紐約聯(lián)儲-美聯(lián)儲透明度和政策預(yù)期錯誤：文本分析方法（英）

Fed

Transparency

andPolicy

Expectation

Errors:A

TextAnalysis

ApproachNO.

1081NOVEMBER

2023EricFischer

|

Rebecca

McCaughrin

|

Saketh

Prazad

|MarkVandergonFedTransparency

andPolicyExpectationErrors:ATextAnalysisApproachEricFischer,RebeccaMcCaughrin,SakethPrazad,andMarkVandergonFederalReserveBank

ofNew

YorkStaffReports,no.1081November2023/10.59576/sr.1081AbstractThispaperseekstoestimatethe

extenttowhichmarket-impliedpolicyexpectationscouldbeimprovedwith

furtherinformationdisclosurefromtheFOMC.Using

textanalysismethodsbased

onlargelanguagemodels,we

show

thatif

FOMCmeeting

materialswithfive-yearlaggedreleasedates—likemeetingtranscriptsandTealbooks—wereaccessibleto

thepublicin

realtime,

marketpolicyexpectationscouldsubstantiallyimproveforecastingaccuracy.Most

of

thisimprovementoccursduringeasing

cycles.Forinstance,at

thesix-monthforecasting

horizon,the

marketcould

havepredictedasmuch

as125basispointsof

additionaleasingduringthe2001and2008recessions,equivalenttoa40-50percentreductioninmean

squarederror.Thispotentialforecastingimprovementappearstobe

relatedto

incompleteinformation

abouttheFed’sreactionfunction,particularlywithrespecttofinancialstabilityconcernsin2008.Incontrast,

havingenhancedaccessto

meeting

materialswouldnot

haveimprovedthe

market’spolicyrateforecasting

duringtighteningcycles.JELclassification:

E43,E52,E58,C80Keywords:interestrates,

monetarypolicy,centralbankandpolicy,sentimentanalysis_________________Fischer,

McCaughrin:MarketsGroup,FederalReserveBankof

NewYork(email:eric.fischer@,

mccaughrin@).Prazad:ResearchAnalyst,ResearchGroup,FederalReserveBank

ofNewYork(email:saketh.prazad@).Vandergon:TechnologyGroup,FederalReserveBank

ofNewYork(email:mark.vandergon@).TheauthorsthankMichaelBauer,RyanBush,RichardCrump,MatthewGentzkow,SeungLee,EricOffner,ChiaraScotti,Adam

Shapiro,DerekTang,andseminarparticipantsattheFederalReserveBank

ofNewYork,theFederalReserveBoard,

theFederalReserveBankofAtlantaQuantitativeSkills

Conference,

andtheWesternEconomicAssociationInternational

Conferenceforhelpfulcomments.

TheyalsothankFederalReserveBankofNewYorkMarketsGroupanalystsforassistance

with

sentenceannotationandThiernoDialloforexcellentresearchassistance.Thispaperpresentspreliminaryfindingsand

is

beingdistributedto

economistsand

otherinterestedreaderssolely

to

stimulatediscussionandelicitcomments.Theviews

expressedinthispaperare

thoseoftheauthor(s)anddonotnecessarilyreflectthepositionoftheFederalReserveBankofNewYorkortheFederalReserveSystem.Anyerrorsoromissionsaretheresponsibility

ofthe

author(s).Toview

the

authors’disclosurestatements,visit/research/staff_reports/sr1081.html.1

IntroductionOver

the

last

few

decades,

central

banks

have

made

substantial

changes

to

their

communica-tions

and

transparency

practices

in

an

e?ort

to

improve

market

expectations

of

future

monetarypolicy.

Clear

and

e?ective

communications

help

to

ensure

that

policy

shifts

are

properly

transmit-ted

to

?nancial

conditions

and

the

real

economy.

Recent

work

on

monetary

policy

expectations

hasdocumented

that

while

market

expectations

may

often

be

unbiased

predictors

of

the

future

policyrate,

there

can

be

substantial

ex-post

predictability

of

market

expectation

errors,

particularly

dur-ing

signi?cant

policy

easing

episodes

(Cieslak,

2018),

(Schmeling,

Schrimpf

and

Ste?ensen,

2022),(Bauer

and

Swanson,

2023).

