Post on 14-Oct-2020
transcript
1
000
001
002
003
004
005
006
007
008
009
010
011
012
013
014
015
016
017
018
019
020
021
022
023
024
025
026
027
028
029
030
031
032
033
034
035
036
037
038
039
040
041
042
043
044
045
046
047
048
049
050
051
052
053
054
055
056
057
058
059
060
061
062
063
064
065
066
067
068
069
070
071
072
073
074
075
076
077
078
079
080
081
082
083
084
085
086
087
088
089
090
091
092
093
094
095
096
097
098
099
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
Economics Terminologies in FOMC MinutesCSCI699 Introduction to Information Extraction: project final report
Akira MatsuiStudent ID:5421284665amatsui@usc.edu
Abstract
Professional knowledge is widely used tosolve complex problems in our society.In social science, the vast amount of la-bor has been devoted to public policy is-sues. Many policymakers use their pro-fessional knowledge in social sciences tosolve social problems. However, it isunknown whether these professionals usetheir knowledge explicitly in policy mak-ing situations. Here, we show that pro-fessional knowledge plays a pivotal rolewhen policymakers explain their policiesto the public. We find that the importantterminology semantic change when cen-tral banks change their behavior. We se-lect the technical terminology in the Fed-eral Reserves policy goal and compute theword embedding of these terminologies inthe Federal Open Market Committee min-utes. The changes in the embedding vec-tors affect the federal funds rate and mon-etary base. Our results demonstrate thatthe Federal Reserve Bank (FRB) placesconsiderable value on the concepts in itsmonetary policy goal and the technical ter-minologies convey good indicators of theFRBs behavior. We anticipate this studyto be a starting point to explore how weshould use professional knowledge whenexplaining our findings to the public.
1 Introduction
Many research findings are used to solve real-world problems. Economics is a typical example.Tons of economics research is devoted to solv-ing a wide range of economic problems, from sta-ble marriage problems to optimal monetary pol-icy problems. Even though professionals in eco-
nomics use their economics knowledge in theirminds when they tackle problems in the economy,do they use this knowledge explicitly in discus-sions? If so how these are used?
We hypothesize that policymakers utilize con-cepts from economics when they explain difficultdecisions such as policy rule changes. In otherwords, economics terminology plays a pivotal roleto explain their policy. To study this hypothesis,we use the data from Federal Reserve Bank (FRB),which is the central bank in the United States.
We utilize the dataset made from Federal OpenMarket Committee (FOMC) minutes, and an in-formation extraction technique, to extract relation-ships between economics terminologies and FRBbehavior. We focus on the data from FRB datafor the following reason. First, we can expect thatFOMC members use principal economics con-cepts because all of the FOMC members are eco-nomic experts both from academia and practicalbusiness. Second, the FRB is required to keeptheir transparency by the federal law, which meansthat the FRB always have responsibility to explaintheir policy to the public. The minutes of the com-mittee are open to the public and they have specificgoals of monetary policy that enable us to under-stand what they value. Lastly, we can clearly ob-serve the the FRB’s behavior by studying their pol-icy tools to stimulate the economy such as FederalReserve Banks.
After FOMC meeting, the monetary policy de-cision becomes open to the public. At this point,central banks need to announce their decisionproperly. This communication by the centralbanks is considered one of the important factors toimplement a successful monetary policy [7, 32].Therefore, a central bank deeply values their com-munication not only with the market participantsbut also with the general public. Beside, cen-tral banks have to remain transparent by reporting
2
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
how these decisions are being made because ”bet-ter public understanding makes the policy morecredible and effective1.” The importance of Trans-parency is well known in both theoretical and em-pirical economic studies [10].
Although the specific goals slightly vary amongcentral banks, they aim at keeping economic con-ditions healthy and reducing uncertainty. 2 Theaim of monetary policy, more specifically a centralbank’s goal is considered to promote maximumemployment, stable prices, and economic growth.In recent years, monetary policy is usually a setof policy operations using several tools, such aspurchasing government bonds, assets and settingreserve bank rate [6].
