictionary, or a set of keywords specified by the researcher, to classify texts into known categories. The dictionary approachismostappropriatewhenthereisastrongandreliablepriorbeliefthatacertaintopicispresentin thetextbutinformationtoidentifythetopicislimited. Examplesincludecasesinwhichthetopicofinterest doesnotmatchwiththefactorstructureofunsupervisedmethodsorwhenthereisnotrainingdataavailable to fit a supervised model. In principle, our interest in identifying climate-related speeches fits neatly into theseexamples,sothedictionaryapproachseemspromising. However, the challenge we face is that it is not clear what keywords are best to use to construct the dictionary necessary to identify climate-related speeches. Existing studies in economics and finance that use the dictionary approach do not face this problem. For example, studies on the impact of sentiment on economic and financial outcomes benefit from pre-determined dictionaries that are widely available to classify text based on sentiment.13 Some studies that branch outside of sentiment analysis do create custom dictionaries. A well-known example is the Economic Policy Uncertainty Index constructed by Baker, Bloom, and Davis (2016).14 Another example is Correa, et al. (2021), which creates a dictionary tailored 9ExamplesincludeArmelius,etal.(2020),Benchimol,etal.(2021),HansenandMcMahon(2016),Hansen,etal.(2018),and Hansenetal.(2019),andLarsen,etal.(2021). 10There are a variety of extensions that relax the strong assumptions of LDA, such as the dynamic topic model (Blei and Lafferty,2006),thecorrelatedtopicmodel(BleiandLafferty,2005),orthestructuraltopicmodel(Roberts,etal.,2013). While thesemethodsmayhelpwithourspecificapplication,theyallarestilllimitedbythefactthattheyareunsupervised.Unsupervised methodsoffernoguaranteethatoneoftheidentifiedtopicswillbeclimate-related. Someformofsupervisionseemsnecessaryto achievethisobjective. 11LineartextregressionmethodsincludepenalizedlinearmodelssuchasLasso,Ridge,andElasticNetregressions.Non-linear methodsincludegeneralizedlinearmodels(GLM),supportvectormachine(SVM),regressionstrees,anddeeplearning. 12GentzkowandShapiro(2010)andGentzkow,Shapiro,andTaddy(2019)areexamplesofpapersthatusethisapproach. 13See,Tetlock(2007),LoughranandMcDonald(2011),Bollen,etal.(2011),andWisniewskiandLambe(2013)amongothers. 14Thisindexisconstructedbycountingnewsarticlesthatcontainthefollowingtriple: “economic”or“economy”;“uncertain” 7

to a financial stability context. As noted in Gentzkow, Kelly, and Taddy (2019), creating a dictionary from scratch typically involves selecting keywords on an ad hoc basis. In the case of climate change, using an ad hoc approach could lead to mis-identification, so manual validation is required to ensure the keywords actuallydeliverthetypesofspeechesinwhichweareinterested. Thisisprohibitivelytimeconsuminggiven thesizeofourcorpus. 3.2 OurMethodology Our solution is to come up with our own methodology to identify climate-related speeches using an automated approach for topic detection through keywords.15 A detailed description is provided in Appendix B, but the general strategy is as follows. Starting with a single seed word, “climate change”, we use the method of Laver, et al. (2003) and Watanabe (2018) to create a score for all words (after pre-processing) inthecorpus. Thewordscoreisconstructedbycomparingthefrequencywitheachindividualwordshows up in speeches that do include the seed word relative to the frequency with which the word shows up in speechesthatdonotincludetheseedword. Oncewehaveascoreforeverywordinthecorpusweusethese wordscorestoconstructameasureofthepropensityofagivenspeech(whichisjustacollectionofallthe differentwordsthatwescore)tobe“climate-related”bytakingaweightedaverageofwordfrequencyusing the word scores as weights.16 We then select a minimum cutoff of zero and use that as an initial criteria for identifying what we call an unrefined set of climate-related speeches. As shown in Panel A of Table 2, this initial stage identifies 2,337 speeches that have a speech score greater than zero that are potentially “climate-related”. In practice, many of the speeches identified in this initial stage are indeed climate-related but some are misidentified, especiallywhenthespeechscoreisclosetothethreshold.17 Wedealwithmis-identification byintroducinganiterativerefinementstage,detailedinPanelBofthetable. or“uncertainty”;andoneormoreof“congress”,“deficit”,“FederalReserve”,“legislation”,“regulation”or“WhiteHouse”. 15Laver,etal.(2003)introducedanautomatedmethodusingascoringtechnique(calledWordscores)that,inapoliticaleconomy context,detectspoliticalactors’positionswithminimalhumanintervention.King,etal.(2017)showsthatasupervisedalgorithm canbeusedtodetectasetofkeywordsthatidentifytopicswithgreateraccuracythankeywordsgeneratedusingadhocmethods. Watanabe(2018)andWatanabeandZhou(2020)makesimilarargumentsusinganautomatedmethodologywithscoringtechniques. 16Forallspeechesinthecorpus,thescoreaverages−0.27withastandarddeviationof0.37andrangesfrom−2.16to3.70. A positivescoreindicatesaspeechthatcontainsmanywordsthatshowupfrequentlyalongsidetheseedword“climatechange”,a scoreofzeroindicatestheuseofwordsinthespeechareneutraltotheseedword,andanegativescoreindicatesthewordsusedin thespeechtendnottoalsobeusedinspeechesthatincludetheseedword. 17Thescoringmethodtreatswordssimplyasdataratherthanrequiringanyknowledgeoftheirmeaningasusedinthetext. Whilethisisusefulforautomation, somedegreeofmis-identificationisinevitable. GrimmerandStewart(2013)pointoutthat carefulpre-processingoftextswillmitigatetheissuebecausetheresultwilldependstronglyonthereferencetextsthatareused. Lowe(2008)showspotentialbiasesinthescoringmethod,althoughitcanbereasonablysmalliftheresultsseemplausible. 8

Table2: Refinementprocesstoidentifyclimate-relatedspeeches Numberof SpeechScoreStatistics KeywordIdentified speeches Mean St.Dev. Min. Max. inExplorationSet Allspeeches 17,405 -0.27 0.37 -2.16 3.70 PanelA:InitialIdentification Climate-relatedspeeches(BeforeRefinement) 2,337 0.34 0.52 0.00 3.70 PanelB:RefinementStage Highscorespeechesusedforkeywordexploration 427 1.23 0.65 0.48 3.70 Iteration1(s =[“ClimateChange”]) 1 Speechesidentifiedbys 373 1.30 0.66 0.48 3.70 1 Explorationset 54 0.77 0.38 0.48 1.92 “GreenFinance” Iteration2(s =[s |“GreenFinance”]) 2 1 SpeechesIdentifiedbys 389 1.29 0.68 0.48 3.70 2 Explorationset 38 0.67 0.30 0.48 1.92 “Climate-relatedRisk” Iteration3(s =[s |“Climate-relatedRisk”]) 3 2 SpeechesIdentifiedbys 393 1.29 0.65 0.48 3.70 3 Explorationset 34 0.62 0.21 0.48 1.55 “ParisAgreement” Iteration4(s =[s |“ParisAgreement”]) 4 3 SpeechesIdentifiedbys 395 1.28 0.65 0.48 3.70 4 Explorationset 32 0.61 0.21 0.48 1.55 “ClimatePolicy” Iteration5(s =[s |“ClimatePolicy”]) 5 4 SpeechesIdentifiedbys 396 1.28 0.65 0.48 3.70 5 Explorationset 31 0.61 0.21 0.48 1.55 “ClimateRisk” Iteration6(s =[s |“ClimateRisk”]) 6 5 SpeechesIdentifiedbys 397 1.28 0.65 0.48 3.70 6 Explorationset 30 0.61 0.21 0.48 1.55 “Low-carbonEconomy” Iteration7(s =[s |“Low-carbonEconomy”]) 7 6 SpeechesIdentifiedbys 398 1.28 0.65 0.48 3.70 7 Explorationset 29 0.61 0.22 0.48 1.55 “CarbonEmission” Iteration8(s =[s |“CarbonEmission”]) 8 7 SpeechesIdentifiedbys 399 1.28 0.65 0.48 3.70 8 Explorationset 28 0.58 0.12 0.48 1.02 “GreenBond” Iteration9(s =[s |“GreenBond”]) 9 8 SpeechesIdentifiedbys 400 1.28 0.65 0.48 3.70 9 Explorationset 27 0.56 0.09 0.48 0.87 ∅ PanelC:FinalIdentification Climate-relatedspeeches(AfterRefinement) 555 0.98 0.73 0.00 3.70 Source:Author’scalculations. Thisrefinementfocusesonthesubsetof427speechesthathaveaspeechscoregreaterthantwostandard deviations over the mean for all speeches (this is roughly equivalent to any speech with a score that falls in the top 2.5% of the distribution of all speech scores). We use this subset to identify a set of keywords related to climate change in an iterative process. The initial iteration uses a dictionary comprised of the single keyword “climate change” to identify 373 speeches (87% of the 427 total high score speeches) that 9

includethiskeyword,leaving54highscorespeechesthatcannotbeidentified. Weusetheselaterspeeches asan“explorationset”fromwhichweselectthemostimportantclimate-relatedkeyword,whereimportance is judged by frequency (i.e., the number of speeches that include the keyword) and relevance based on our judgement. Asshowninthetable,forthefirstiterationtheresultingkeywordis“GreenFinance”,whichwe then add to our dictionary, expanding it to two keywords, and we repeat the process. This continues until we can no longer find additional keywords that are useful. At that point the automated dictionary, which consistsofthekeywords: "ClimateChange","GreenFinance","Climate-relatedRisk","ParisAgreement", "ClimatePolicy","ClimateRisk","Low-carbonEconomy","CarbonEmission","GreenBond",iscomplete. ThefinaltworowsofPanelBofTable2showthatadictionarycomprisedoftheseninekeywordsidentifies 400(93%)ofthe427highscorespeechesleaving27highscorespeechesunidentified. The final step is to use the automated dictionary to refine our initial set of speeches. We do this using a rule such that a speech is included in the final set of climate-related speeches if it contains any keyword fromthedictionaryinatleasttwoseparatesentenceswithinthespeech. ThelastrowofTable2showsthis refinementreducesthe2,337initiallyidentifiedspeechesdownto555speechesthatconstituteourfinalset ofclimate-relatedspeeches. 3.3 ValidityChecks AsdiscussedinGrimmerandStewart(2013)andGentzkow,KellyandTaddy(2019),ourapproachrequires acarefulvaliditychecktoensurethatthespeechesidentifiedareindeedmeaningfullyclimate-related. Our validity checks take two forms. We start by examining speech score summary statistics and word clouds. Comparingpre-refinementspeeches(PanelAofTable2)topost-refinementspeeches(PanelC),the range of scores remains unchanged but the average in the post-refinement set is higher (0.98 compared to 0.34). This suggests the refinement is effective at separating out low score speeches that are meaningfully relatedtoclimatechangefromthosethatarenot. TheinformationinPanelBofTable2offersdetailsonhowtherefinementachievesthisseparation. For example,thetableshowsthatinthefirstiterationthesinglekeyword“climatechange“issufficienttoidentify 373 of the 427 highest score speeches. Of the 54 that remain unidentified, the average score of 0.77 is relativelylowbutthemaximumremainsquitehighindicatingthatthereremainsomemeaningfullyclimaterelatedspeecheswithinthissubset. ThewordcloudintheupperleftofFigure2confirmsthatclimate-related topicsdoindeedfeatureprominentlyinthissubsetofunidentifiedspeeches.18 Ourmethodologyaddresses this by adding an additional keyword chosen from this set (what we call the keyword exploration set) and then rerunning the identification based on this expanded dictionary. We continue doing this in subsequent iterations until we can no longer identify a useful keyword to add to the dictionary. Panel B of Table 2 shows that both the average and the maximum score for the keyword identification set are decreasing at 18Inwordclouds,sizerepresentsafrequencyofthewordinasetofspeeches. Ratherthanasimplecountofthenumberof timesthatthewordappearsinatext,textanalysispracticessuggestthatputtingmoreweightonmeaningfulwordsmakesthem moreusefultointerpret. Followingsuchpractices,Figure2usesadocumentfrequency(i.e.,thenumberofspeechesthatinclude aword)thatismultipliedbythecorrespondingwordscoretoputmoreweightsonimportantwordsfordetectingclimate-related speeches. Figure 3 and thereafter use TF-IDF (term frequency-inverse document frequency), which is widely used in practice. TD-IDFadjustsasimplefrequencybyputtinglessweightsonwordsthatcommonlyappearacrosstheentiresetofspeeches. 10

each iteration as it becomes increasingly difficult to identify additional keywords. As can be seen in the bottom right panel of Figure 2 the most frequent words used in the keyword identification set at the final iterationhavelittletodowithclimatechange. Figure2: Wordcloudsofspeechesusedforkeywordexploration Source:Author’scalculations. After applying our completed dictionary to the set of pre-refinement speeches, the post-refinement set includes367ofthe427totalhighscorespeechesusedforthekeywordexploration(i.e.,therefinementfilters out 60 speeches as not meaningfully climate related). Of those, 27 (with an average speech score of 0.56) areremovedbecausetheydonotcontainanyofthekeywordsandanadditional33(withanaveragespeech scoreof0.66)arefilteredbecausetheydonotcontainatleastonekeywordintwoseparatesentences. The refinement also adds an additional 188 speeches that have low scores (average speech score of 0.28) but nonetheless did satisfy the dictionary-based refinement criteria. This is an indication that the refinement addsvaluetotheidentificationprocessbeyondasimplehighscorecutoff. Figure3showswordcloudsfor climate-related speeches before refinement (top left) and after refinement (top right). The most frequent words in the post-refinement set are more narrowly related to climate-relevant topics compared to the prerefinement set. The lower left panel shows a word cloud for those 1,782 speeches filtered out of the prerefinement set. These speeches tend to focus on topics like technology, financial sector innovation, and Brexit,forexample. Finally,thelowerrightpanelshowsawordcloudforallspeechesthatarenotclimaterelated. Theytendtofocusonstandardcentralbanktopicssuchasinflationandmonetarypolicy. Inadditiontospeechstatisticsandwordclouds,wealsohandvalidateasamplingofbothclimate-related andnon-climaterelatedspeeches. Handvalidationistime-consuming,soweconcentrateonfoursubsetsof speechesthatbestdefinetheidentificationboundary. Westartwiththe27speechesfilteredoutofthesetof highestscorespeechesbecausetheydonotcontainanykeywordfromthedictionary. Thesespeechesfocus ontopicslikecyberrisk,socialresponsibility,ortheinsuranceindustry(seethebottomrightpanelofFigure 11