These

expectation

errors

are

believed

to

be

ex-ante

unpredictable

anddriven

by

an

underestimation

of

the

central

bank’s

sensitivity

to

economic

downturns.

In

this

pa-per,

we

argue

that

Fed

transparency

plays

some

role

in

these

expectation

errors,

and

demonstratethat

a

portion

of

the

expectation

errors

are

predictable

ex-ante,

given

enhanced

access

to

informa-tion

from

monetary

policy

meetings.We

de?ne

transparency

in

terms

of

information

loss

between

the

public

and

the

Fed

with

respectto

expectations

of

future

monetary

policy.

In

a

procedural

sense,

central

banks

are

not

fully

trans-parent.

Monetary

policy

meeting

deliberations

are

con?dential.

Transcripts

and

related

brie?ngmaterials

are

only

released

to

the

public

with

a

signi?cant

lag.

To

be

sure,

central

banks

have

legit-imate

reasons

to

maintain

con?dentiality

or

institute

lags

in

providing

access

to

information.

Suchopacity,

which

we

de?ne

as

the

inverse

of

transparency,

may

be

deemed

necessary,

for

instance,to

maintain

independence

and

encourage

objective

and

vigorous

debate

among

policymakers.

Ifinternal

discussions

and

sta?

economic

forecasts

were

to

become

immediately

available

to

the

pub-lic,

market

participants

could

perceive

such

forecasts

as

a

commitment

to

certain

future

actions

bythe

central

bank,

which

could

in

turn

reduce

the

?exibility

of

the

central

bank

in

the

future.In

lieu

of

immediate

access

to

meeting

materials,

the

public

receives

a

wide

range

of

commu-nications

from

the

Fed,

including

statements,

meeting

minutes,

speeches,

interviews,

and

pressconferences.

If

these

communications

can

be

used

to

accurately

predict

the

Fed’s

future

policy

de-cisions,

despite

the

con?dentiality

of

meetings,

then

we

posit

that

there

is

no

information

loss.

Inthis

paper,

we

do

not

attempt

to

weigh

the

trade-o?s

of

increased

transparency

against

potentialimprovements

in

market

policy

forecasts.

Rather,

our

goal

is

to

highlight

potential

costs

of

opacitythat

have

been

previously

unacknowledged

by

the

literature.1We

measure

information

loss

by

comparing

actual

market

policy

expectations

with

our

pre-dictions

of

what

market

expectations

would

have

been

in

a

counterfactual

world

in

which

the

Fedreleases

meeting

transcripts

and

Tealbooks

immediately

instead

of

with

a

?ve

year

lag.1

Thesecounterfactual

predictions

are

generated

using

our

“FedSpeak

model”,

a

forecasting

model

thatpredicts

the

change

in

the

fed

funds

rate

h

months

after

meeting

t,

using

fed

funds

futures

fromthe

day

after

meeting

t,

con?dential

sentiment

and

topic

content

from

meeting

t,

and

con?dentialTealbook

forecasts

from

meeting

t.2

One

day

after

a

meeting,

the

futures

market

should

have

pricedin

all

relevant

information

released

before

the

meeting,

including

meeting

minutes,

speeches,

andmacroeconomic

data

releases.We

produce

time-varying

measures

of

sentiment

and

topic

content

of

meeting

transcripts

andTealbooks.

Topic

modeling

and

sentiment

analysis

are

standard

machine

learning

techniques

thatallow

for

a

detailed

quantitative

assessment

of

how

subjective

content

of

text

changes

over

time.We

measure

sentiment

and

topic

at

the

sentence-level.

A

sentence

can

have

a

sentiment

label

of“Positive”,

“Negative”,

or

“Neutral”.

The

same

sentence

can

also

have

a

topic

label

of

“EconomicGrowth”,

“In?ation”,

“Labor

Markets”,

“Financial

Stability”,

and

“Monetary

Policy”.

A

sentencemay

also

have

no

associated

topic

label.

The

text

analysis

is

conducted

using

supervised

machinelearning

methods

and

large

language

models,

which

we

train

using

a

dataset

of

sentences

fromFOMC

speeches,

statements,

minutes,

transcripts,

and

interviews.