To test this hypothesis, we use FOMC min-utes data from 1993Q1 to 2018Q1, and map theFRB mandate to the terminologies in economicsdictionary to select the important terminologies.In addition, we use the historical data of FederalFunds Rate (FFR) and Monetary Base (MB) to seechanges in the semantics of the terminologies re-flect the policy changes. To extract the informa-tion of the semantic of the terminologies, we useWord2Vec to obtain the word embedding of theterminologies. Then, we study how the changes inthe word embedding affect FRB policy changes.
2 Related Research
In this section, we survey the related research thatuse the documents published by the central banks.Most works use NLP techniques to extract infor-mation from the minutes and study some impor-tant concepts in economics such as communica-tion, uncertainty and transparency. Others are in-terested in extracting features from the minutes toforecast economic outcomes. As discussed in theIntroduction, we are interested more generally inhow professionals utilize the professional knowl-edge to explain their behavior in a policy makingsituation.
2.1 Communication
Most studies about central banks’ communicationsuse NLP to extract features from the documents bythe central bank and find the relationships between
1ECB website, Transparency https://www.ecb.europa.eu/ecb/orga/transparency/html/index.en.html
2To understand why uncertainty matters in monetary pol-icy see Bekaert et al.(2013). [4]
these features and economic indicators such as in-terest rates or monetary policy goals.
Lucca and Trebbi (2009) use sentiment analy-sis with the FOMC minutes and show how thesentiments in the minutes affect Treasury rates.To find the effect of sentiment on Treasury rates,they use Vector Auto Regression (VAR) model,which is the mainstream econometric method inMacroeconomics3. They find that communica-tion4has an impact on Treasury rates and also theminutes tell us the critical implication in mone-tary policy such as rule-based interest rate and theTaylor-rule 5. Tang (2017) studies whether thecentral bank communications affect interest rates[30]. The study presents the methods with theNaive Bayes model and find a strong relationshipthe labor-related contents in the minutes and mon-etary policy response to labor news. Hendry andMadeley (2010) study more specific aspect of therelationship between a central bank communica-tion and Treasury rates by using Latent SemanticAnalysis(LSA) [12]. They study how the commu-nication of the Bank of Canada (BOC) affects re-turns and volatility in short-term as well as long-term interest rate markets6. They find that strongeffects appear in interest rates.
There are studies on the communication dif-ferences in the central banks. Eyup and Odabas(2016) classify the communication strategies [13].They classify the policy statements into severalcategories from the FRB, the European CentralBank (ECB) and Central Bank of the Republic ofTurkey (CBRT). Keida and Takeda (2017) showthe changes in monetary policy communicationamong the two different governors [17]. As ananecdotal evidence, Mr. Kuroda, the BOJ gover-nor, uses the completely different style of speechand interviews than the former governor. Keidaand Takeda (2017) use tf-idf and similarity evalu-ation using vector space model (SVM), and thencompare the differences among two BOJ gover-nors speeches.
3For example, Kuttner (2001) studies the impact of mon-etary policy actions using VAR with a brief explanation ofVAR [14].
4They argues that their semantic analysis extracts thechanges in the monetary policy.
5To understand these rules, see 9.4.2 Monetary policyrules in Kuttner (2018) [15]
6In monetary policy, an attenuating volatility of a interestrate is considered as significant since high volatility meanshigh uncertainty in future economic conditions. For example,see 9.4.1 Monetary policy implementation in Kuttner (2018)[15].
3
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
2.2 Uncertainty and Transparency
Some researchers use NLP to create uncertaintyindex. Andres (2017) use LDA to compose eco-nomic policy uncertainty index from news text [3].Saltzman and Yung (2018) construct uncertaintymeasurements from the FOMC documents cover-ing the long period (from 1970 to 2018) [24]. Theyalso use VAR to find the relation between their un-certainty measurements and economic condition.
Acosta (2015) use LSA to make a transparencymeasurement over 32 years period from FOMCminutes. By using their measurements, this studyshows that the Act’s requirement increased thecentral bank’s transparency [1].