2).19 Itiscommonforclimatechangetobementionedalongsidethesetopics,andindeedeveninthesame sentence,andthisdrivesupthespeechscore. Wethinkitiswarrantedtoremovethemfromthefinalsetof climate-relatedspeeches. Figure3: Wordcloudsforclimate-relatedspeeches,beforeandafterrefinement Source:Author’scalculations. Next,weturntothe33highscorespeechesthatwereexcludedatthefinalrefinementstep(speechesthat contain at least one keyword from the dictionary, but not in at least two sentences). These speeches cover a variety of topics that only lightly touch on climate issues.20 For example, some discuss the insurance sectorandtangentiallymentionclimateasapotentialissueofconcern. Otherspeechesfocusonsustainable finance and mention climate issues only in passing. Under a different identification scheme some of these speeches might have been included in the final set, but in general they tend to treat climate-related topics onlylightly,soitseemsreasonabletoexcludethem. Focusingontheboundarybetweenclimate-relatedandnon-climate-relatedspeeches,thethreeclimaterelatedspeecheswiththelowestspeechscores(i.e.,thelastthreein)alldiscussclimateissuestosomeextent but in no sense is it the main theme of the speech. For example, Nabiullina (2021) “Speech at Association of Russian Banks” (web link), which scored a 0.0002, touches on climate change only to the extent that it isanissuethatbankingsystemfaces. Olsen(2021)“Monetarypolicystrategy-frommandatetodecisions” (web link) scored a 0.02 and only briefly touches on climate in the context of the central bank mandate. Finally,Gang(2021)“HongKong’spositioningandprospectasaninternationalfinancialcentre”(weblink) scoreda0.03anddiscussesgreenfinanceasaparticularstrengthofHongKong’sfinancialmarkets. Onthe 19Examples include: Hendar (2015) “Increasing cooperation between Indonesia’s central bank and state police“ (web link); Lautenschlager(2017)“Cyberresilience-abankingsupervisor’sview”,(weblink);andWaas(2014)“Handlingallegedpayment systemandcurrencyexchangecrime.”(weblink). 20Examplesinclude:Diokno(2021)“Sustainabilityininvesting“(weblink);Elderson(2018)“Letthefutureoffinancebethat offinancingthefuture”(weblink); Lane(2016)“Dualperspectivesontheinsurancesector–consumerprotectionandfinancial stability”(weblink). 12

othersideofthecutoff,allofthehighestscoringnon-climate-relatedspeecheswereallexcludedduringthe speechrefinementstage,sotheyhavebeenalreadyreviewedabove. Finally,weexamined11speechesthatdonotsatisfytheinitialcriteriaofhavingaspeechscoregreater than zero, but do satisfy the refinement criteria of having a keyword in at least two sentences. These negativescorespeechestendtofocusonmonetarypolicyandfinancialmarketdevelopments.21 Climate-related wordsareusedlargelyinthecontextofconcernsaboutcommodityprices, forexample, orthroughdiscussionofhowenhancementofgreenfinancemightbefavorabletofinancialmarketdevelopment. Inprinciple, thesetypesofspeechesarecandidatesforinclusion. However, afterreviewingeachofthem, ourjudgment is that the treatment of climate change is only cursory (which is what leads to the negative score) and for thatreasonwechoosenottoincludetheminourfinalset. Some caveats are worth noting. First, selection of the seed word itself, as well as the speech score thresholds, will affect the final results. For example, changing the seed word to something other than “climatechange”affectsthedistributionofspeechscoresandthismattersforthefinalsetofspeeches. Second, we chose a threshold of 0 for identifying pre-refinement speeches. Raising this threshold would lead to fewer pre-refinement speeches and that would also affect the final set. By the same token, lowering the threshold should lead to more pre-refinement speeches, but this is less worrisome because the refinement stage would likely filter many of these additional speeches out. The third caveat is that we focused the keywordidentificationsetonthepre-refinementspeechesthathadaspeechscoresgreaterthantwostandard deviations over the mean. Raising (or lowering) this threshold will decrease (or increase) the size of the keyword exploration set and that, in turn, matters for how the dictionary is constructed. A fourth caveat is thattheselectionofkeywords,thoughguidedbysomecomputationaltools,isultimatelybasedonjudgment. Differentkeywordsmightleadtoadifferentsetofidentifiedspeeches. Giventhesecaveats,somedegreeof handvalidationseemsinevitable.22 Finally, there are instances where the identification methodology is not perfect. For example, Haldane (2010)“The$100billionquestion”(weblink)isidentifiedbyourmethodologyasaclimate-relatedspeech but a close read reveals that climate change is only referred to in the speech as an example of a negative externality, which is then used to motivate why central banks should take action against systemic financial risk. The speech is not about central banking and climate change per se. This sort of mis-identification seems unavoidable, although it is comforting that the score assigned to this speech was very low at 0.081 (amongthebottomtwentyspeechesincluded). 3.4 SummaryStatistics Summary statistics are presented in Table 3. Roughly one half of all central banks (50 out of 108 total in the sample, or 46.2%) have given at least one speech on climate change amounting to just over 3% of 21Examplesinclude: Trichet(2008)“Monetarypolicyinchallengingtimes”(weblink);Yue(2021)“HongKong’spositioning andprospectasaninternationalfinancialcentre”(weblink);Macklem(2021)“Thelongandshortofit-abalancedvisionforthe internationalmonetaryandfinancialsystem”(weblink) 22Weexploredthesensitivityofourfinalsetofclimate-relatedspeechestotheseparameters. Themostsensitivearetheinitial seedwordandthethresholdusedforidentifyingthepre-refinementspeechset. Incontrast,thesetofhighscorespeechesusedfor therefinementstageisrobustacrossthechoiceofparametersand,asaresult,theidentifiedkeywordsarerelativelyinsensitive. 13

all speeches in the dataset. The average speech score is 0.98. Central banks in advanced economies are more active compared to emerging markets. Nearly 80% of all advanced economy central banks (29 out of37total)havegivenatleastoneclimate-relatedspeechwhichaccountfor3.8%ofallspeechesgivenby advancedeconomycentralbanks. Incontrast,onlyabout30%ofemergingmarketcentralbanks(21outof 71total)havegivenatleastoneclimate-relatedspeechandamountingtoaround1.5%ofallspeechesgiven. Theaveragespeechscoreofadvancedeconomyspeechesis0.99and,whilelowerforemergingmarketsat 0.90,thedifferenceisnotstatisticallysignificant. Table3: Summarystatistics,Climate-relatedspeeches #of Percentof #ofClimate- Percentof Average CentralBanks CentralBanks relatedSpeeches TotalSpeeches SpeechScore Total 50 46.3% 555 3.2% 0.98 AdvancedEconomy 29 78.4% 462 3.8% 0.99 EmergingMarketEconomy 21 29.6% 93 1.7% 0.90 Source:Author’scalculations. The cumulative number of climate-related speeches given by each central bank over the entire sample period,conditionalongivingatleastfiveclimate-relatedspeeches,isshowninthetoptwopanelsofFigure 4. ThetopleftpanelshowsthefivemostactiveadvancedeconomycentralbanksaretheECB,theBankof England, the Deutsche Bundesbank, the Bank of France, and Bank of Italy, each of which have given over thirtyspeechesrelatedtoclimatechange.23 The top right panel shows the most active emerging market central banks are the central banks of Malaysia, the Philippines, China, India, and Fiji. The bottom two panels show the same information for the average speech scores. Amongst advanced economies, the Federal Reserve, the Netherlands Bank, and the Bank of Greece have the three highest average speech scores despite having given a relatively low number of climate-related speeches. A similar dynamic exists amongst the emerging market economies as the Central Bank of Kenya and the Central Bank of the Phillipines rank highly on average scores across a relativelylownumberoftotalspeeches. 23ItisworthnotingthattheECBidentifies46climate-relatedspeechesonitswebsite(weblink)andtheBankofJapanidentifies 10(weblink). Forcomparison,ourmethodologyidentifiesawidersetofclimate-relatedspeechesforbothcentralbanks,with72 identifiedfortheECBand18identifiedfortheBankofJapan. TakingintoaccountthattheECBandtheBankofJapanbothlist oneclimate-relatedspeechontheirrespectivewebsitesthatarenotreportedtotheBIS,andthereforedonotshowupinourdataset, weidentify90percentofthespeechesreportedbytheECB(40of45speeches)and100percentofthespeechesreportedbythe BankofJapan(9of9). 14

Figure4: Totalnumberofclimate-relatedspeechesandaveragespeechscorebycentralbank,1997-2021 Notes:Centralbankswithlessthanfiveclimate-relatedspeechesaredroppedfromthisfigure. Source:Author’scalculations. Figure5showshowcommunicationaboutclimatechangehasevolvedovertime. Byallaccounts,ithas expandedsharplyoverthepastfiveyears. Morecentralbanksarespeakingaboutclimateandtheyaredoing it with increasing frequency. The top left panel shows the number of central banks giving climate-related speeches has increased with a notable jump following the 2015 Paris Agreement.24 The top right panel shows an even sharper increase in the number of speeches given each year. Over the past five years, the shareofclimate-relatedspeeches(lowerleftpanel)hasshotup,risingfromunder5percentin2017to20% of speeches given by emerging market central banks andmore than 30% for advanced economies. Finally, thebottomrightpanelshowsthetheaveragespeechscore,whichisaproxyfortheextenttowhichclimaterelatedtopicsarecoveredwithinagivenspeech(i.e.,theintensivemargin),hasalsoincreasedsteadily.25 24InApril2015, G20FinanceMinistersandCentralBankGovernorsaskedtheFinancialStabilityBoard(FSB)“toconvene public- and private- sector participants to review how the financial sector can take account of climate-related issues” (Financial StabilityBoard2015). 25Thespikeintheaveragespeechscoreforadvancedeconomycentralbanksin2011isentirelydrivenbyasinglespeechwith averyhighscoreof3.13. 15

Figure5: Climate-relatedspeechesovertime,1997-2021 Source:Author’scalculations. Identifyingthemostrelevantspeechesineachyearoverthepasttwentyyearsoffersanotherperspective on the evolution of climate-related speeches over time. Table 4 shows “Environmental issues and their implicationsforfinancialinstitutionsinHongKong”,givenbytheHongKongMonetaryAuthority(HKMA) wasnotonlythemostrelavantclimate-relatedspeechgivenin2000,butalsothefirstclimate-relatedspeech identifiedoverallinourset. Followingthisearlycontribution,therewasadearthofclimate-relatedspeeches until the late 2000s, at which point the issue seems to have resurfaced, perhaps owing to the Nobel Peace PrizebeingawardedtotheIntergovernmentalPanelonClimateChangeandAlGoreJrin2007. From2007 to 2014 the speech score for the most relevant climate-related speeches averaged just under 1.20, but this averageissupportedbytwoveryhighscoringspeechesgivenbyrepresentativesofBankIndonesiaandthe Bank of Greece in 2008 and 2011, respectively. Removing these two speeches pushes the average score down to a more modest 0.61, suggesting that while central banks were talking about climate in this early period they were quite limited in what they had to say. This changed notably following Mark Carney’s 2015speechon“BreakingtheTragedyoftheHorizon: ClimateChangeandFinancialStability”. Speeches that followed began to address climate topics more extensively with the average speech score for the most relevantclimate-relatedspeechinagivenyearrisingto2.81overtheperiod2015to2021. 16

Table4: Mostrelevantclimate-relatedspeechesbyyear,1997-2021 Total#of Climate Climate- Year CentralBank SpeakerandTitle Date Speech related Score Speeches 2000 HongKongMonetaryAuthority DavidCarse:Environmentalissuesandtheirimplicationsforfinancial 2000-11-29 0.52 1 institutionsinHongKong 2007 MonetaryAuthorityofSingapore GohChokTong:StayingaheadoftheAsiancurve 2007-11-01 0.15 1 2008 BankIndonesia Boediono: Macroeconomicimpactofclimatechange-opportunities 2008-08-02 2.76 11 andchallenges 2009 BankofItaly MarioDraghi:Thefinancialcrisis-impactandresponses 2009-04-26 0.76 2 2010 BankofItaly AnnaMariaTarantola:Womennurturingsustainabledevelopment 2010-10-22 1.05 4 2011 BankofGreece GeorgeAProvopoulos:TheimpactofclimatechangeinGreece 2011-06-01 3.13 1 2012 CentralBankofMalaysia MuhammadbinIbrahim:Roleandopportunitiesofthefinancialsys- 2012-10-02 0.68 2 teminsupportinggreentechnology 2013 CentralBankofTrinidadandTobago Jwala Rambarran: Generating more inclusive economic growth 2013-06-05 0.48 5 throughscienceandtechnology 2014 ReserveBankofIndia GPadmanabhan:Corporatesustainabilityapanaceaforgrowth-val- 2014-10-17 0.54 1 ues,convictionsandactions 2015 BankofEngland MarkCarney: Breakingthetragedyofthehorizon-climatechange 2015-09-29 2.25 10 andfinancialstability 2016 BankofEngland MarkCarney:RemarksonthelaunchoftheRecommendationsofthe 2016-12-14 2.26 6 TaskForceonClimate-relatedFinancialDisclosures 2017 BankofGreece YannisStournaras:Climatechange-challenges,risksandopportuni- 2017-06-30 3.70 15 ties 2018 BankofEngland MarkCarney:Atransitioninthinkingandaction 2018-04-06 2.91 31 2019 CentralBankofKenya Patrick Njoroge: The importance of green finance guidelines as 2019-02-20 2.84 134 Nairobiseekstobecomeaglobalhub 2020 ReserveBankofNewZealand AdrianOrr: Progressingclimateactionbydrivingtransformational 2020-10-28 2.91 108 change 2021 BankofJapan HaruhikoKuroda:Addressingclimate-relatedfinancialrisks–froma 2021-03-25 2.77 223 centralbank’sperspective Source:Author’scalculations. Alltold,centralbankcommunicationaboutclimatechangehasexpandedrapidly,bothattheextensive margin(ascapturedbythenumberofspeechesfocusingonclimatechangeandthenumberofcentralbanks givingthesespeeches)andattheintensivemargin(ascapturedbytheaveragespeechscore). 4 What do Central Banks Talk About with Regard to Climate? Wesummarizesometopicsthatcentralbanksaddressintheirspeechesaboutclimatechange,startingwith financialstabilityandmacroprudentialpolicybeforeturningtoabroadersetoftopics. 4.1 ClimateChange,FinancialStability,andMacroprudentialPolicy We are particularly interested in how central banks discuss the use of macroprudential tools to address financial stability risks associated with climate change. To focus on the overlap between these three topics weappliedthesamemethodologyasdescribedinSection3.2toidentifyfinancialstability-relatedspeeches andspeechesrelatedtomacroprudentialpolicy. 17

