6A 1A JAVIER TUREN Lumpy Forecasters

Lumpy Forecasts
Isaac Baley Javier Turén

UPF, CREI, BSE, CEPR PUC Chile
Banco Central do Brasil – Annual Conference – May, 2024

Motivation
• Expectations and belief formation are central in macroeconomics
1 / 27
Motivation
• Recently, exponential use of survey data to:

I elicit beliefs and information sets
I test theories of expectations (e.g., rational, sticky, behavioral...)
Assumption: Forecasts (what agents disclose) = Beliefs (what they truly expect)
1 / 27
Motivation

• We ask: Does this assumption hold for inflation surveys? If not, what drives the difference?
1 / 27
Motivation

• We ask: Does this assumption hold for inflation surveys? If not, what drives the difference?
• Crucial implications for conducting monetary policy
1 / 27
What do we do
• New facts: Survey of professionals that forecast end-of-year inflation
I Forecasts are lumpy: inaction + large revisions

II Frequency and size of revisions fall with the horizon
III Distance to average (“consensus”) triggers revisions
2 / 27
What do we do

• New model: Bayesian learning + Fixed revision costs + Strategic concerns

I Ss-type model for forecasts, not beliefs
I Structural estimation that matches empirical facts
I Quantify the role of each friction
2 / 27
What do we do

• New model: Bayesian learning + Fixed revision costs + Strategic concerns

I Ss-type model for forecasts, not beliefs
I Structural estimation that matches empirical facts
I Quantify the role of each friction
• Three applications:
1. Heterogeneity across forecaster types (if enough time...)
2. Forecast rationality tests
3. Changes in fundamental volatility (if enough time...)
2 / 27
Contributions
• Forecast lumpiness
Mankiw & Reis (02), Reis (06), Andrade & Le Bihan (13), Gaglianone, Giacomini, Issler & Skreta (22)
? We provide direct evidence with high frequency data
3 / 27
Contributions

• Inaction with learning
Álvarez, Lippi & Paciello (11), Baley & Blanco (19), Baley, Figueiredo & Ulbricht (22), Bonomo, et. al. (23)
? We develop Ss forecasting model and discipline it with microdata
3 / 27
Contributions


• Strategic concerns
Ottaviani & Sørensen (06), Hansen, McMahon & Velasco (14), Broer & Kohlhas (22), Valchev & Gemmi (23)
? We quantify the strength of strategic concerns with microdata
3 / 27
Contributions


• Strategic concerns
Ottaviani & Sørensen (06), Hansen, McMahon & Velasco (14), Broer & Kohlhas (22), Valchev & Gemmi (23)
? We quantify the strength of strategic concerns with microdata

• Rationality tests
Nordhaus (87), Vives (93), Patton & Timmermann (10, 12), Capistrán and López-Moctezuma (13), Coibion
& Gorodnichenko (12, 15), Giacomini, Skreta & Turen (20), Bordalo, Gennaioli, Ma & Schleifer (20)
? We show that lumpiness accounts for behavioral biases

3 / 27
Roadmap
1 Data and fixed-event forecasting
2 Term structure of forecast revisions and errors
3 A model of lumpy forecasts
4 Applications
Data description
• CPI inflation in the US
4 / 27
Data description

• Bloomberg’s ECFC survey of professional forecasters
1 High-frequency: monthly observations (aggregated from weekly)
2 Fixed-event forecasting: end-of-year inflation
3 Consensus: average forecast in real time
4 / 27
Data description

• Around 100 forecasters per year × 14 years ∼ 12,600 obs

I Normal years: 2010–2019 (main analysis) & Turbulent years: 2008–2009, 2020–2021
4 / 27
Data description


• Four types of forecasters:

I banks, financial institutions, consulting companies, universities/research centers
4 / 27
Data description


• Four types of forecasters:

I banks, financial institutions, consulting companies, universities/research centers
• Evidence that inflation expectations (collected from Bloomberg) drives trading behavior of
forecasters, Bahaj et.al. (23)
4 / 27
Fixed-event forecasting
• Fixed event is end-of-year inflation π ∼