2.2.1 Attitude of committee membersSuda et al. (2018) extract disagreements amongFOMC members and study its effect on asset price[28]. They classify the topics in FOMC mem-bers speeches and define disagreement as the dis-persion of the quantified sentiments. The studyshows that the disagreement among FOMC mem-bers dilutes the announcement of the future mone-tary policy direction 7.
Apel and Grimaldi (2012) use semantic analysisto study the attitudes of the committee members[2]. They use the minutes of the Swedish cen-tral bank and find the attitudes of the committeemembers are useful to predict monetary policy de-cisions.
2.3 Economic prediction
There is a lot of NLP studies that predict macroe-conomic outcomes8. Here, I focus on the stud-ies related to monetary policy. Rohlfs et al.(2016) predict monetary policy targets and Erffec-tive Federal Funds Rate (EFFR)9 by extracting thetopics in FOMC statements with Latent DirichletAllocation (LDA) [23].
Moniz and Jong (2014) predict interest rateexpectations using machine learning. They usethe Bank of England Monetary Policy Commit-tee Minutes (MPCM) from 1997 to 2014 and en-semble methods to predict interest rate expecta-tions in the financial market [20]. Ramachandran
7This type of announcement is described Forward guid-ance, an important tool in today’s monetary policy. To under-stand forward guidance, see 12.5.2 Quantitative easing andforward guidance in Kuttner (2018) [15]
8One of the early works is Wuthrich et al. [33]9To understand Target and Federal Funds Rate, see 9.4
An interest rate-centered view of monetary policy in Kuttner(2018) [15]
and DeRose (2018) classify FOMC meetings in2017 [21]. Stajner et al. (2016) focus on the spe-cific classification problem. They classify specu-lative and non-speculative speech in the transcriptsof FOMC monetary policy meetings [25]. Millerand McCoy (2014) study the content changes inFOMC transcripts and detect the changes betweenthe per and the post-financial crisis[19]. They useLDA to extract topic from the transcripts and clas-sify them. Tan and Lee (2018) study the effectof emphasis on the listeners’ reception in FOMCmembers’ speech [21]. To this aim, they examinerhetoric patterns, hedging, in the transcripts of allFOMC meetings from 1977 to 2008.
While the related researches use the conceptsin a specific discipline or features such as senti-ments, we are focusing on the minimum unit oflanguage, words. Therefore, our study is moregeneral than other related research. The implica-tion of our study can be applied to outside of eco-nomics as well: professional terminologies conveyimportant information to understand their behav-ior.
3 Data
In this section, we provide a brief overview of datacollection and describe the data. We crawl theFOMC minutes from 1967 to 2018 and the eco-nomics terminology dictionary.
The FOMC minutes are the minutes of theFOMC meeting10 to announce the monetary pol-icy decision to the public. As discussed in the Re-lated Research section, the minutes by the centralbanks provide the important information about thecentral bank behavior. The FOMC minute playsthe pivotal role in monetary policy. The FOMCminutes include not only the detail of policy deci-sion making but also the perspectives and the out-look of the economic condition at that time.
To have the FOMC minute datasets, we use theexisting datasets 11 containing the minutes from1967 to 2007. In addition, we write a crawlerto obtain the minutes from 2008 to 2018. Wealso collect the economic terminologies from thewebsite that gives brief explanations of basic eco-nomics concepts, The Economics Classroom12.This website is run by the professor in economics
10https://www.federalreserve.gov/monetarypolicy/fomc_historical_year.htm
11https://stanford.edu/˜rezab/useful/fomc_minutes.html
12https://econclassroom.com/
4
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
Figure 1: Number of words in minutes in eachquarter
After preprocessing the FOMC minute, we de-cided to focus on the minutes in 1993Q1-2018Q1for two reasons. Fist, the minutes before 1993 aremuch shorter than those minutes after 1993. Sec-ond, some minutes before 1993 do not have theterminologies that we select in the Methods sec-tion. The Figure 1 shows the the umber of wordsin minutes over the entire periods. This plot de-scribes the length of minute increases over time.The striking point is between 1992 and 1993 peri-ods. The FRB changed the style of the minute andthey started to publish digital version minutes in1993. To prevent this fact from affecting our anal-ysis, we study the minute from 1993Q1 to 2018Q1in the rest of this paper.