Startingwiththeseedword,“financialstability”,wecreateascoreforeveryspeechanduseittoidentify anunrefinedsetoffinancialstability-relatedspeeches. Thissetisthenrefinedusinghighscorespeechesto createa“financialstability”dictionary,whichconsistsofthefollowingkeywords: “financialstability”;“systemicrisk”;“macroprudentialpolicy”;“resolutionregime”;“macroprudentialinstrument”;“countercyclical capital buffer”; and “macroprudential perspective”. We apply this dictionary using the same identification rule from Section 3.2 (at least one keyword in two separate sentences within the speech) to identify 3,566 speeches (just over 20% of all speeches) that focus on the topic of financial stability. Similarly, we use the seed word “macroprudential policy” to create a dictionary consisting of the keywords: “macroprudential policy”; “systemic risk”; “macroprudential regulation”; “macroprudential instrument”; “financial vulnerability”; “countercyclical capital buffer”; “macroprudential measure”; “macroprudential supervision”; and “macroprudential analysis” and use it to identify 1,050 speeches (6% of all speeches) focusing on macroprudentialpolicy.26 After classifying speeches into those that are climate-related, financial stability-related, and macroprudentialpolicy-related,weassesstheextentofoverlapbylookingforspeechesattheintersectionofthethree topics. Figure6showsthatofour555totalclimate-relatedspeeches,alittleunderone-third(168,or30.0%) also touch on financial stability, but only a small handful (37, or 6.7%) address macroprudential policy. In comparison,overonequarterofallfinancial-stabilityrelatedspeeches(986,or27.7%)alsotouchonmacroprudentialpolicyissues. Thatsaid,muchofthefocusintheremainderofthissectionisonthe36speeches attheintersectionofallthreetopics. Figure6: Climate-related,financialstability-related,andmacroprudentialpolicy-relatedspeeches Source:Author’scalculations. 26The“financialstability”and“macroprudentialpolicy”dictionariessharesomeofthesamekeywords,sooverlapbetweenthe twosetsofspeechesistobeexpected.Thatsaid,thespeechscores,whichareusedtocreatetheunrefinedspeechsetstowhichthe dictionariesareapplied,arebasedondifferentseedwords.Asaresult,anyoverlapintroducedfromcommonkeywordswouldonly beintroducedintherefinementstage. 18

To better understand this subset of speeches Table 5 shows average speech scores across the three topics for different speech subsets. The top panel compares climate-related speeches to non-climate-related speeches. Asexpected,theaverageclimate-relatedspeechscoreismuchhigherforclimate-relatedspeeches relative to non-climate related speeches and the difference is statistically significant. It is also true that climate-related speeches have higher financial stability speech scores and higher macroprudential speech scores, althoughthedifferenceissmallerinmagnitude(albeitstillstatisticallysignificant). Inotherwords, the typical climate-related speech is more likely to address both financial stability- and macroprudentialrelatedtopicsrelativetothetypicalnon-climate-relatedspeech. Thebottompanelfocusesonthesubsetof speeches that address both financial stability and macroprudential policy issues. Climate-related speeches thattouchonbothtopicstendtotreatthemmorelightlyrelativetospeechesthatarenotclimate-related. Table5: Speechscoresfordifferentsubsetsofclimate-andnon-climate-relatedspeeches Climate-related FS-related Macropru-related SpeechScore SpeechScore SpeechScoer Climate-relatedspeeches,(n=555) 0.98 0.15 -0.22 Non-climate-relatedspeeches,(n=16,850) -0.31 -0.03 -0.32 Difference 1.28 0.18 0.10 p-value 0.000 0.000 0.000 Climate-FS-Macropru,(n=36) 0.51 0.39 0.23 Non-climate-FS-Macropru,(n=950) -0.21 0.48 0.35 Difference 0.72 -0.09 -0.12 p-value 0.000 0.003 0.000 Source:Author’scalculations. Figure 7 compares speech scores for different subsets of speeches in a series of scatter plots. The leftpanelcomparesthe3,566financialstability-and1,050macroprudentialpolicy-relatedspeeches. Those focusingexclusivelyonfinancialstabilityorexclusivelyonmacroprudentialpolicyaredepictedbythegreen andbluedots,respectively. Speechesattheintersectionofthetwotopicsaredepictedbythereddots. The relationshipbetweenthefinancialstabilityandmacroprudentialspeechscoresinallthreesubsetsispositive (the green, blue, and red dots are all upward sloping), suggesting that the two topics are often discussed closelytogether. Themiddlepanelshowscomparesour555climate-relatedspeechestofinancialstabilityrelatedspeeches,themajorityofwhicharenotrelatedtoclimatechange(asdepictedbythebluedots). For thosethateitherfocusexclusivelyonclimate(thegreendots)ortouchonbothtopicsinthesamespeech(the red dots) there is a positive correlation between the speech scores suggests that financial stability is often discussed in the context of climate change. The final panel on the right compare climate-related speeches tomacroprudentialpolicy-relatedspeechesandindoingsoshowsanimportantdichotomy. Climate-related speeches(thegreendotshugthey-axis)rarelyaddressmacroprudentialtopicsandviceversa(thebluedots hug the x-axis). In the rare case the two topics overlap (37 speeches total), the relationship between the climatespeechscoreandthemacropruendtialpolicyspeechscoreisweak. Tosummarize,financialstabilityandmacroprudentialpolicytopicsdotendtoshowupmorefrequently in climate-related speeches relative to speeches that are not related to climate change. However, when we focusonthesubsetofspeecheswherewemightexpecttofindthemostextensivediscussionofthesetopics 19

inthecontextofclimatechange,wefindthatcentralbankstendtotreatbothtopicsonlylightly.27 Figure7: Topicrelevanceinclimate-,financialstability-andmacroprudentialpolicy-relatedspeeches Source:Author’scalculations. Diggingalittledeeper,weexploreoursetofclimate-relatedspeechestoidentifyanymentionofspecific policy tools possibly associated with climate change. To do this, we collected a total of 1,498 climateor green-related keywords (i.e., “Climate XXX” and “Green XXX”) found in our set of climate-related speeches and judgmentally selected 65 relevant policy tools that show up in this set of words.28 These 65 policy tool keywords were then grouped into six different categories: climate disclosure, climate risk management,greentaxonomy,climateresearch,climatescenarioanalysis/stresstesting,andgreenmonetary policy/capitalallocation. With these dictionaries in hand, we then identify 183 speeches that address climate-related topics and alsocontainsomementionofanyofthesepolicytoolkeywords. Table6showsthedistributionacrossthesix categories. Themostfrequentreferencestopolicytoolstendtobemicroprudentialinnature. Forexample, overtwentypercentofallclimate-relatedspeeches(114outof555total)containsomekeywordassociated with climate disclosures, meaning either voluntary or mandatory reporting of climate risk exposure which mightallowcentralbanks,supervisors,andprivatemarketparticipantstobetterunderstandclimate-related 27Twoexamplesofrecentspeechesthattouchonbothtopics,butonlylightlyareLagarde(2021),“Macroprudentialpolicyin Europe–thefuturedependsonwhatwedotoday”(weblink))“Effectiveclimatepolicieswouldalsobenefitthefinancialsector. Butmacroprudentialpolicymakershaveaparttoplay,too,inidentifyingandmitigatingthefinancialaspectsofclimate-related risks.” andLane(2019)“ClimatechangeandtheIrishfinancialsystem”(weblink)“Thiscallsforongoingmonitoringofclimate risks,togetherwiththedevelopmentofclimate-drivenscenarioanalysesandstresstests.Byextension,theresultsofsuchanalyses maycallfortheappropriatemacroprudentialpoliciestomitigatetheserisks.” Thislighttreatmentstandsoutgiventhetypically tightrelationshipbetweenfinancialstabilityandmacroprudentialpolicyobservedinnon-climate-relatedspeeches. 28Mostof1,498climate-orgreen-relatedkeywords, "climatecrisis"and"greeninvestment"forexample, arenotobviously relatedtoclimatepolicytools. Inaddition,wealsoremovewords,suchas“climatepolicyinstrument”and“greenfinancepolicy” thatdonotindicatespecifictools.SeeAppendixCforalistofall65keywordsofspecificpolicytoolsandtheirclassificationsinto 6categories. 20

Table6: Keywordsforidentifyingspecificclimatepolicytoolsmentionedinclimate-relatedspeeches Category Number Average Average Average of Climate FS Macropru speeches Speech Speech Speech Score Score Score AllClimate-relatedSpeeches 555 0.98 0.14 -0.22 AnyClimatePolicyTools 183 1.48 0.27 -0.19 ClimateDisclosure 114 1.60 0.30 -0.17 ClimateRiskManagement 48 1.51 0.28 -0.25 GreenTaxonomy 26 1.58 0.20 -0.27 ClimateResearch 12 1.49 0.31 -0.13 ClimateScenarioAnalysis/StressTest 74 1.65 0.34 -0.16 GreenMonetaryPolicy/CapitalAllocation 12 1.33 0.17 -0.14 Source:Author’scalculations. financialrisks. Anadditional10%(48speeches)touchonclimateriskmanagement. Typically,thelanguage in these speeches is geared toward identifying and monitoring to get a better understanding of both macro and/or microprudential implications of climate risk.29 Green taxonomy accounts for 5% (26 speeches) of identified speeches. The discussion typically focuses on developing a common set of definitions regarding the environmental profile of assets and financial market participants and is closely associated with green/sustainable finance.30 Climate research comes up in 2% (12 speeches) of identified speeches likely reflectingthatcentralbankslacksufficientunderstandingofclimateriskandneedtobetterunderstandthose risksbeforetakingpolicyaction.31 Climatescenarioanalysis/stresstestingismentionedinroughly15%of allclimate-relatedspeeches(74speeches). Manydiscusstheusefulnessofscenarioanalysisfromamicroprudentialperspective,forexampleinthecontextofriskassessment.32 Thatsaid,ahandfulofspeechesdo touch on top down stress testing as a means of assessing the systemic implications of climate risk with an eyetowardguidingmacroprudentialpolicy.33 29Forexample,Stiroh(2020)“Climatechangeandriskmanagementinbanksupervision”(weblink),“...supervisorscanfocus on identifying and managing risks, both microprudential and macroprudential, that emerge along a transition path to a more sustainableeconomy”,andElderson(2021)“Overcomingthetragedyofthehorizon”(weblink),“Gettingbankstodevelopaction planstocomplywiththeexpectationsthattheECBsetoutinitsguideonCEriskswasthefirststepingettingthemtoassessand developtheirriskmanagementcapabilities.” 30Menon(2021)“Beingthechangewewanttosee-asustainablefuture”(weblink)“Developingacleartaxonomyfortransition activitiesisespeciallyrelevantforAsia. Asianeedstosustaineconomicandsocialdevelopmentwhileshiftingtoalowercarbon future. A taxonomy that includes both green and transition activities can support a progressive shift to greater sustainability.”; Elderson(2018)“Let’sdance”(weblink)“Weshouldremoveallunnecessaryobstaclestothistransition. Andassistthesectorin creatingcommondefinitionsandstandards.Aswellasprovidethenecessaryguidance.Thedevelopmentofasustainabletaxonomy isagoodexampleofthis.” 31Stournaras(2018)“Climatechange-threats,challenges,solutionsforGreece”(weblink)“Greece,alongwithothersmall countriesintheclimate-sensitiveMediterraneanregion,isexpectedtoincuradverseeffectsfromclimatechange. Acknowledging thisfact,theBankofGreecehasbeenoneofthefirstcentralbanksworldwidetoactivelyengageintheissueofclimatechangeand investsignificantlyinclimateresearch.” 32Breeden (2019) “Avoiding the storm - climate change and the financial system” (web link) “Measuring these future risks fromclimatechangetotheeconomyandtothefinancialsystemisacomplextask. Amyriadofpossibleclimatepathways–with different physical and transition effects – need to be translated into economic outcomes and financial risks looking ahead over manydecades.Tosimplifythatchallenge,weneedtofocusnotonwhatwillhappenbutwhatmighthappen.Todothatwecanuse scenarioanalysis–datadrivennarrativesthathelpanchorourassessmentsofrisk.” 33Forexample,deGuindos(2019)“Implicationsofthetransitiontoalow-carboneconomyfortheeuroareafinancialsystem” (weblink)“Thepilottestframeworkwillbemacroprudentialinnature, andallowustoanalysethesystem-widematerialityof 21