The

topic

and

sentiment

la-bels

for

this

training

dataset

were

determined

by

analysts

from

the

Markets

Group

at

the

FederalReserve

Bank

of

New

York.

We

show

that

our

machine

learning-based

approaches

generate

im-provements

in

accuracy

on

a

held-out

test

set

relative

to

the

lexical

approaches

commonly

used

ineconomics.We

show

that

the

out-of-sample

predictions

of

the

FedSpeak

model

signi?cantly

outperformfed

funds

futures

at

three-,

six-,

and

nine-month

forecasting

horizons,

thus

indicating

opacity.

Theoutperformance

we

?nd

for

the

full

1996

to

2016

period

reveals

an

important

asymmetry

betweentightening

and

easing

cycles.

During

the

2001

and

2008

easing

cycles,

the

gap

between

fed

fundsfutures

and

the

FedSpeak

model’s

predictions

was

as

much

as

125

basis

points

at

the

six

month1Before

2010,

the

two

main

sta?

brie?ng

documents

were

the

Greenbook,

which

summarized

economic

develop-ments,

and

the

Bluebook,

which

summarized

monetary

policy

options.

In

2010,

the

two

documents

were

merged

intothe

Tealbook.

For

the

rest

of

this

paper,

when

we

refer

to

“Tealbook”,

we

refer

to

the

pre-2010

Greenbooks

and

thepost-2010

Tealbooks.

We

do

not

consider

Bluebooks.2In

this

counterfactual

world,

we

assume

that

the

Fed

does

not

adjust

the

contents

of

its

meetings

in

response

togreater

transparency.2horizon,

equivalent

to

a

40-50%

reduction

in

mean

squared

error.

We

?nd

no

such

outperformanceduring

tightening

cycles

and

during

the

zero

lower

bound.

These

results

imply

that

Fed

communi-cations

may

be

less

informative

during

easing

cycles,

to

the

extent

that

meeting

materials

containedpolicy-relevant

information

not

re?ected

in

market

pricing.One

potential

explanation

for

our

results

is

that

although

the

market

takes

into

account

com-munications

released

before

meeting

t,

a

large

portion

of

the

important

information

within

meetingtranscripts

is

released

to

the

public

in

the

speeches

and

minutes

released

after

meeting

t.

If

post-meeting

communications

are

highly

informative,

then

the

opacity

we

identify

in

our

main

resultsmay

be

temporary.

To

account

for

this

possibility,

we

implement

our

forecasting

exercise

with

amodi?ed

FedSpeak

model

that

pairs

market

expectations

one

day

after

meeting

t

with

meeting

ma-terials

from

meeting

t

?1

(rather

than

from

meeting

t).

If

post-meeting

communications

eliminateopacity,

then

information

from

meeting

t

?1

should

no

longer

be

important

since

Fed

communi-cations

released

between

meeting

t

?

1

and

meeting

t

would

have

been

priced

in

by

the

market.Instead,

we

?nd

the

modi?ed

FedSpeak

model

continues

to

outperform

the

market

during

easingcycles,

though

by

a

somewhat

lower

magnitude.

This

suggests

that

opacity

persists

many

weeksafter

a

meeting.Using

an

alternative

metric

of

transparency,

we

verify

the

implication

that

the

outperformanceof

the

FedSpeak

model

re?ects

less

informative

communications

during

easing

cycles.

To

do

this,we

test

whether

a

hypothetical

investor

could

have

used

meeting

minutes

in

real

time

to

predictthe

contents

of

con?dential

meeting

transcripts.

The

investor

estimates

the

historical

relationshipbetween

the

contents

of

meetings

minutes

and

the

contents

of

meeting

transcripts

and

then

ap-plies

that

historical

relationship

to

new

minutes

observations

in

order

to

generate

predictions

oftranscript

content.

We

compare

these

predictions

with

actual

transcript

content

in

order

to

assesstransparency.

We

continue

to

?nd

an

asymmetry

between

tightening

cycles

and

easing

cycles.