To identfy the relation between the economicterminology usages and the monetary policychanges, we use the historical data of the FFR andthe MB. As discussed in the Related Research,FFR and MB are the important monetary policyoperation tools and the FRB change them to con-duct their policy packages. Therefore, we can as-sume that the changes in these policy tools can beinterpreted as the changes in monetary policy. TheFigure 2 shows log difference of FFR and MB.The most clear change of the monetary policy wasmade after the financial crisis in 2008.
4 Methods
In this section, we will explain the methods toidentify the relationships between changes in theterminology usage by the central bank and in thecentral bank behavior. First, we compute the theterminologies semantic among periods. We utilizeWord2Vec to obtain the word embedding for termsand define the semantic change as the cosine dis-
Figure 2: Log difference over time: Federal FundsRate and Monetary Base
tance between embedding vectors. Second, we useVector Autoregression to identify the central bankbehavioral response to the terminology meaningchanges in the minutes.
4.1 Word2Vec
Word2Vec is a well-known word embeddingmethod introduced by Mikolov et al. (2013) [18].This methods use Skip-gram model to learn syn-tactic and semantic word representations. Ourmethods are highly inspired by Hamilton et al.(2016) [11]. They use word embedding methodsto capture word semantic changes from historicalcorpora and they study the statistical law of theword semantic.
To identify the word semantic transitions acrosstime periods, we construct word embedding foreach time period. We construct word embed-ding from each FOMC minute quarterly corporaby gensim [22]. Then, we aligned the vectorsinto the same coordinate axes. As Hamilton etal. (2016) [11] discuss, this is because low dimen-sion embedding vectors can yield arbitrary orthog-onal transformations. To compare word vectors ofthe same word from different periods, we followHamilton et al. (2016) [11] and use orthogonalProcrustes to align the embedding vectors.
We also select the terminologies to be used inthe analysis by the following procedure. First, wefocus on the words in the economic dictionary.The words in the dictionary can be interpreted asimportant terminologies in economics. Second,we select the words on witch the FRB place im-portance from the dictionary. To this aim, I mapconcepts in the Fed’s mandate 13to the words in the
13https://www.federalreserve.gov/faqs
5
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
dictionary. The Fed’s mandate describes what theFederal Reserve seek to achieve through its mone-tary policy and this can be interpreted as what theyplace importance on. The Fed’s mandate says
The Federal Reserve works to promote astrong U.S. economy. The Congress hasdirected the Fed to conduct the nation’s1) monetary policy to support three spe-cific goals: maximum 2) sustainable em-ployment, 3) stable prices, and moder-ate long-term interest rates.14
We map the three terms in the Fed’s mandateto the economic terminologies: monetary policy,sustainable employment, and stable prices. Thefirst term, ”monetary policy”, is straightforwardbecause we have the same word in the dictionary.However, we need to translate the other two, ”sus-tainable employment” and ”stable prices” as theyare not in the dictionary and high level concepts.We translate ”sustainable employment” as ”labor”because ”sustainable employment” is about the la-bor conditions in the economy. What we can learnfrom ”sustainable employment” is the FRB paysclose attention to ”labor.” For ”stable prices”, Wechoose ”inflation” because inflation is a typical in-dicator of price level and we also assume that theFRB pays close attention to ”inflation”.
We compute how the word embedding vectorsof the selected words vary across the time periods.We use cosine distances between the word vectorsfor the same word,s at time period t as
yst = 1−ws
t ·wst+1
‖wst+1‖‖ws
t ‖
where wst is the word embedding for word s at
time period t. Note that yst is non-negative, i.e.)0 ≤ yst ≤ 2 for any word s and 0 ≤ t. By using thecosine distance, we can focus on how much wordembeddings change between time t and t+ 1.
4.2 Vector AutoregressionTo test our hypothesis that the FRB change theirlanguage when they change their policy behav-ior, we use a statistical model. Since the dataused in this paper is timeseries data, we will useVAR, which is a generalized autoregressive(AR)model[26, 27]15. By using VAR, we can consider
14the numbers are assigned by the author15The lecture note by Zivot provides a good introduction.