Speeches that address green monetary policy/capital allocation make up only 2% (12 speeches) of all climate-relatedspeeches. Thesespeechesstandoutbecausetheyreferenceanumberofpolicytoolsthatwe tendtothinkofastraditionallymacroprudentialinnature. Thetoolsmentionedcanbesplitintotwogroups. First, some speeches discuss green monetary policy in the form of quantitative easing or tailoring central bankassetpurchasestotheenvironmentalprofileofthoseassets.34 Atleastonespeechraisesthepossibility of allowing for favor green collateral in monetary policy credit operations, but raises this possibility with a good deal of skepticism.35 The second group of climate-related tools operate through the banking sector and includes risk weighting “discounts” for green assets and “penalizing factors” for brown assets as well as capital requirements tailored to carbon intensive exposures.36 Regardless of whether it is green monetary policy or capital requirements, these tools tend to be discussed more in the context of supporting the transitiontoanetzerocarbonworldbysupportinggreenfinanceasopposedtotoolsdirectedtoaddressing climate-relatedfinancialstabilityrisks. Tothisend,atleastsomediscussioncentersaroundthequestionof how effective central banks can be in achieving climate-related objectives and highlights the possibility of unintendedconsequences,includingdisintermediationandrisk-shifting.37 All told, the evidence suggests that central banks have very little to say at this point about specific macroprudential policy tools in the context of climate-related financial stability. Most discussion of policy toolsoractionstendstobemorefocusedonmicroprudentialpolicywithaparticularemphasisonmandatory disclosureand/orclimatescenarioanalysis. Ininstanceswherethepossibilityofsystemicriskdoesfeature moreprominently,thediscussionofavailabletools—suchasstresstestingthatgoesoutsideofthebanking sector—stillblursthelinesbetweenmicroandmacroprudentialpolicy. transitionrisksforbanks’solvency,alongwiththeirlendingcapacityandtheimplicationsfortheoveralleconomy.”;Lane(2019) “Climate change and the Irish financial system” (web link) “This calls for ongoing monitoring of climate risks, together with thedevelopmentofclimate-drivenscenarioanalysesandstresstests. Byextension,theresultsofsuchanalysesmaycallforthe appropriatemacroprudentialpoliciestomitigatetheserisks.”;Carney(2019)“TCFD:strengtheningthefoundationsofsustainable finance”(weblink), “Supervisorsoffinancialsectorfirmswillalsoneedtoconsiderwhichmetricsaremostusefulfordifferent levelsofassessment:-Microprudential,toassesshowindividualfirmsaremanagingclimate-relatedrisks-forexampletheimpact ofaphysicalortransitionriskonaloanbook.-Macroprudential,toconsiderhowandwhetherindividualexposurescouldscaleup tosystemicrisk-Macroeconomic,tohelpunderstandhowthefinancialsystemandeconomyinteractindifferentclimatetransition scenarios.” 34Villeroy de Galhau (2021) “The role of central banks in the greening of the economy” (web link) and Mauderer (2019) “Centralbanks-acrisismanagerfortheclimate?”(weblink) 35Weidmann(2019)“Climatechangeandcentralbanks”(weblink) 36Bailey(2021)“Tacklingclimateforreal-theroleofcentralbanks”(weblink),Matsen(2019)“Climatechange,climaterisks andNorgesBank”(weblink) 37Weidmann (2019) “Climate change and central banks” (web link) and Diaz de Leon (2021) “Remarks - How can central bankersandsupervisorssupportclimaterisksandgreenfinanceandmanagerisks?”(weblink) 22

4.2 OtherTopicsinClimate-relatedSpeeches Centralbankscoverawidevarietyofclimate-relatedtopicsthatgobeyondfinancialstabilityand/ormacroprudential policy. Focusing on the 386 climate-related speeches shown in Figure 6 that have no overlap with financial stability or macroprudential policy, we use cluster analysis at the sentence level to identify some other topics that feature prominently.38 The topics we identified are: Climate Impact/Transition; Supervision and Regulation; Financial System; Sustainable Finance; Financial Innovation; Asset Allocation; MonetaryPolicy; andCentralBankMandate. WordcloudsforeachtopicarepresentedinFigure8togive abettersenseofthesubjectmatterineachcluster. Figure8: Wordcloudsforclimate-relatedtopicsidentifiedviaclusteranalysis Source:Author’scalculations. Figure 9 shows the distribution of sentence topic shares in individual speeches (i.e., the number of sentencesinaspeechfocusingonagiventopicofinterestexpressedasafractionofthetotalnumberofsentencesinthatspeech)forthetopicsweidentify. Foreachtopic,thedotistheaveragesentenceshareacross allspeeches,thedashisthemedian,andtheboxrepresentedthe25th and75th percentile. Sentencesshares are distributed relatively evenly across all eight topics, although some are more prevalent than others (for example, Sustainable Finance relative to Climate Impact/Transition) and some distributions have notably longerandfattertailsthanothers(forexample,SupervisionandRegulationrelativetoMonetaryPolicy). 38Weapplyk-meansclusteringmethodtoall43,372sentencesin386climate-relatedspeecheswhicharerepresentedby300dimensionalvectorsthataremappedusingwordembeddingtechniques(seeAppendixBforadetaileddescriptionofourmethodology).AsdiscussedinSection3.1,unsupervisedtopicmodelssuchasLDAcouldbealternativewaystodiscoverunknowntopics. LDAcouldprovidesimilarclusteringresultstothosedemonstratedinsection4.2,buttheywerelessinterpretableandnotrobustin thesensethatresultswereheavilydependingonhyperparametersthatarearbitrarysetbeforeestimation,samplingalgorithmsand waysofwordpre-processing. 23

Figure9: Sentencesharedistribution,bytopic Source:Author’scalculations. Table 7 shows correlations in sentences shares across topics. A positive correlation across two topic pairs suggests that both topics tend to show up together in the same speech whereas a negative correlation suggests the opposite. Additional details on the types of issues discussed in each topic are provided below, but the correlation table makes it clear that certain topics naturally tend to be discussed together. For example, sentences that touch on Climate Impact/Transition tend to show up together in the same speech withdiscussionofSupervisionandRegulationandtheFinancialSystem. Monetarypolicyinthecontextof climate change tends to be discussed in close association with the central bank mandate. Some degree of correlation across topics is to be expected because the metrics are constructed at the level of an individual sentencewithinaspeechandcentralbankspeechescovermanytopics. Table7: Correlationofsentencesharesacrosstopics Climate SupervisionFinancial Sustainable Financial Asset Monetary Central Impact and Reg- System Finance Innova- Alloca- Policy Bank /Transi- ulation tion tion Mandate tion ClimateImpact/Transition 1.00 0.22∗∗∗ 0.27∗∗∗ 0.04 −0.14∗∗ −0.21∗∗∗ −0.18∗∗∗ 0.03 SupervisionandRegulation . 1.00 0.35∗∗∗ 0.15∗∗ −0.12∗ 0.08 −0.32∗∗∗ −0.11∗ FinancialSystem . . 1.00 −0.23∗∗∗ −0.21∗∗∗ 0.13∗∗ −0.13∗ -0.06 SustainableFinance . . . 1.00 0.18∗∗∗ −0.15∗∗ −0.39∗∗∗ −0.29∗∗∗ FinancialInnovation . . . . 1.00 −0.03 −0.26∗∗∗ −0.33∗∗∗ AssetAllocation . . . . . 1.00 −0.10∗ −0.06 MonetaryPolicy . . . . . . 1.00 0.39∗∗∗ CentralBankMandate . . . . . . . 1.00 Notes:Asterisksdenotestatisticalsignificance,***indicatesp-value<0.001;**indicatesp-value<0.01;*indicatesp-value<0.05. Source:Author’scalculations. 24

Climate Impact/Transition. Speeches with a high share of sentences in this category touch on the impactofclimatechangeontheeconomyandthefinancialsystem. Thelong-termimpactofclimatechange isoftendiscussedinthecontextofamanagedtransitiontoazeroemissionsglobaleconomy,whichexplains the positive correlation in sentence share with the Supervision and Regulation category.39 Speeches also touch on using the leverage of the financial system to help to manage this transition, hence the positive correlationwiththeFinancialSystemcategory. Intermsofcentralbankactions,measurement,transparency, and additional research on the topic are common. In addition, there are strong calls for global cooperation inbetterunderstandingtheeconomicandfinancialimplicationsofclimatechange. SupervisionandRegulation. Thistopicfocusesonriskmanagementpractices,forsupervisedinstitutionsaswellasforsupervisorsthemselves,andthealignmentofbankactionswithsupervisoryexpectations features prominently.40 Some speeches step outside the banking sector to address risk management in the insurance industry and liability risk is discussed in this context. Potential policy actions include improved measurement, promoting transparency in reporting, and international cooperation in standard setting. In addition, scenario analysis focused on the resilience of individual financial institutions (as opposed to topdownclimatestresstestingfromasystemicrisk,financialstabilityperspective)isdiscussed. Inadditionto Climate Impact/Transition and Financial System, sentences related to Supervision and Regulation tend to alsoshowupinspeechesthataddressSustainableFinance. FinancialSystem. Thefinancialsystemisdiscussedinclimate-relatedspeechesinwaysthatgobeyond financial stability. For example, it is raised in the context of a lack of understanding of risk, often with an emphasisonuncertainty.41 Thelackofawell-developedanalyticframeworkfeedsthislackofunderstanding and is a concern for central banks, regulators, and financial market participants alike and this point is frequentlyraised. Somespeechesarecarefultopointoutthegreentransitioncreatespotentialopportunities forinvestors. Withregardtopolicytoolsoractions,transparency,globalcooperation,andmoreresearchon thetopiccomeupfrequently. Scenarioanalysis/stresstestingisdiscussedinbothamicroprudentialaswell as a macroprudential context. In addition to Climate Impact/Transition and Supervision and Regulation, sentencesrelatedtoFinancialSystemtendtoalsoshowupinspeechesthataddressAssetAllocation. Sustainable Finance. Sustainable finance captures the idea that central banks and regulators play an centralroleinsupportingthealignmentoffinancialindustryresponseswithnationalclimatestrategiesand priorities.42 Theobjectiveistoencouragethedevelopmentofasustainablefinancialsystemonthetransition path to a low-carbon economy. Coordination between the private sector, various domestic government agencies,includingfinancialregulators,andinternationalbodiesisimportant. Speecheswithahighshareof 39Thethreespeecheswiththehighestsentenceshareinthiscategoryare:Stournaras(2017)“Climatechange-challenges,risks andopportunities”(weblink);Stournaras(2019)“Climatechange-threats,challenges,solutionsforGreece”(weblink);Elderson (2021)“Forestsandfinance”(weblink). 40Elderson(2021)“Overcomingthetragedyofthehorizon”(weblink);deGuindos(2019)“Speakingnotesonclimate-related risks”(weblink);Jain(2021)“BuildingamoreresilientfinancialsysteminIndiathroughgovernanceimprovements”(weblink). 41Jain(2021)“BuildingamoreresilientfinancialsysteminIndiathroughgovernanceimprovements”(weblink);Poloz(2019) “Release of the Financial System Review” (web link); Breeden (2019) “Avoiding the storm - climate change and the financial system”(weblink). 42NorShamsiahMohdYunus(2021)“Remarks-GreenSwan2021GlobalVirtualConference”(weblink),Yue(2020)“ManagingclimaterisksinHongKong”(weblink),andChew(2021)“Openingremarks-WorldBank’sSustainableExchangEDevelopmentSeries(SEEDS)”(weblink) 25

sentencesrelatedtoSustainableFinancealsotouchonSupervisionandRegulationandFinancialInnovation. Financial Innovation. Climate-related speeches discuss financial innovation largely in the context of promoting a sustainable and inclusive financial system.43 (Sustainable Finance is the only category that shows a positive correlation with Financial Innovation in Table 7.) Examples of topics discussed include the growth of green investment products, the role of fintech in promoting green investing, and the need for increased global emissions trading. At least one speech touches on the potential environmental impact of thehighlevelsofenergyconsumptionrequiredinthecrypto-assetmarket. Asset allocation. Asset allocation, or management, is discussed in the context of market pricing of risk, the balance sheet structure of insurers, or the role of non-bank financial intermediaries in adapting to uncertain climate risk.44 Some speeches discuss incorporating Environmental, Social, and Governance (ESG) factors in investment portfolios sometimes with an emphasis on potential investment opportunities arising from energy transition. Central banks also talk about the carbon footprint of their own activities, eitherthecompositionofthebalancesheetorasitrelatestothemanagementofdomesticpensionfunds.45 Central Bank Mandate. Many speeches touch on how climate may or may not fit into the primary mandate on price stability but in some cases they also touch on secondary mandates in relation to environmental or sustainable growth objectives.46 The thought that ensuring an adequate carbon pricing system is ajobforelectedofficials,notcentralbanks,issomethingthatcomesupinahandfulofspeeches. Monetarypolicy. Finally,speechesdiscussmonetarypolicyinatleasttwocontexts: eitherthroughthe directeffectofclimatechangeoneconomicactivityorthroughthepossibilitythatclimatechangecouldlead to structural change in the transition to net zero.47 Structural change, in turn, could have implications for productivityandthenaturalrateofinterest. Somespeechesareveryskepticalaboutthesuitabilityofmonetarypolicytoaddressclimate-relatedissues,whileothersnotethatcentralbankshavealreadycommittedto incorporatingclimatechangeintocurrentmonetarypolicyframeworks. 4.3 AdvancedversusEmergingMarketEconomies Ourreviewofclimate-relatedtopicsrevealedsomeimportantdifferencesinhowcommunicationishandled inadvancedeconomycentralbanksrelativetotheircounterpartsinemergingmarketeconomies. Figure 10 shows distributions for the climate, financial stability, and macroprudential policy speech scores broken out across the two cohorts of banks. The climate speech score distributions (left panel) are broadlysimilarinthatbotharebi-modal. Thissuggeststherearetwotypesofclimate-relatedspeechesfor both advanced economy and emerging market central banks, those that treat climate relatively lightly (the 43Perrazzelli(2021)“LaunchoftheG20TechSprintongreenfinanceandsustainableeconomy”(weblink);Carney(2019)“A platformforinnovation”(weblink);Santiprabhob(2019)“Thailand2025-dealingwithmajortrends”(weblink). 44Gerken(2021)“DevelopmentsinthePRA’ssupervisionofannuityproviders”(weblink); Gerken(2021)“ThePRA’srole in improving the processes that support insurers’ investment” (web link); de Guindos (2020) “The euro area financial sector opportunitiesandchallenges”(weblink) 45See,forexample,Breman(2020)“HowtheSverigesRiksbankcancontributetoclimatepolicy”(weblink)andKuroda(2021) “TheBankofJapan’sstrategyonclimatechange.”(weblink) 46Lagarde(2021)“PressConference”(weblink);Scicluna(2021)“Pricestabilityandbeyond-understandingtheimpactofthe ECB’sStrategyReview”(weblink);Schnabel(2021)“Societalresponsibilityandcentralbankindependence”(weblink). 47SomeexamplesincludeHernandezdeCos(2021)“TheEuropeanCentralBank’snewmonetarypolicystrategy”(weblink) andMacklem(2021)“RenewaloftheMonetaryPolicyFramework”(weblink) 26