Inparticular,

predicted

transcript

sentiment

tends

to

be

more

positive

than

actual

transcript

sentimentduring

easing

cycles,

which

we

posit

may

explain

why

the

market

underestimated

future

rate

cuts.We

o?er

two

main

caveats

to

our

analysis.

First,

our

results

cannot

identify

whether

recentinnovations

in

Fed

communications,

like

the

Summary

of

Economic

Projections

(SEP)

and

pressconferences,

have

improved

Fed

transparency.

Due

to

the

lagged

release

of

meeting

transcripts,we

can

simultaneously

observe

the

SEP,

press

conferences,

and

meeting

transcripts

only

from

20113through

2017.

These

years

were

mostly

characterized

by

the

zero

lower

bound

(ZLB)

and

explicitforward

guidance,

both

of

which

e?ectively

reduce

the

scope

for

market

expectation

errors.

Evalu-ating

the

e?ects

of

the

SEP

and

press

conferences

would

require

more

post-ZLB

meeting

transcriptsto

be

released.

Second,

our

results

do

not

consider

the

possibility

that

FOMC

members

could

re-spond

to

enhanced

transparency

by

altering

the

content

discussed

at

monetary

policy

meetings.These

behavioral

responses

are

emphasized

in

Hansen,

McMahon

and

Prat

(2018),

who

study

theresponse

of

policymakers

to

the

transparency-enhancing

reforms

in

1993.The

asymmetry

we

?nd

between

tightening

and

easing

cycles

is

consistent

with

the

results

ofCieslak

(2018)

and

Schmeling,

Schrimpf

and

Ste?ensen

(2022),

who

?nd

large

excess

returns

inTreasuries

and

fed

funds

futures

during

easing

cycles,

but

not

during

tightening

cycles.

They

at-tribute

the

excess

returns

to

market

expectation

errors,

rather

than

to

changing

risk

premia.

Cieslak(2018)

argues

that

while

these

forecasting

errors

may

be

predictable

ex-post,

they

are

di?cult

topredict

in

real-time,

even

for

policymakers.

Other

papers

?nding

asymmetries

in

monetary

policyexpectations

between

tightening

and

easing

cycles

include

Bauer,

P?ueger

and

Sunderam

(2022),who

?nd

that

professional

forecasters

perceive

policy

decisions

to

be

less

dependent

on

macroe-conomic

conditions

during

easing

cycles

and

therefore

less

predictable.We

also

contribute

to

a

long-standing

literature

on

Fed

transparency.

Most

work

on

trans-parency

and

central

bank

communication

has

studied

the

optimal

level

of

transparency

and

theconditions

under

which

signaling

the

path

of

future

rates

is

welfare-enhancing.3

Less

work

hasbeen

done

on

the

extent

to

which

the

Fed

achieves

transparency

in

practice.

We

?ll

this

gap

byempirically

testing

the

extent

to

which

markets

could

have

historically

improved

their

policy

rateforecasting

with

broader

access

to

information

from

central

bank

meetings.Many

prior

studies

of

central

bank

transparency

have

focused

on

relative

improvements

intransparency.

For

example,

Swanson

(2006)

?nds

that

the

private

sector

has

become

better

at

fore-casting

monetary

policy

since

the

1980s,

likely

due

to

improved

Fed

transparency.

But

these

studiescan

only

measure

relative

changes

in

transparency

and

cannot

tell

us

how

transparent

communica-tions

are

overall.

The

only

studies

that

attempt

to

measure

central

bank

transparency

in

an

absolutesense

rely

on

a

qualitative

lens,

such

as

Eij?nger

and

Geraats

(2006)

and

Dincer

and

Eichengreen(2018),

who

develop

qualitative

indices

of

transparency

over

time

for

central

banks

around

the3For

details,

see

Woodford

(2005),

Cukierman

(2009),

Morris

and

Shin

(2005),

among

others.4world.

While

these

indices

are

useful

for

comparing

a

very

diverse

set

of

institutions,

they

rely

oncoarse

binary

criteria.Our

second

set

of

results

identi?es

the

speci?c

pieces

of

information

within

meeting

transcriptsand

Tealbooks

that

explain

the

large

gaps

between

the

FedSpeak

model’s

predictions

and

marketexpectations

during

easing

cycles.