[34]
biases typically in time series data such as simul-taneity bias. The VAR model to be estimated hasthe following structure,
yt = c+ Φ1yt−1 + · · ·+ Φpyt−p + εt, εt ∼W.N.(Σ) (1)
yt = (yfft , ymbt , yst )ᵀ, (2)
where c is n×1 interception vector , Φ is n×ncoefficient matrix, and Σ is variance-covariancematrix. The vector, yt, has three endogenous vari-ables at time t. yff is a log difference of Fed-eral Fund Rate, ymb is a log difference of Mone-tary Base. In addition, we incorporate yst into themodel, which is a cosine difference of word em-bedding of terminology s.
VAR allows us to identify the central bank be-havior respond to a terminology semantic changein the minute by using impulse response. Impulseresponse function, IRFij(k), is the function oftime k, which shows the response of variable yito the changes in variable yj at time k.
IRFij(k) =∂yi,t+k
∂εjt(3)
By using the impulse response function,IRFij(k), we can study how a change16 in a vari-able, yj , affects other yi over time periods. Inthis analysis, we focus on how a change in ys af-fects the other variables, yff , ymb. In other words,we focus on how the changes in the word embed-ding of the same words affects the FRB policychanges.
5 Results
5.1 Case StudyBefore answering the research question, we con-duct a case study to examine a typical exam-ple of when the FRB change both their behav-ior and terminologies semantic. We study thedata around the financial crisis in 2008, whichcaused serious economic problems. It is clearthat FRB changed their policy during this time pe-riod [8, 31, 5]. Therefore, we give how the wordembedding changes around these time periods tostudy the case of the FRB changing its behavior.Table 1 shows the top three similar words of ”la-bor” by Word2Vec from annual corpora. We can
16In this paper, one standard deviation change
6
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
see, after the crisis, the most similar word changedfrom ”ebullient” to ”disorderly”. Although thereare some negative words before the crisis, we cansee more negative words after the crisis such as”tip(s)”,”fragile”, ”disorderly” and ”foreclosed”.This case study shows that looking into an impor-tant words provides a clue to understand the FRB’sperspectives and behavior at that time.
5.2 Impulse Response Analysis
I estimate the VAR model and compute the im-pulse responses with a 95% confident interval.Figures 3, 4 and 5 show the impulse response foreach model. The most notable result is for themodel with s = ”labor”, Figure 3. The secondrow in Figure 3 shows a change in the word em-bedding of ”labor” largely decreases FF and thethird row shows that a change in the ”labor” em-bedding increases MB. This means that when theychange the meaning of ”labor”, they also changetheir policy behavior to stimulate the economy.Although the other two columns shows that thesame results, the impulse responses are weak. Thelast column shows the changes in the word embed-ding of ”monetary policy” has a weak effect on thepolicy behavior. These results show that each ter-minology affects the policy behavior to differentdegree.
The other interesting finding is that in all mod-els (Figures 3, 4 and 5) a change in the terminol-ogy word embedding does not have a long effecton itself. The change in a terminology seman-tic not largely affect the semantic of that termi-nology in the following time periods. In the allfirst rows in Figure 3, 4 and 5, the impulse re-sponses converge faster than other rows. This indi-cate that when they change the meaning of a term,they will adhere to that meaning in the followingterms. This is consistent with the previous stud-ies that show that keeping a central bank’s policyconsistent is optimal monetary policy [16, 29, 9].
6 Conclusion
In today’s complex society, more and more peo-ple are obtaining higher educations to solve socialproblems. However, whether their professionalknowledge help them to explain their solution toour society has not been explained well.