firstpeakatlowerscore)orthosethattreatthetopicmoreintensely(thesecondpeakatthehighscore). One possibleexplanationforthisisthatasaspeechtopic,theprominenceofclimatechangehasgrownovertime. Early speeches touched on it only lightly, but as momentum in the central banking community has grown, thetopicisnowbeingtreatedmoreextensively. Theevidenceonthehighestscoringclimatespeechperyear showninTable4supportsthisexplanationasthespeechscoresriseovertime. Theclimate-scoredistribution for advanced economies has a fatter right tail relative to the emerging market speeches, which supports the higher mean reported in Table 8 (although the two are not statistically different from one another). The middle and right panels shows that advanced economy central banks treat both financial stability and macroprudential policy more intensively then their emerging market counterparts. The difference in the averagespeechscoresreportedinTable8foreachofthesecategoriessupportsthisconclusion. Alltold,the evidencesuggeststhatcentralbanksinadvancedeconomiesseemmorewillingtotalkaboutbothfinancial stability and macroprudential policy in the context of climate change relative to central banks in emerging marketeconomies. Figure10: Speechscoredistribution,advancedeconomyandemergingmarketcentralbanks Source:Author’scalculations. Table8: Averagespeechscore,advancedeconomyandemergingmarketcentralbanks Advanced EmergingMarket Economies Economies Difference p-value ClimateSpeechScore 0.99 0.90 0.09 0.27 FSSpeechScore 0.36 0.28 0.08 0.00 MacropruSpeechScore 0.34 0.29 0.05 0.02 Source:Author’scalculations. The difference in topic treatment extends across beyond financial stability and macroprudential policy, Figure 11 shows the average sentence share for each of our six topics in non-financial stability, non- 27

macroprudential policy speeches broken out across the two subsets of central banks. Advanced economies have a higher average sentence share for Climate Impact/Transition (4.13 for advanced economies versus 3.16 for emerging market economies), Financial System (6.63 versus 5.24), Supervision and Regulation (9.83 versus 9.16), and Asset Allocation (5.56 versus 4.83). For the first two categories, the difference in meansisstatisticallysignificant,butquantitativelysmallandforthesecondtwocategoriesthedifferenceis not statistically different from zero. However, a more meaningful difference comes from the fact that advancedeconomiesaddressMonetaryPolicy(4.64versus1.75foradifferenceof2.78)andtheCentralBank Mandate (5.57 versus 2.50 for a difference of 3.07) notably more than emerging market central banks. At thesametime,emergingmarketeconomiesaddressSustainableFinance(6.44versus13.42foradifference of−6.97)andFinancialInnovation(4.67versus8.14foradifferenceof−3.47)muchmoreextensivelythan centralbanksinadvancedeconomies. Anexplanationforthedifferenceintopictreatmentlikelycomesback to the fact that many central banks in emerging market economies have wider ranging mandates that may includethecentralbanksupportingsustainabledevelopmentobjectives.48 Figure11: Averagesentencesharebytopic,advancedeconomyandemergingmarketcentralbanks Notes: Asterisks denote statistical significance of difference in means between Advanced and Emerging Market Economies, *** indicates pvalue<0.01;**indicatesp-value<0.05;*indicatesp-value<0.10. Source:Author’scalculations. Thisdivergenceintopiccoverageacrossthetwocohortsofbanksisadevelopmentthathasbeenevolving over time. To illustrate this, Figure 12 compares the average sentence share for each non-financial stability, non-macropruprudential policy topic for advanced economies (left panel) and emerging market economies(rightpanel). Theredbarsineachpanelshowthetopicsentenceshareforthefirst100climaterelatedspeeches(coveringtheperiod2000toMarch2019)andtheredbarsshowthesameinformationfor the most recent 100 speeches (covering July 2019 to 2021). Topic coverage for advanced economies has 48SeeCampiglio,etal. (2018)andDikauandVolz(2021)foranextensivetreatmentofhowclimatefitsintothemandatesof variouscentralbanks. 28

moved away from issues related to the Financial System and Financial Innovation to concentrate more on climate in the context of Monetary Policy and the Central Bank Mandate. At the same time, topic coverageforemergingmarketeconomiesasmovedawayfromtalkingaboutSupervisionandRegulationandthe CentralBankMandatetowardSustainableFinanceandAssetAllocation. Figure 12: Average sentence share by topic for first 100 speeches versus recent 100 speeches, advanced economyandemergingmarketcentralbanks Notes: Asterisks denote statistical significance of difference in means between Advanced and Emerging Market Economies, *** indicates pvalue<0.01;**indicatesp-value<0.05;*indicatesp-value<0.10. Source:Author’scalculations. 5 Ambiguity in Climate-related Language Inthissection,weteststatisticallywhetherlanguageusedinclimate-relatedspeechesismateriallydifferent from that used in other speeches. The motivation is based, in part, on the finding from Section 4.1 that the link between financial stability and macroprudential policy for climate-related speeches is out of line with the link typically found in non-climate-related speeches. Moreover, our reading of the treatment of many topicsaddressed inSection 4.2suggestsa hesitancyof centralbanksto addressclimate-related issueswith precise,definitivelanguage. Our focus is on the use of speculative language. The text analysis literature has a number of different ways to characterize speculative language, but we opt for a dictionary approach using keywords that com- 29

monly imply uncertainty: may, might, can and could.49 Using this dictionary, we construct a speculative languagespeechscoreusingasimplecountingmethod. Inthiscase,wecountthenumberofsentencesina givenspeechthatcontainatleastonekeywordfromthedictionaryandexpressitasashareofthetotalnumberofallsentencesinthespeech. Thehighertheshare,thegreaterthedegreeofambiguity. Wecalculated ascoreforeveryspeechinthedatabaseandthencomparedclimate-relatedspeechestoallotherspeeches. Figure13: Qualitativeaspectsoflanguageinclimateversusnon-climate-relatedspeeches Source:Author’scalculations. TheresultsareshowninFigure13whichplotsprobabilitydensityfunctions(PDFs)forourspeculative language score for both climate-related and non-climate-related speeches in the left panel. The right panel showscumulativedistributionfunctions(CDFs)forthetwosetsofspeeches. BoththeCDFandthePDFfor thespeculativelanguagescoreforclimate-relatedspeechesliestotherightofnon-climaterelatedspeeches. Asimplet-testrejectsthenullhypothesisthatthemeans(10.6forclimate-relatedspeechesand9.5fornonclimate related speeches) of these two distributions are same with a very high level of confidence (p-value of< 0.001). Moreover,aKomolgorov-Smirinovtestunderthenullhypothesisthatthetwodistributionsare the same is overwhelming rejected. This statistical evidence suggests that central banks do indeed tend to usemorespeculativelanguagewhencommunicatingaboutclimate-relatedissues. Climate change is a relatively new topic that is outside the traditional areas of central bank expertise. Therisksassociatedwithclimatechangereflectanexceptionallyhighdegreeofuncertainty,aredifficultto measure, and past experience is not helpful for understanding how these risks might evolve in the future. For these reasons, some degree of ambiguity about the topic in public communication seems appropriate. Thatsaid,ascentralbanksdeepentheirknowledgeaboutclimate-relatedrisks,thelanguageusedinpublic communicationshouldevolve. Goingforward,sharpercommunicationwithmorepreciselanguagecanhelp establishcredibilitythatcentralbanksarebeingresponsivetoclimatechange. 49Focusingonuseofmodalauxiliaries(modalities)inspeechesandtextsisanareaofspecialitywithinthefieldofsociolinguistics(see,forexample,Torres(2021),JaimeandPerez-Guillot(2015),deWaardandMaat(2012)).Withintheeconomicsliterature, severalpapersapplythesemethodstoexplorecentralbankcommunication; MundayandBrookes(2021)usestheproportionof wordsthataremodalverbsinanewsarticletogaugetheuncertaintycomponentofthenewsvalueofBankofEnglandpublications. Kawamuraetal,(2019)utilizemodalityintheJapaneselanguagetomeasureambiguityintheBankofJapan’smonthlyeconomic report. 30

6 Conclusion Textanalysistoolswereappliedtoasetofnearly17,000speechesgivenbyover100centralbanksoverthe past twenty-five years. We presented a novel methodology to identify the subset of central bank speeches that address issues related to climate change. Once identified, these speeches were analyzed to assess the extenttowhichcentralbankstalkabouttheuseofmacroprudentialpolicytoolstoaddressfinancialstability risks associated with climate change. We also examined a range of climate-related topics that go beyond financialstabilityandmacroprudentialpolicy. Ourresultsshowthatcentralbankshaveincreasedcommunicationwiththepublicaboutclimatechange, withanespeciallysharpincreaseinrecentyears. Centralbankstouchonfinancialstabilityconcernsassociatedwithclimatechange,buttheyrarelydiscussmacroprudentialpolicy. Outsideofthesetwotopics,central bankstouchonawidevarietyofotherclimate-relatedtopics,including: theimpactofclimatechangeonthe economy,climateinthecontextofthefinancialsystemandsupervisionandregulation,financialinnovation, andassetallocation. Advancedeconomycentralbankstendtofocusmoreattentiononhowclimatechange fitsinwithmonetarypolicyandthecentralbankmandate. Emergingmarketcentralbanksconcentratemore on sustainable finance. To the extent that communication touches on direct central bank actions, the focus tendstobeonmicroprudentialsupervisionandregulationincludingtopicssuchassupervisoryexpectations, stresstesting,andmandatorydisclosuresofclimateriskexposure. Wealsoprovidestatisticalevidencethat suggestscentralbankstendtousevagueandspeculativelanguageinclimate-relatedspeechesinawaythat standsoutrelativetootherspeeches. Lookingforward,additionalresearchisneededtobetterunderstandtheroleofcentralbanksinreacting tooraddressingclimate-relatedissues. Thisadditionalclaritywillhelpcentralbankstocommunicatemore clearlyaboutanurgentpublicpolicyissueand, indoingso, willhelpcrediblyestablishthattheyarebeing responsivetoclimatechange. 31

References [1] Armelius,H.,C.Bertsch,I.HullandX.Zhang(2020)“SpreadtheWord: InternationalSpilloversfrom CentralBankCommunication,”JournalofInternationalMoneyandFinance,Volume103,102-116. [2] Baker, S., N. Bloom and S. Davis (2016) “Measuring Economic Policy Uncertainty,” The Quarterly JournalofEconomics,Volume131,Issue4,Pages1593-1636,November2016. [3] Benchimol,J.,S.Kazinnik,andY.Saadon(2021)“FederalReserveCommunicationandtheCOVID- 19Pandemic,”CovidEconomics: VettedandReal-TimePapers,Issue79,218-256. [4] Bholat, David, N.Broughton, J.Ter MeerandE.Walczak(2019)“EnhancingCentralBankCommunicationsusingSimpleandRelatableInformation,”JournalofMonetaryEconomics,vol.108,1-15. [5] Bholat,David,S.Hansen,P.Santos,andC.Schonhardt-Bailey(2015)“TextMiningforCentralBanks: Handbook,”CentreforCentralBankingStudies(33).pp.1-19. [6] Blei,DavidandJohnLafferty(2005)“CorrelatedTopicModels,”Proceedingsofthe18thInternational ConferenceonNeuralInformationProcessingSystems. [7] Blei,David,andJohnLafferty(2006)“DynamicTopicModels,”Proceedingsofthe23rdInternational ConferenceonMachineLearning,113-120. [8] Blei, David, A. Ng and M. Jordan (2003) “Latent Dirichlet allocation.” Journal of Machine Learning Research,3,993-1022. [9] Blinder, Alan, M. Ehrmann, M. Fratzscher, J. de Haan and D. Jansen (2008) “Central Bank Communication and Monetary Policy: A Survey of Theory and Evidence,” Journal of Economic Literature, Volume46,Issue4,910-945. [10] Bollen,Johan,HuinaMaoandXiaojunZeng(2011)“TwitterMoodPredictstheStockMarket,”JournalofComputationalScience,2(1): 1-8. [11] Bolton, Patrick, Morgan Despres, Luiz Awazu Pereira da Silva, Frédéric Samama and Romain Svartzma (2020) “The Green Swan: Central Banking and Financial Stability in the Age of Climate Change,”BankforInternationalSettlements,January2020. [12] Born, Benjamin, Michael Ehrmann and Marcel Fratzscher (2014) “Central Bank Communication on FinancialStability,”TheEconomicJournal,124,no.577: 701-734. [13] Campiglio,Emanuele,YannisDafermos,PierreMonnin,JoshRyan-Collins,GuidoSchottenandMisa Tanaka(2018)“ClimateChangeChallengesforCentralBanksandFinancialRegulators,”NatureClimateChange,Volume: 8,462-468. 32