We

generate

variable

importance

measures

from

the

FedSpeakmodel

to

determine

which

pieces

of

information

within

meeting

deliberations

would

have

beenmost

valuable

for

policy-sensitive

rate

markets

to

have

known

in

real-time.

In

2007-2008,

the

mostimportant

variables

were

?nancial

stability

topic

frequency,

?nancial

stability

sentiment,

economicgrowth

sentiment,

and

the

sentiment

of

the

FOMC’s

leadership.

In

2000-2003,

the

most

importantvariables

were

aggregate

sentiment,

economic

growth

sentiment,

and

the

sentiment

of

leadership.In

both

periods,

the

importance

of

these

variables

often

far

surpassed

the

importance

of

marketexpectations.We

interpret

the

variable

importance

results

through

the

lens

of

a

simple

monetary

policy

rule,where

the

future

fed

funds

rate

is

determined

by

the

committee’s

economic

outlook

and

a

time-varying

reaction

function.

According

to

Nakamura

and

Steinsson

(2018),

Campbell,

Evans

andJustiniano

(2012),

and

Romer

and

Romer

(2000),

the

Fed

has

an

“information

advantage”

aboutthe

economy

and

can

therefore

make

better

forecasts

than

the

public

about

future

economic

condi-tions.

Based

on

this

view,

the

FedSpeak

model

might

outperform

the

market

because

the

meetingmaterials

contain

important

information

about

the

Fed’s

economic

forecasts.The

view

that

the

Fed

has

stronger

forecasting

abilities

than

market

participants

has

been

chal-lenged

by

Hoesch,

Rossi

and

Sekhposyan

(forthcoming)

and

Bauer

and

Swanson

(2023),

amongothers,

who

?nd

that

Tealbook

forecasts

have

not

been

more

accurate

than

private

sector

forecastsin

recent

years.

Consistent

with

these

?ndings,

we

show

that

Tealbook

forecasts

were

relativelyunimportant

within

the

FedSpeak

model

during

easing

cycles

and

pointed

towards

tighter

policyrather

than

looser

policy.

Since

Tealbook

forecasts

are

commonly

assumed

within

the

literature

tore?ect

the

FOMC’s

economic

outlook

(see,

for

example,

Shapiro

and

Wilson

(2019)

and

Bauer

andSwanson

(2023)),

the

low

importance

of

these

forecasts

suggests

that

outlook-related

informationwas

unlikely

to

be

responsible

for

reduced

transparency

during

easing

cycles.As

an

additional

test

of

the

importance

of

outlook-related

information,

we

re-estimate

the

Fed-Speak

model

under

a

hypothetical

scenario

where

the

Fed

has

perfect

foresight

of

future

economic5conditions.

We

modify

the

FedSpeak

model

to

include

macroeconomic

data

releases,

like

nonfarmpayrolls

and

CPI

in?ation,

from

the

month

after

meeting

t.

Even

with

such

extreme

assumptionsabout

the

Fed’s

knowledge

of

the

economy,

the

topic

and

sentiment

variables

retain

at

least

75%of

their

predictive

power.

This

suggests

that

the

text-derived

variables

are

capturing

informationunrelated

to

the

state

of

the

economy

and

are

thus

providing

information

about

policymakers’reactions

to

incoming

economic

news.We

?nd

that

sentiment

and

topic

frequency

variables

are

often

more

important

than

marketexpectations

for

forecasting

future

policy

during

easing

cycles,

providing

a

compelling

a?rmativecase

for

a

reaction

function-based

explanation

of

easing

cycle

opacity.

Sentiment

is

important

be-cause

it

allows

us

to

directly

observe

the

reactions

of

FOMC

members

to

incoming

economic

datarather

than

having

to

infer

their

reaction

based

on

historical

relationships

between

macroeconomicvariables

and

monetary

policy.

Topic

frequency

variables

should

also

be

interpreted

as

related

tothe

reaction

function.

Holding

topic-speci?c

sentiment

constant,

if

the

FOMC

discusses

a

certaintopic

more

often,

then

the

members

may

implicitly

be

weighting

that

topic

more

heavily

in

theirreaction

function.Our

results

emphasizing

incomplete

information

about

policymakers’

reactions

are

consistentwith

a

growing

literature.