Using the dataset from FOMC minute and ainformation extraction technique, we found thatthe policy maker use their knowledge to explain
Figure 3: Impulse responses in the embeddingwords vector distance ”labor” with a 95% confi-dence interval
Figure 4: Impulse responses in the embeddingwords vector distance ”inflation” with a 95% con-fidence interval
7
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
Year 2005 2006 2007 2008 2009 2010
1st ebullient(0.308) exchange-rate(0.31) disorderly(0.272) occur(0.259) tips(0.293) disorderly(0.333)2nd incremental(0.245) opportunity-cost(0.278) emerge(0.265) resulting(0.246) nonresidential(0.273) attention(0.26)3rd disorderly(0.218) financial(0.25) exchange-rate(0.26) housing(0.244) fragile(0.265) foreclosed(0.258)
Table 1: The top three similar words of ”labor” with cosine similarity around the financial crisis in 2008.
Figure 5: Impulse responses in the embeddingwords vector distance ”monetary policy” with a95% confidence interval
their policy direction. When they change theusage of the terminologies, their policy changesaccordingly. The degree of effect of semanticchanges differ among terminologies, which meansthe importance of the terminologies varies as well.Moreover, once they change the semantic of theterminologies, they tend to follow the new seman-tic.
This study describes how the professionals useeconomic concepts in practice and how we shouldunderstand economics in order to understand mon-etary policy management by bridging the gap be-tween different research domains, economics, andcomputer science.
The role of the professional terminologies inthis analysis raises a number of further openquestions. Although there is a clear relation-ship between the usage changes and the behaviorchanges, how they changes the usage when theywant to change their behavior is still an open ques-tion. The other problem is that we do not discussthe source of the FRB’s behavior changes. The al-teration in the FRB policy may come from changes
in economic conditions or may come from theFRB policy strategy changes.
6.1 Note related to the class projects
The FRB also provides other documents such astranscripts of FOMC. Although they are informa-tive, they are not analyzed in the report. Sincemost these documents are published five years af-ter the meeting, they are not for communicationwith the public. However, we provide the data andthe code for them in the repository. 17.
References[1] Miguel Acosta. Fomc responses to
calls for transparency. Available atSSRN:https://ssrn.com/abstract=2647486, 2015.
[2] Mikael Apel and Marianna Grimaldi. The informa-tion content of central bank minutes. Available atSSRN: https://ssrn.com/abstract=2092575, 2012.
[3] Andres Azqueta-Gavaldon. Developing news-based economic policy uncertainty index with un-supervised machine learning. Economics Letters,158:47–50, 2017.
[4] Geert Bekaert, Marie Hoerova, and Marco Lo Duca.Risk, uncertainty and monetary policy. Journal ofMonetary Economics, 60(7):771–788, 2013.
[5] Ben Bernanke. The courage to act: A memoir ofa crisis and its aftermath. WW Norton New York,2015.
[6] Ben S Bernanke and Vincent R Reinhart. Conduct-ing monetary policy at very low short-term inter-est rates. American Economic Review, 94(2):85–90,2004.
[7] Alan S Blinder, Michael Ehrmann, MarcelFratzscher, Jakob De Haan, and David-Jan Jansen.Central bank communication and monetary policy:A survey of theory and evidence. Journal ofEconomic Literature, 46(4):910–45, 2008.
[8] Stephen G Cecchetti. Crisis and responses: the fed-eral reserve in the early stages of the financial cri-sis. Journal of Economic Perspectives, 23(1):51–75,2009.
17For detail, please read README in the repository.
8
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
ACL 2018 Submission ***. Confidential Review Copy. DO NOT DISTRIBUTE.
[9] Richard Clarida, Jordi Gali, and Mark Gertler. Mon-etary policy rules and macroeconomic stability: ev-idence and some theory. The Quarterly journal ofeconomics, 115(1):147–180, 2000.
[10] Petra M Geraats. Central bank transparency. Theeconomic journal, 112(483):F532–F565, 2002.
[11] William L Hamilton, Jure Leskovec, and Dan Ju-rafsky. Diachronic word embeddings reveal sta-tistical laws of semantic change. arXiv preprintarXiv:1605.09096, 2016.
[12] Scott Hendry and Alison Madeley. Text min-ing and the information content of bank ofcanada communications. Available at SSRN:https://ssrn.com/abstract=1722829, 2010.