[14] Chenet,Hugues,JoshRyan-CollinsandFrankvanLerven(2021)“Finance,Climate-ChangeandRadicalUncertainty: TowardsaPrecautionaryApproachtoFinancialPolicy,”EcologicalEconomics,Volume183,106957. [15] Correa, Ricardo, Keshav Garud, Juan Londono and Nathan Mislang (2021) “Sentiment in Central Banks’ Financial Stability Reports,” Review of Finance, Volume 25, Issue 1, Pages 85-120, February 2021. [16] de Waard, A., and H. P. Maat (2012) “Epistemic Modality and Knowledge Attribution in Scientific Discourse: A Taxonomy of Types and Overview of Features.” In: Proceedings of the 50th Annual MeetingoftheAssociationforComputationalLinguistics,pages47–55. [17] Dikau, Simon, and Ulrich Volz (2021) “Central Bank Mandates, Sustainability Objectives and the PromotionofGreenFinance,”EcologicalEconomics,Volume184,107022. [18] Financial Stability Board (2015) “FSB to establish Task Force on Climate-related Financial Disclosures,”PressreleaseonDecember4th,2015. [19] Gentzkow,M.,B.KellyandM.Taddy(2019)“TextasData,”JournalofEconomicLiterature,57(3): 535-74. [20] Gentzkow,M.,J.M.Shapiro,andM.Taddy(2019)“MeasuringGroupDifferencesinHighdimensional Choices: MethodandApplicationtoCongressionalSpeech”Econometrica,Vol.87,No.4(July,2019), 1307–1340. [21] Gentzkow,Matthew,andJesseShapiro(2010)“WhatDrivesMediaSlant? EvidencefromU.S.Daily Newspapers,”Econometrica,Vol.78,No.1,pp.35-71. [22] Grimmer, J., and B. Stewart (2013) “Text as Data: The Promise and Pitfalls of Automatic Content AnalysisMethodsforPoliticalTexts,”PoliticalAnalysis,21(3),267-297. [23] Gürkaynak, Refet, Brian Sack and Eric Swanson (2005) “Do Actions Speak Louder than Words? the Response of Asset Prices to Monetary Policy Actions and Statements,” International Journal of CentralBanking,vol.1(1). [24] Guthrie, Graeme, and Julian Wright (2000) “Open Mouth Operations,” Journal of Monetary Economics,vol.46(2),pages489-516. [25] Haldane, Andrew, and Michael McMahon (2018) “Central Bank Communications and the General Public,”AEAPapersandProceedings,108: 578-83. [26] Hansen, Lars Peter, (2022) “Central Banking Challenges Posed by Uncertain Climate Change and NaturalDisasters,”JournalofMonetaryEconomics,Volume125,Pages1-15,January2022. 33

[27] Hansen, S., and M. McMahon (2016) “Shocking Language: Understanding the Macroeconomic EffectsofCentralBankCommunication,” JournalofInternationalEconomics, Volume99, Supplement 1,PagesS114-S133,March2016. [28] Hansen, S., M. McMahon and A. Prat (2018) “Transparency and Deliberation Within the FOMC: A Computational Linguistics Approach,” The Quarterly Journal of Economics, Volume 133, Issue 2, Pages801-870,May2018. [29] Hansen, S., M. McMahon and M. Tong (2019) “The Long-Run Information Effect of Central Bank Communication,”JournalofMonetaryEconomics,Volume108..Pages185-202. [30] Hansson, Magnus, (2021) “Evolution of Topics in Central Bank Speech Communication,” Working Papers,UniversityofGothenburg,DepartmentofEconomics. [31] Jaime,A.,andC.Perez-Guillot(2015)“AComparisonAnalysisofModalAuxiliaryVerbsin´TechnicalandGeneralEnglish,”Procedia-SocialandBehavioralSciences,Volume212,Pages292-297. [32] Kawamura,K.,Y.Kobashi,M.ShizumeandK.Ueda(2019)“StrategicCentralBankCommunication: Discourse Analysis of the Bank of Japan’s Monthly Report,” Journal of Economic Dynamics and Control,Volume100,Pages230-250. [33] King, Gary, Patrick Lam and Margaret Roberts (2017) “Computer-Assisted Keyword and Document SetDiscoveryfromUnstructuredText,”AmericanJournalofPoliticalScience,61: 971-988. [34] Kryvtsov,Oleksiy,andLubaPetersen(2021)“CentralBankCommunicationthatWorks: Lessonsfrom LabExperiments,”JournalofMonetaryEconomics,vol.117,pages760-780. [35] Larsen,V.H.,L.A.Thorsrud,andJ.Zhulanova(2021)“News-driveninflationexpectationsandinformationrigidities,”JournalofMonetaryEconomics,Volume117,2021,Pages507-520. [36] Laver, Michael, Kenneth Benoit and John Garry (2003) “Extracting Policy Positions from Political TextsusingWordsasData,”AmericanPoliticalScienceReview,97(2): 311-331. [37] Loughran, Tim, and Bill McDonald (2011) “When is a Liability not a Liability? Textual Analysis, Dictionaries,and10-Ks,”JournalofFinance,66(1): 35-65. [38] Lowe,Will,(2008)“UnderstandingWordscores,”PoliticalAnalysis,16(4): 356-371. [39] McKibbin, Warwick, Adele Morris, Peter Wilcoxen and Augustus Panton (2021) “Climate Change and Monetary Policy: Issues for Policy Design and Modelling,” Oxford Review of Economic Policy, Volume36,Issue3,Autumn2020,Pages579-603,August2021. [40] Montes, Gabriel, and Rodolfo Nicolay (2017) “Does Clarity of Central Bank Communication Affect Credibility? Evidences Considering Governor-Specific Effects,” Applied Economics, 49:32, 3163- 3180. 34

[41] Moschella, Manuela, and Luca Pinto (2018) “Central Banks’ Communication as Reputation Management: How the Fed Talks under Uncertainty,” Public Administration, Volume 97, Issue 3, Pages 513-529. [42] Munday,T.andJ.Brookes(2021)“Markmywords: thetransmissionofcentralbankcommunication tothegeneralpublicviatheprintmedia,”BankofEnglandStaffWorkingPaperNo.944. [43] Nakamura, Emi, and Jon Steinsson (2018) “High-Frequency Identification of Monetary Non- Neutrality: The Information Effect,” The Quarterly Journal of Economics, vol. 133(3), pages 1283- 1330. [44] Pennington, J., R. Socher, C. Manning (2014) “GloVe: Global Vectors for Word Representation,” Proceedingsofthe2014ConferenceonEmpiricalMethodsinNaturalLanguageProcessing(EMNLP), October2014,pages.1532–1543. [45] Roberts, Margaret, Brandon Stewart, Dustin Tingley, and Edoardo Airoldi (2013) “The Structural TopicModelandAppliedSocialScience,”AdvancesinNeuralInformationProcessingSystemsWorkshoponTopicModels: Computation,Application,andEvaluation,Volume4,pp.1-20. [46] Shapiro,Adam,andDanielWilson(2021)“TakingtheFedatItsWord: ANewApproachtoEstimating CentralBankObjectivesusingTextAnalysis,”TheReviewofEconomicStudies,December2021. [47] Torres,P.J.(2021)“TheRoleofModalsinPolicies: TheUSOpioidCrisisasaCaseStudy,”Applied CorpusLinguistics,100008. [48] Tetlock, Paul C., (2007) “Giving Content to Investor Sentiment: The Role of Media in the Stock Market,”JournalofFinance,62(3): 1139-68. [49] Watanabe, K., (2018) “Newsmap: A Semi-Supervised Approach to Geographical News Classification,”DigitalJournalism,6(3),294-309. [50] Watanabe,K.,andY.Zhou(2020)“Theory-DrivenAnalysisofLargeCorpora: SemisupervisedTopic ClassificationoftheUNSpeeches,”SocialScienceComputerReview,February2020. [51] Weitzman, Martin, (2009) “On Modeling and Interpreting the Economics of Catastrophic Climate Change,”ReviewofEconomicsandStatistics,91(1),1-19. [52] Wisniewski,Piotr,andBrendanLambe(2013)”TheRoleofMediaintheCreditCrunch: TheCaseof theBankingSector,”JournalofEconomicBehaviorandOrganization,85: 163-75. 35

Appendix A. List of Central Banks in Speech Dataset Our analysis covers 108 central banks each of which has at least one speech in our Central Bank Speech Dataset. Weclassifythesecentralbanksintoadvancedandemergingmarketeconomiesasinthelistbelow. ThisclassificationisbasedonIMF’sclassificationfortheirWorldEconomicOutlook(seehttps://www.imf. org/external/pubs/ft/weo/2021/01/weodata/groups.htm). A.1 AdvancedEconomies: EuropeanCentralBank,BankofFrance,BankofEngland,DeutscheBundesbank,NetherlandsBank,Bank of Italy, Bank of Spain, Monetary Authority of Singapore, Central Bank of Ireland, Reserve Bank of New Zealand, Bank of Japan, Bank of Finland, Federal Reserve System, Central Bank of Norway, Bank of Greece,BankofCanada,HongKongMonetaryAuthority,SwissNationalBank,ReserveBankofAustralia, SverigesRiksbank,DanmarksNationalbank,BancodePortugal,CentralBankofMalta,BankofLithuania, Bank of Israel, Central Bank of Iceland, Bank of Korea, National Bank of the Republic of Austria, Czech National Bank, Central Bank of Luxembourg, National Bank of Belgium, Monetary Authority of Macao, BankofEstoniaBankofLatvia,NationalBankofSlovakia,BankofSlovenia,CentralBankofCyprus. A.2 EmergingMarketEconomies: Central Bank of Malaysia, Central Bank of the Philippines, Reserve Bank of India, Reserve Bank of Fiji, Bank of Mexico, The People’s Bank of China, Bank Indonesia, Bank of Thailand, Bank of Mauritius, Central Bank of Trinidad and Tobago, National Bank of Poland, Central Bank of The Bahamas, Bank of Albania,CentralBankofKenya,NationalBankofSerbia,CentralBankofBarbados,CentralBankofNigeria, Central Bank of Kuwait, South African Reserve Bank, Bank of Zambia, Bank of Uganda, State Bank ofPakistan,CentralBankofChile,CentralBankoftheRepublicofTurkey,NationalBankoftheRepublic of North Macedonia, Central Bank of Sri Lanka, National Bank of Romania, Bank of Papua New Guinea, Bank of Ghana, Central Bank of Bahrain, Bank of Botswana, Bulgarian National Bank, Central Bank of theRussianFederation,BankofNamibia,CentralBankofCuraçaoandSintMaarten,CentralBankofArgentina,SaudiCentralBank,ReserveBankofMalawi,BankofJamaica,CentralBankofSolomonIslands, NationalBankofUkraine,EasternCaribbeanCentralBank,CentralBankoftheRepublicofKosovo,Central Bank of Bosnia and Herzegovina, Central Bank of Nepal, Central Bank of Seychelles, Bank of Sierra Leone,CentralBankofBrazil,MagyarNemzetiBank,CentralBankoftheUnitedArabEmirates,Croatian National Bank, Bank of Algeria, Central Bank of Samoa, Bank Al-Maghrib (Central Bank of Morocco), Bank of Mozambique, Maldives Monetary Authority, Cayman Islands Monetary Authority, Central Bank of Colombia, Bank of Guyana, Reserve Bank of Vanuatu, Bank of Guatemala, Bank of Tanzania, Central BankofArmenia,CentralBankofAruba,CentralBankofBelize,CentralBankofBolivia,CentralBankof Ecuador,CentralBankofJordan,CentralBankofTheGambia,CentralBankofUruguay,NationalBankof Cambodia. 36

Appendix B. Text Analysis Tools Thismethodologicalappendixdescribetextanalysistechniquesusedinourpaper. B.1 Textpre-processing Inourtextpre-processingstep,wefollowstandardpracticessuchasmethodologiesdescribedinBholat,et al. (2015),Benoit,etal. (2018),andGrimmer,etal. (2022).50 Therawspeechtextcontainsmanysymbols forpunctuationandcommonwordssuchas“the”and“and”whichdonotaddanalyticalvalue. Assuch,it iscommontopre-processtherawtexttoreduceitdowntoitsmostvaluablecontent. Textpre-precessingis regardedasanimportantstepintextanalysistomaketextdatahandyandinformative. First, we strip out symbols and numbers. Second, we remove a subset of words that are verycommon. Verycommonwords,so-called“stopwords,”includearticles(“the,”“a”),conjunctions(“and,”“or”),forms oftheverb“tobe,”andsoon. Weusethestop-wordlistsbasedontheSMART(SystemfortheMechanical AnalysisandRetrievalofText)InformationRetrievalSystemdevelopedatCornellUniversityinthe1960s. Next, we recast words into their common linguistic root using a Part-of-Speech (PoS) tagger of the spaCy package for Python (Honnibal, et al., 2020). It can identify the grammatical class (part-of-speech) to which each word belongs—nouns, pronouns, verbs, etc.—and convert them into their base form by a lemmatizationalgorithmthatusessomerulesandadictionaryofirregularpatterns. Forexample. theword “banks”willbeconvertedto“bank,”theword“saw”willbeconvertedto“see,”andtheword“meeting”is converted to “meet” or “meeting” depending on its use in a sentence. Lemmatization reduces the number of unique types of words that carry the same information and makes a large corpus more manageable.51 Moreover, the PoS tagger can identify noun phrases by dependency parsing. It is extremely useful in our analysistotreatnounphrasessuchas“climatechange”asacombinedwordratherthantwoseparatewords “climate”and“change”. Thenounphrase“financialstability”hasaspecificmeaningforcentralbanks,and weprefertouseitratherthantwocommonwords“financial”and“stability.”52 Finally,weremoveasubsetofwordsthatareveryrare. Ourspeechcorpusincludes2,176,150sentences and 992,740 unique word types after pre-processing as described above. We remove less frequent words from the corpus depending on the characteristics of the analysis, especially for those that require large computational resources and/or where low frequent words do not play an important role (but rather give noise) in results. Specifically, truncating words that appear less than 5 times for word embedding or 100 timesforscorecalculationreducesthetypesofwordsto123,518or10,965,respectively. 50WeusetheRpackagequantedav.3.1.0(Benoit,etal.,2018)toprocesstextdataandcalculatestatisticsacrossspeechesand sentences. 51Wordstemming,anotherpopulartechniquethatdiscardstheendofawordusingasimplerule,isregardedasanapproximation oflemmatization. 52Another popular way to deal with noun phrases is to take all combinations of n-grams, which is a phrase of length n in a sentence.However,usinghigherordern-gramssignificantlyincreasesthenumberofwordtypesinthevocabulary,soitrequiresan additionaltreatmentforremovinglessinformativephrases. 37