Bauer

and

Swanson

(2023)

propose

a

“Fed

response

to

news”

channelfor

explaining

monetary

policy

surprises.

Schmeling,

Schrimpf

and

Ste?ensen

(2022)

?nd

thatmarket

expectation

errors

occur

contemporaneously

with

Taylor

Rule

deviations.

Cieslak

(2018)?nds

that

a

large

portion

of

unexpected

easing

comes

from

unscheduled

FOMC

meetings,

suggest-ing

that

FOMC

members

eased

more

aggressively

than

markets

expected

in

response

to

surpriseeconomic

news.

Bauer,

P?ueger

and

Sunderam

(2022)

?nd

that

professional

forecasters

updatetheir

beliefs

about

the

Fed’s

reaction

function

in

response

to

monetary

policy

shocks,

indicatingimperfect

information

about

the

reaction

function.

We

contribute

to

this

literature

by

providingthe

?rst

quantitative

evidence

of

the

magnitude

of

these

information

frictions.A

reaction

function-based

explanation

of

the

FedSpeak

model’s

performance

has

quite

di?erentimplications

for

Fed

transparency

policy

than

a

forecast

or

outlook-based

explanation.

Informa-tion

about

forecasts

may

be

relatively

easy

to

convey

to

market

participants,

through

instrumentslike

the

Summary

of

Economic

Projections.

But

as

Woodford

(2005)

emphasizes,

conveying

infor-mation

about

the

Fed’s

reaction

function

to

the

public

is

di?cult

because

of

the

large

number

of6di?erent

contingencies

and

scenarios

that

may

arise,

each

demanding

a

di?erent

response

from

thecentral

bank.

The

easing

cycles

that

we

emphasize

may

be

examples

of

infrequent

contingencies,where

conveying

information

about

the

Fed’s

response

in

advance

may

be

di?cult

in

practice.

Ina

similar

vein,

Schmeling,

Schrimpf

and

Ste?ensen

(2022)

argue

that

the

Fed’s

response

to

nega-tive

macroeconomic

shocks

is

inherently

di?cult

for

markets

to

learn

because

of

the

relatively

fewobservations.Finally,

we

contribute

to

a

large

literature

that

uses

textual

analysis

of

FOMC

documents

andcommunications

to

study

FOMC

communications

and

the

transmission

of

communications

to

?-nancial

markets.

Some

important

papers

include

Hansen

and

McMahon

(2016),

Acosta

(2015),Gardner,

Scotti

and

Vega

(2022),

Schmanski

et

al.

(2023),

Chernulich,

Li

and

McGinn

(forthcom-ing),

and

Hansen

and

Kazinnik

(2023).The

rest

of

the

paper

proceeds

as

follows:

in

Section

2,

we

discuss

the

text

analysis

methodsused

to

summarize

the

qualitative

content

of

Fed

documents.

In

Section

3,

we

describe

the

con-struction

of

the

FedSpeak

model

and

show

its

superior

forecasting

ability

relative

to

market

ex-pectations.

In

Section

4,

we

use

variable

importance

measures

to

argue

that

market

forecastingerrors

during

easing

cycles

were

due

to

incomplete

information

about

the

FOMC’s

reaction

func-tion

rather

than

the

economic

outlook.

Section

5

o?ers

concluding

remarks.2

Text

Analysis

MethodologyIn

this

section,

we

describe

the

text

analysis

methods

we

use

for

topic

classi?cation

and

sentimentanalysis.

We

de?ne

topic

as

the

subject

of

a

speaker’s

sentence,

as

it

relates

to

the

economy

ormonetary

policy.

For

example,

in

the

phrase

“market

liquidity

is

worsening”,

the

topic

is

clearlymarket

functioning,

or

?nancial

stability

more

broadly.

In

the

phrase,

“the

CPI

release

shows

mixedsignals”,

the

topic

is

clearly

in?ation.We

de?ne

sentiment

as

the

subjective

attitude

that

a

speaker

conveys

through

their

language.For

example,

the

phrase

“market

liquidity

is

worsening”

conveys

negative

sentiment