[13] Eyup Kahveci and Aysun Odabas. Central bankscommunication strategy and content analysis ofmonetary policy statements: The case of fed, ecband cbrt. Procedia-Social and Behavioral Sciences,235:618–629, 2016.
[14] Kenneth N Kuttner. Monetary policy surprises andinterest rates: Evidence from the fed funds futuresmarket. Journal of monetary economics, 47(3):523–544, 2001.
[15] Kenneth N Kuttner. Macroeconomicss.https://econ.williams.edu/faculty-pages/teaching-resources/, 2018.
[16] Susanne Lohmann. Optimal commitment in mone-tary policy: credibility versus flexibility. The Amer-ican Economic Review, 82(1):273–286, 1992.
[17] KEIDA Masayuki, TAKEDA Yosuke, et al. A se-mantic analysis of monetary shamanism: A case ofthe boj’s governor haruhiko kuroda. RIETI Discus-sion Paper Series, 2017.
[18] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg SCorrado, and Jeff Dean. Distributed representationsof words and phrases and their compositionality. InAdvances in neural information processing systems,pages 3111–3119, 2013.
[19] John E Miller and Kathleen F McCoy. Changingfocus of the fomc through the financial crisis. NLPUnshared Task in PoliInformatics. URL: http://www.academia. edu/download/34123959/fomc. pdf,2014.
[20] Andy Moniz and Franciska de Jong. Predicting theimpact of central bank communications on financialmarket investors interest rate expectations. In Eu-ropean Semantic Web Conference, pages 144–155.Springer, 2014.
[21] Harish Gandhi Ramachandran and Dan DeRose Jr.A text analysis of federal reserve meeting minutes.arXiv preprint arXiv:1805.07851, 2018.
[22] Radim Rehurek and Petr Sojka. Software Frame-work for Topic Modelling with Large Corpora. InProceedings of the LREC 2010 Workshop on NewChallenges for NLP Frameworks, pages 45–50, Val-letta, Malta, May 2010. ELRA. http://is.muni.cz/publication/884893/en.
[23] Christopher Rohlfs, Sunandan Chakraborty, andLakshminarayanan Subramanian. The effects ofthe content of fomc communications on us treasuryrates. In Proceedings of the 2016 Conference onEmpirical Methods in Natural Language Process-ing, pages 2096–2102, 2016.
[24] Bennett Saltzman and Julieta Yung. A machinelearning approach to identifying different types ofuncertainty. Economics Letters, 171:58–62, 2018.
[25] Sanja Stajner, Nicole Baerg, Simone PaoloPonzetto, and Heiner Stuckenschmidt. Automaticdetection of speculation in policy statements. Auto-matic detection of speculation in policy statements,2016.
[26] James H Stock and Mark W Watson. Vector au-toregressions. Journal of Economic perspectives,15(4):101–115, 2001.
[27] James H Stock, Mark W Watson, et al. Introduc-tion to econometrics, volume 104. Addison WesleyBoston, 2003.
[28] Shintaro Suda, Ryo Ito, and KiyoshiIzumi. Disagreement among fomc mem-bers and asset prices. Available atSSRN:https://ssrn.com/abstract=3109243, 2018.
[29] Lars EO Svensson. Inflation targeting as a mon-etary policy rule. Journal of monetary economics,43(3):607–654, 1999.
[30] Jenny Tang. Fomc communication and in-terest rate sensitivity to news. Available atSSRN:https://ssrn.com/abstract=3075842, 2017.
[31] John B Taylor. The financial crisis and the pol-icy responses: An empirical analysis of what wentwrong. Technical report, National Bureau of Eco-nomic Research, 2009.
[32] Michael Woodford. Central bank communicationand policy effectiveness. Technical report, NationalBureau of Economic Research, 2005.
[33] Beat Wuthrich, D Permunetilleke, Steven Leung,W Lam, Vincent Cho, and J Zhang. Daily predictionof major stock indices from textual www data. Hkietransactions, 5(3):151–156, 1998.
[34] Eric Zivot. Lecture notes in time serieseconometrics, vector autoregressive mod-els for multivariate time series. https://faculty.washington.edu/ezivot/econ584/notes/varModels.pdf, 2006.