B.2 Methodologyforidentifyingclimate-relatedspeeches AsdescribedinSection3.2,ourmethodologytoidentifyclimate-relatedspeechesconsistsoftwosteps: the firststepistocalculateaclimatespeechscorebasedonaseedword(“climatechange”)usedtodetectaset ofpotentialclimaterelated-speeches(pre-refinement);thesecondstepistorefinethesetofclimate-related speechesusingourowndictionarydevelopedbyasemi-automatedway. B.2.1 ClimateSpeechScore ThebasicideaofourautomatedscoringmethodisbasedonLaver,etal. (2003)andWatanabe(2018). Ina politicaleconomycontext,Laver,etal. (2003)constructwordscoresusingpre-selectedreferencetextsthat arefromoppositepoliticalpositionsandusethescorestoscaleideologicalpositionsoftargettexts. Watanabe(2018)usessimpleseedwordstogeneratealargetrainingdatasetforscoringwithoutadditionalhuman interventions. WatanabeandZhou(2020)advocatethatWatanabe’smethodisusefulfortopicclassification inthesensethatresearcherscaneffectivelyclassifytextsintomultiplecategoriesconsistentwiththeirspecificinterests. Wemodifythesemethodologiestofitourtaskofsingletopicidentification. Concretely,our methodology ofcalculating aclimate speechscore and usingit to identify(pre-refinement) climate-related speechesisthefollowing: 1. Usingasingleseedword(“climatechange”),wedivideallspeechesinto2sets: speechesthatinclude theseedword (S )andspeechesthatdonotincludetheseedword (S¯ ). seed seed 2. WeconstructclimatewordscoreforeachwordibycalculatingrelativewordfrequencybetweenS seed andS¯ asfollows: seed ClimateWordScore = log(wf )−log(wf ) , i i,Sseed i,S¯ seed wherewf isanormalisedwordfrequencyofwordiinasetofspeechesS calculatedasthenumber i,S oftimeswordiappearedinasetofspeechesS,dividedbythetotalnumberofwordsinthosespeeches. 3. Usingtheclimatewordscores,wecalculateclimatespeechscoreforeachspeech jbytakingweighted averageofwordscoresusing(normalized)wordfrequenciesasweights, (cid:88) ClimateSpeechScore = (ClimateWordScore × wf ) . j i i,j i 4. Weidentifyasetof(pre-refinement)climate-relatedspeeches(S )thathaveaclimate climate-related(pre) speechscoregreaterthan0(i.e.,ClimateSpeechScore > 0 for ∀l ∈ S ). l climate-related(pre) B.2.2 KeywordDictionaryforRefinement As described in Section 3.2, we find that the pre-refinement set of climate-related speeches successfully includesmostspeechesthatwewanttoidentifiedasclimate-related,butitalsoincludeslotsofnon-relevant speeches as well. We manually examine non-relevant speeches that have high climate speech scores and 38

select keywords to exclude these speeches. Therefore, our refinement step can be thought of as defining the boundary of climate-related speech identification. We take the following iterative process for defining akeyworddictionaryandobtainingthefinalsetofclimate-relatedspeecheswiththegoalofautomatingas muchoftheprocessaspossible. 1. We define high score speeches for keyword exploration (S ⊂ S ) using an arbitrary 0 climate-related(pre) threshold for the climate speech score, set to two standard deviations over the mean for all speeches (0.48forclimatespeechscore). 2. We explore a set of keywords (D = {k , k , ..., k }) from the high score speeches (S ) by the ref 1 2 N 0 followingiterativeprocess. (a) In Iteration n, we select the single most climate-relevant keyword k from a exploration set n of speeches S based on word frequency (the number of speeches in S that include the n−1 n−1 keyword) and our judgement (where we know that keywords that have a higher climate word scorearemorelikelyclimate-relevant). (b) As an exploration set for the next iteration (S ), we collect speeches that do not include the n selectedkeywordk fromthepreviousexplorationsetS . n n−1 (c) Continuetoitterateuntilwecannolongerfindanyclimate-relevantwordintheexplorationset. 3. Usingthekeyworddictionary(D ),weidentifyasetof(after-refinement)climate-relatedspeeches ref the subset of (pre-refinement) climate-related speeches (S ⊂ S ) which climate-related climate-related(pre) containatleasttwoseparatesentenceswithanyidentifiedkeywords. B.3 Methodologyforidentifyingothertopicsinclimate-relatedspeeches InSection4.2,weemploythek-meansalgorithmtoidentifytopicsinclimate-relatedspeeches. Thismethod, whichisone ofthemoststandard toolsforclustering, isused inconjunctionwitha wordembeddingtechniquefollowingworksbyMikolov,etal. (2013)(knownasword2vec)andPennington,etal. (2014)(known asGloVe). ThecombinationofthesetechniquesissuggestedbyGrimmer,Roberts,andStewart(2022).53 Thewordembeddingtechniquemapsalargecorpusintolow-dimensionalwordvectorssuchthatwords which have similar meaning/usage are represented by similar vectors. More specifically, the GloVe algorithmisbasedonthemodelthattakestheform, P(k|i) (w −w )Tw = log , i j k P(k|j) where: w denotesawordvectorassignedforwordi,and P(k|i)indicatesawordco-occurrenceprobability i thatwordkappearsinthecontextofwordi. Ifwordiand jareusedforthesimilarcontexts,P(k|i)/P(k|j) → 1 and (w − w ) → 0. Based on the model, the GloVe algorithm finds the word vectors by conducting a i j weightedleastsquareregressionthatassignssmallweightstofrequentco-occurrences. 53SeealsoKozlowski,Taddy,andEvans(2019)andDieng,RuizandBlei(2020)forpotentialapplicationsofwordembedding otherthanusingforclustering. 39

To measure the word co-occurrence probability (as a matrix form C), we count occurrences of any wordslocated10wordsbeforeor10wordsafterthetargetword,weightedbyadecayingfactor(1/"distance between the words"). We set the dimension of the word vectors to 300. In our corpus, the number of word types in vocabulary is 123,518, so the 123,518×123,518 sized word co-occurrence matrix can be representedby300×123,518sizedsetoftheestimatedvectorswhereC = vTv. Whenwehavevectorrepresentationsofthecorpus,wecanthenapplythestandardk-meansalgorithm to the vectors in order to classify them. In our implementation, we calculate sentence vectors by simply averaging the word vectors assigned in words in the corresponding sentence.54 Using the sentence vectors asaninput,thek-meansalgorithmminimizesthefollowingobjectivefunction: (cid:88)N (cid:88)T f(µ,t,v) = d(v,µ)·I(c = t) , i t i i=1 t=1 where: Nisthenumberofsentences,Tisthenumberoftopics(clusters),I(c = t)isaclusterindicatorthat i equalsto1ifsentenceiisassignedtotopict,andd(v,µ)isthedistancebetweenavectorofsentenceiand i t aclustercenterµoftopictisgivenbysquaredEuclideandistance: (cid:88)M d(v,µ) = (v −µ )2 , i t i,m t,m m=1 where: M is the dimension of the sentence vectors (set to 300 in our word embedding step) and v is an i,m element of sentence vector for sentencei. Using the results of the optimization, every sentence is assigned toasingletopicwhoseclustercenteristheclosesttothesentencevector. WesetthenumberofclustersK = 20,andmanuallylabeltheestimatedclustersbasedonwordcloudsof sentencesinthecorrespondingclustersasdescribeinSection4.2. Wediscover8meaningfultopicsshown in Figure 8, while other 12 clusters include specific names of countries and central banks, language frequentlyusedinspeeches(suchas“today,”“gentleman,”and“attention”)orgeneralwords(suchas“figure,” “percentage”,and“year”),andarethereforenotofourinterests(i.e.,theyareso-called“garbagetopics”). References [1] Benoit, K., K. Watanabe, H. Wang, P. Nulty, A. Obeng, S. Müller and A. Matsuo (2018) “quanteda: An R package for the quantitative analysis of textual data.” Journal of Open Source Software, 3(30), 774. [2] Bholat,David,S.Hansen,P.Santos,andC.Schonhardt-Bailey(2015)“TextMiningforCentralBanks: Handbook,”CentreforCentralBankingStudies(33).pp.1-19. [3] Dieng, A., F. Ruiz and D. Blei (2020) “Topic Modeling in Embedding Spaces.” Transactions of the AssociationforComputationalLinguistics,2020;8439–453. 54Weusesentencevectorsforclustersratherthanspeech-levelaggregationofwordvectors.Everyspeechcouldincludemultiple topicsinthespeech,sotheclusteringresultsforspeechestendtobemoreambiguous. 40

[4] Grimmer, Justin, Margaret Roberts and Brandon Stewart (2022) Text as Data: A New Framework for MachineLearningandtheSocialSciences,PrincetonUniversityPress. [5] Honnibal,Matthew,InesMontani,SofieVanLandeghem,andAdrianeBoyd(2020)“spaCy: Industrial strengthNaturalLanguageProcessinginPython.” [6] Kozlowski,A.C.,M.Taddy,andJ.A.Evans(2019)“TheGeometryofCulture: AnalyzingtheMeaningsofClassthroughWordEmbeddings”,AmericanSociologicalReview,84(5),pp.905–949. [7] Laver, Michael, Kenneth Benoit and John Garry (2003) ”Extracting Policy Positions from Political TextsusingWordsasData,”AmericanPoliticalScienceReview,97(2): 311-331. [8] Mikolov, T., I. Sutskever, K. Chen, G. Corrado and J. Dean (2013) “Distributed Representations of WordsandPhrasesandTheirCompositionality.”InProceedingsofthe26thInternationalConference onNeuralInformationProcessingSystems,3111–19. [9] Pennington, J., R. Socher, C. Manning (2014) “GloVe: Global Vectors for Word Representation,” Proceedingsofthe2014ConferenceonEmpiricalMethodsinNaturalLanguageProcessing(EMNLP), October2014,pages.1532–1543. [10] Watanabe, K., (2018) “Newsmap: A Semi-Supervised Approach to Geographical News Classification,”DigitalJournalism,6(3),294-309. [11] Watanabe,K.,andY.Zhou(2020)“Theory-DrivenAnalysisofLargeCorpora: SemisupervisedTopic ClassificationoftheUNSpeeches,”SocialScienceComputerReview,February2020. 41

Appendix C. Climate-specific Policy Tools InouranalysisinSection4.1,weidentify65keywordsthatindicateclimate-specificpolicytoolsfromatotal of1,498climate-orgreen-relatedkeywords(i.e.,"ClimateXXX"and"GreenXXX")foundinourcorpus. We classify them into 6 categories: climate disclosure, climate risk management, green taxonomy, climate research, climate scenario analysis/stress testing, and green monetary policy/capital allocation. Table C.1 showsacompletelistofthose65keywordsandtheirclassificationsinto6categories. TableC.1: KeywordsforClimate-specificPolicyTools Category Climate- Keywords related Speeches AllSpeeches 555 ClimateDisclosure 113 climaterelatefinancialdisclosures, climatedisclosure, climaterelatedisclosure, climaterelatefinancialdisclosure, climaterelatedfinancialdisclosures, climateriskdisclosure, climatereporting, climaterelatereporting, climaterelatereportingobligation, climaterelatereportingrequirement, climaterelateriskdisclosure,climaterelatedisclosurerequirement ClimateRiskManagement 48 climateriskmanagement, climateriskanalysis, climateriskassessment, climaterelateriskmanagement, climateassessment, climatechangeriskassessment,climaterelateriskanalysis,climaterelateriskassessment GreenTaxonomy 29 greentaxonomy,greentaxonomygreenbondstandard,greenlabel,greenloantaxonomy, greencertification, greencredential, greencriterion, climaterelatecriterion,greenstandard ClimateResearch 12 climateresearch, climatemodel, climatemodelling, climatechangeanalysis,climatevulnerabilityassessment,climatedataanalysis,climateriskassessmentmodel Climate Scenario Analysis 74 climatescenario, climatescenarioanalysis, climatestresstest, climat- /StressTest estresstesting, climaterelatescenario, climatescenarioexercise, climaterelatestresstest, climateriskscenario, climatebiennialexploratoryscenario, climateriskstresstest, climatechangescenario, climaterelatescenarioanalysis, climateriskmanagementscenarioanalysis, climateriskscenarioanalysis, climatedrivescenarioanalysis, climateriskstresstestanalysis, climatesensitivityanalysis,climatestresstestingsensitivityanalysis Green Monetary Policy / 12 greenmonetarypolicy, greenqe, greenassetpurchaseprogramme, climate- CapitalAllocation greeningmonetarypolicypolicy, greeningmonetarypolicy, climatechangefundprovisioningmeasure, greenmonetarypolicyassetportfolio, climatebasecentralbankpurchase, climateorientpurchase, greenassetpurchase, climatelinkcapitalinstrument 42

Appendix D. List of Reviewed Speeches SpeechesReviewedinSection3.3(ValidityChecks) [1] Hendar (2015) “Increasing cooperation between Indonesia’s central bank and state police,“ Bank Indonesia,https://www.bis.org/review/r150309c.pdf. [2] Lautenschlager, Sabine, (2017) “Cyber resilience - a banking supervisor’s view,” European Central Bank,https://www.bis.org/review/r170629a.pdf. [3] Waas, Ronald, (2014) “Handling alleged payment system and currency exchange crime,“ Bank Indonesia,https://www.bis.org/review/r141023b.pdf. [4] Diokno, Benjamin E., (2021) “Sustainability in investing,“ Central Bank of the Philippines, https: //www.bis.org/review/r200807m.pdf. [5] Elderson, Frank, (2018) “Let the future of finance be that of financing the future,” Netherlands Bank, https://www.bis.org/review/r180503h.pdf. [6] Lane,PhilipR.,(2016)“Dualperspectivesontheinsurancesector–consumerprotectionandfinancial stability,”CentralBankofIreland,https://www.bis.org/review/r160613c.pdf. [7] Nabiullina, Elvira, (2021) “Speech at Association of Russian Banks,” Bank of Russia, https://www. bis.org/review/r210427e.pdf. [8] Olsen, Øystein, (2021) “Monetary policy strategy - from mandate to decisions,” Norges Bank, https: //www.bis.org/review/r211112f.pdf. [9] Gang,Yi,(2021)“HongKong’spositioningandprospectasaninternationalfinancialcentre,”People’s BankofChina,https://www.bis.org/review/r211216o.pdf. [10] Trichet, Jean-Claude, (2008) “Monetary policy in challenging times,” European Central Bank, https: //www.bis.org/review/r080606c.pdf. [11] Yue,Eddie,(2021)“HongKong’spositioningandprospectasaninternationalfinancialcentre,”Hong KongMonetaryAuthority,https://www.bis.org/review/r211209d.pdf. [12] Macklem,Tiff,(2021)“Thelongandshortofit-abalancedvisionfortheinternationalmonetaryand financialsystem,”BankofCanada,https://www.bis.org/review/r211008c.pdf. [13] Haldane, Andrew G., (2010) “The $100 billion question,” Bank of England, https://www.bis.org/ review/r100406d.pdf. 43

SpeechesReviewedinSection3.4(SummaryStatistics) [1] Carse, David, (2000) “Environmental issues and their implications for financial institutions in Hong Kong,”HongKongMonetaryAuthority,https://www.bis.org/review/r001129c.pdf. [2] Goh,ChokTong(2007)“StayingaheadoftheAsiancurve,”MonetaryAuthorityofSingapore,https: //www.bis.org/review/r071105b.pdf. [3] Boediono (2008) “Macroeconomic impact of climate change - opportunities and challenges,” Bank Indonesia,https://www.bis.org/review/r080901c.pdf. [4] Draghi, Mario, (2009) “The financial crisis - impact and responses,” Bank of Italy, https://www.bis. org/review/r090507b.pdf. [5] Tarantola, Anna Maria, (2010) “Women nurturing sustainable development,” Bank of Italy, https: //www.bis.org/review/r101208g.pdf. [6] Provopoulos, George A., (2011) “The impact of climate change in Greece,” Bank of Greece, https: //www.bis.org/review/r110614a.pdf. [7] Ibrahim,Muhammad,(2012)“Roleandopportunitiesofthefinancialsysteminsupportinggreentechnology,”CentralBankofMalaysia,https://www.bis.org/review/r121004g.pdf. [8] Rambarran,Jwala,(2013)“Generatingmoreinclusiveeconomicgrowththroughscienceandtechnology,”CentralBankofTrinidadandTobago,https://www.bis.org/review/r130806a.pdf. [9] Padmanabhan, G., (2014) “Corporate sustainability a panacea for growth - values, convictions and actions,”ReserveBankofIndia,https://www.bis.org/review/r141021d.pdf. [10] Carney,Mark,(2015)“BreakingtheTragedyoftheHorizon: ClimateChangeandFinancialStability,” BankofEngland,https://www.bis.org/review/r151009a.pdf. [11] Carney,Mark,(2016)“RemarksonthelaunchoftheRecommendationsoftheTaskForceonClimaterelatedFinancialDisclosures,”BankofEngland,https://www.bis.org/review/r161216h.pdf. [12] Stournaras, Yannis, (2017) “Climate change - challenges, risks and opportunities,” Bank of Greece, https://www.bis.org/review/r170726e.pdf. [13] Carney, Mark, (2018) “A transition in thinking and action,” Bank of England, https://www.bis.org/ review/r180420b.pdf. [14] Njoroge, Patrick, (2019) “The importance of green finance guidelines as Nairobi seeks to become a globalhub,”CentralBankofKenya,https://www.bis.org/review/r190321d.pdf. [15] Orr,Adrian,(2020)“Progressingclimateactionbydrivingtransformationalchange,”ReserveBankof NewZealand,https://www.bis.org/review/r201104f.pdf. 44

[16] Kuroda,Haruhiko,(2021)“Addressingclimate-relatedfinancialrisks–fromacentralbank’sperspective,”BankofJapan,https://www.bis.org/review/r210326c.pdf. SpeechesReviewedinSection4.1(ClimateChange,FinancialStability,andMacroprudential Policy) [1] Lagarde, Christine, (2021) “Macroprudential policy in Europe – the future depends on what we do today,”EuropeanCentralBank,https://bis.org/review/r211208i.pdf.“Effectiveclimatepolicieswould alsobenefitthefinancialsector.Butmacroprudentialpolicymakershaveaparttoplay,too, inidentifyingandmitigatingthefinancialaspectsofclimate-relatedrisks.” [2] Lane,PhilipR.,(2019)“ClimatechangeandtheIrishfinancialsystem,”BankofIreland,https://www. bis.org/review/r190206b.pdf.“This calls for ongoing monitoring of climate risks, together with the development of climate-driven scenario analyses and stress tests. By extension, the results of such analysesmaycallfortheappropriatemacroprudentialpoliciestomitigatetheserisks.” [3] Stiroh, Kevin, (2020) “Climate change and risk management in bank supervision,” Federal Reserve BankofNewYork,https://www.bis.org/review/r200309b.pdf.“...supervisorscanfocusonidentifying and managing risks, both microprudential and macroprudential, that emerge along a transition path toamoresustainableeconomy” [4] Elderson, Frank, (2021) “Overcoming the tragedy of the horizon,” European Central Bank, https: //www.bis.org/review/r211118d.pdf. “Getting banks to develop action plans to comply with the expectationsthattheECBsetoutinitsguideonCEriskswasthefirststepingettingthemtoassessand developtheirriskmanagementcapabilities.” [5] Menon, Ravi, (2021) “Being the change we want to see - a sustainable future,” Monetary Authority ofSingapore,https://www.bis.org/review/r210609g.pdf.“Developingacleartaxonomyfortransition activitiesisespeciallyrelevantforAsia.Asianeedstosustaineconomicandsocialdevelopmentwhile shifting to a lower carbon future. A taxonomy that includes both green and transition activities can supportaprogressiveshifttogreatersustainability.” [6] Elderson, Frank, (2018) “Let’s dance,” Netherlands Bank, https://www.bis.org/review/r180904c.pdf. “Weshouldremoveallunnecessaryobstaclestothistransition.Andassistthesectorincreatingcommon definitions and standards. As well as provide the necessary guidance. The development of a sustainabletaxonomyisagoodexampleofthis.” [7] Stournaras, Yannis, (2018) “Climate change - threats, challenges, solutions for Greece,” Bank of Greece, https://www.bis.org/review/r190412g.pdf. “Greece, along with other small countries in the climate-sensitive Mediterranean region, is expected to incur adverse effects from climate change. Acknowledgingthisfact,theBankofGreecehasbeenoneofthefirstcentralbanksworldwidetoactively engageintheissueofclimatechangeandinvestsignificantlyinclimateresearch.” 45

[8] Breeden,Sarah,(2019)“Avoidingthestorm-climatechangeandthefinancialsystem,”BankofEngland,https://www.bis.org/review/r190430k.pdf.“Measuringthesefuturerisksfromclimatechangeto the economy and to the financial system is a complex task. A myriad of possible climate pathways – withdifferentphysicalandtransitioneffects–needtobetranslatedintoeconomicoutcomesandfinancialriskslookingaheadovermanydecades.Tosimplifythatchallenge, weneedtofocusnotonwhat willhappenbutwhatmighthappen.Todothatwecanusescenarioanalysis–datadrivennarratives thathelpanchorourassessmentsofrisk.” [9] de Guindos, Luis, (2019) “Implications of the transition to a low-carbon economy for the euro area financial system,” European Central Bank, https://www.bis.org/review/r191122e.pdf. “The pilot test framework will be macroprudential in nature, and allow us to analyse the system-wide materiality of transition risks for banks’ solvency, along with their lending capacity and the implications for the overalleconomy.” [10] Carney, Mark, (2019) “TCFD: strengthening the foundations of sustainable finance,” Bank of England, https://www.bis.org/review/r191008a.pdf. “Supervisors of financial sector firms will also need toconsiderwhichmetricsaremostusefulfordifferentlevelsofassessment: -Microprudential, toassesshowindividualfirmsaremanagingclimate-relatedrisks-forexampletheimpactofaphysicalor transition risk on a loan book. - Macroprudential, to consider how and whether individual exposures could scale up to systemic risk - Macroeconomic, to help understand how the financial system and economyinteractindifferentclimatetransitionscenarios.” [11] VilleroydeGalhau,François,(2021)“Theroleofcentralbanksinthegreeningoftheeconomy,”Bank ofFrance,https://www.bis.org/review/r210211g.pdf. [12] Mauderer, Sabine, (2019) “Central banks - a crisis manager for the climate?,” Deutsche Bundesbank, https://www.bis.org/review/r191029b.pdf. [13] Bailey,Andrew,(2021)“Tacklingclimateforreal-theroleofcentralbanks,”BankofEngland,https: //www.bis.org/review/r210602a.pdf. [14] Matsen, Egil, (2019) “Climate change, climate risks and Norges Bank,” Norges Bank, https://www. bis.org/review/r191108a.pdf. [15] Weidmann,Jens,(2019)“Climatechangeandcentralbanks,”DeutscheBundesbank,https://www.bis. org/review/r191029a.pdf. [16] DiazdeLeon,Alejandro,(2021)“Remarks-Howcancentralbankersandsupervisorssupportclimate risks and green finance and manage risks?,” Bank of Mexico, https://www.bis.org/review/r210727a. pdf. 46

SpeechesReviewedinSection4.2(OtherTopicsinClimate-relateSpeeches) [1] Stournaras, Yannis, (2017) “Climate change - challenges, risks and opportunities,” Bank of Greece, https://www.bis.org/review/r170726e.pdf. [2] Stournaras, Yannis, (2019) “Climate change - threats, challenges, solutions for Greece,” Bank of Greece,https://www.bis.org/review/r190412g.pdf. [3] Elderson, Frank, (2021) “Forests and finance,” European Central Bank, https://www.bis.org/review/ r211123a.pdf. [4] Elderson, Frank, (2021) “Overcoming the tragedy of the horizon,” European Central Bank, https: //www.bis.org/review/r211118d.pdf. [5] de Guindos, Luis, (2019) “Speaking notes on climate-related risks,” European Central Bank, https: //www.bis.org/review/r191018h.pdf. [6] Jain, Mahesh Kumar, (2021) “Building a more resilient financial system in India through governance improvements,”ReserveBankofIndia,https://www.bis.org/review/r210817a.pdf. [7] Poloz, Stephen S., (2019) “Release of the Financial System Review,” Bank of Canada, https://www. bis.org/review/r190517a.pdf. [8] Breeden,Sarah,(2019)“Avoidingthestorm-climatechangeandthefinancialsystem,”BankofEngland,https://www.bis.org/review/r190430k.pdf. [9] Nor Shamsiah Mohd Yunus (2021) “Remarks - Green Swan 2021 Global Virtual Conference,” Bank NegaraMalaysia,https://www.bis.org/review/r210609a.pdf. [10] Yue, Eddie, (2020) “Managing climate risks in Hong Kong,” Hong Kong Monetary Authority, https: //www.bis.org/review/r201109b.pdf. [11] Chew,JessicaChengLian,(2021)“Openingremarks-WorldBank’sSustainableExchangEDevelopmentSeries(SEEDS),”BankNegaraMalaysia,https://www.bis.org/review/r211109i.pdf. [12] Perrazzelli,Alessandra,(2021)“LaunchoftheG20TechSprintongreenfinanceandsustainableeconomy,”BankofItaly,https://www.bis.org/review/r210510a.pdf. [13] Carney, Mark, (2019) “A platform for innovation,” Bank of England, https://www.bis.org/review/ r190430c.pdf. [14] Santiprabhob,Veerathai,(2019)“Thailand2025-dealingwithmajortrends,”BankofThailand,https: //www.bis.org/review/r191202g.pdf. [15] Gerken, Charlotte, (2021) “Developments in the PRA’s supervision of annuity providers,” Bank of England,https://www.bis.org/review/r210429e.pdf. 47

[16] Gerken, Charlotte, (2021) “The PRA’s role in improving the processes that support insurers’ investment,”BankofEngland,https://www.bis.org/review/r211130s.pdf. [17] de Guindos, Luis, (2020) “The euro area financial sector - opportunities and challenges,” European CentralBank,https://www.bis.org/review/r200206c.pdf. [18] Breman, Anna, (2020) “How the Sveriges Riksbank can contribute to climate policy,” Sveriges Riskbank,https://www.bis.org/review/r200304g.pdf. [19] Kuroda, Haruhiko, (2021) “The Bank of Japan’s strategy on climate change,” Bank of Japan, https: //www.bis.org/review/r210804d.pdf. [20] Lagarde, Christine, (2021) “Press Conference,” European Central Bank, https://www.bis.org/review/ r210709a.pdf. [21] Scicluna,Edward,(2021)“Pricestabilityandbeyond-understandingtheimpactoftheECB’sStrategy Review,”CentralBankofMalta,https://www.bis.org/review/r210722a.pdf. [22] Schnabel, Isabel, (2021) “Societal responsibility and central bank independence,” European Central Bank,https://www.bis.org/review/r210528e.pdf. [23] HernandezdeCos,Pablo,(2021)“TheEuropeanCentralBank’snewmonetarypolicystrategy,”Bank ofSpain,https://www.bis.org/review/r210923d.pdf. [24] Macklem, Tiff, (2021) “Renewal of the Monetary Policy Framework,” Bank of Canada, https://www. bis.org/review/r211229l.pdf. 48