P12
= h=1 xh
I Forecast horizon: h ∈ {12, 11, 10, ..., 2, 1}

I Monthly year-on-year inflation: xh = 1 cpih −cpih+12
12 cpih+12
5 / 27
Consensus observed in real time
1
PN
• Consensus (average forecast): Fh = N i=1 fh
i
6 / 27
Example of three forecasters in 2019
2019
2.5
Forecast Inflation (%)
1.5 2
JP Morgan
Standard Chartered
Wells Fargo
Consensus
1
12 11 10 9 8 7 6 5 4 3 2 1
Forecast Horizon
7 / 27
Roadmap
1 Data and fixed event forecasting
4 Applications
(I) Forecast revisions are lumpy
• Revisions: ∆fhi ≡ fhi − fh+1

i Stylized facts - Revisions
I Frequency: Pr[∆fhi 6= 0] = 0.43 (5 revisions/year, duration 1.6 months)

I Size: E |∆fhi | adjust = 0.25

(a) Frequency of revisions (b) Size of non-zero revisions
8 / 27
(II) Error dispersion and hazard rate fall with horizon
• Errors: ehi ≡ π − fhi Stylized facts - Errors
I Mean squared error: E[(ehi )2 ] = 0.24
• Hazard: h(a) = Pr[∆f 6= 0|a], a = forecast age (fixed effects: year + forecaster type)
(a) Mean Squared Error (b) Hazard Rate
9 / 27
(III) Gap to consensus triggers revisions
• Gap to consensus: chi ≡ fh+1

i
− Fh
• Residual prob. of adjustment (fixed effects: year + horizon + forecaster)
(a) Prob. of Upward Adjustment (b) Prob. of Downward Adjustment
10 / 27
Robustness
A1. Weekly data
A2. Rounding
A3. Alternative survey – Consensus Economics
A4. Longer horizon (18 months)
11 / 27
Roadmap
4 Applications
Setup
• N forecasters i choose inflation forecast fhi to minimize sum of monthly losses

" 1 #
(f − π) + r (fh − Fh ) + κ1{f i 6=f i }
X
i 2 i 2
min E
{fhi }1h=12 | h {z } | {z } h h+1
h=12 | {z }
accuracy strategic
stability
12 / 27
Setup

" 1 #
(f − π) + r (fh − Fh ) + κ1{f i 6=f i }
X
i 2 i 2
min E
{fhi }1h=12 | h {z } | {z } h h+1
h=12 | {z }
accuracy strategic
stability
P12
• End-of-year inflation: π = h=1 xh
PN
• Consensus: Fh = N −1 i=1 fh
i
12 / 27
Setup

" 1 #
(f − π) + r (fh − Fh ) + κ1{f i 6=f i }
X
i 2 i 2
min E
{fhi }1h=12 | h {z } | {z } h h+1
h=12 | {z }
accuracy strategic
stability
P12
◦ AR(1) structure: xh = cx + φx xh+1 + εxh , εxh ∼ N (0, σx2 ) (one period delay)
iid
◦ Private signal: xehi = xh + ζhi , idiosyncratic noise ζhi ∼ N (0, σζ2 )
PN
i
◦ AR(1) structure: Fh = cF + φF Fh+1 + εFh , εFh ∼ N (0, σF2 ) (one period delay)
12 / 27
Setup

" 1 #
(f − π) + r (fh − Fh ) + κ1{f i 6=f i }
X
i 2 i 2
min E
{fhi }1h=12 | h {z } | {z } h h+1
h=12 | {z }
accuracy strategic
stability
P12
◦ AR(1) structure: xh = cx + φx xh+1 + εxh , εxh ∼ N (0, σx2 ) (one period delay)
iid
◦ Private signal: xehi = xh + ζhi , idiosyncratic noise ζhi ∼ N (0, σζ2 )
PN
i
◦ AR(1) structure: Fh = cF + φF Fh+1 + εFh , εFh ∼ N (0, σF2 ) (one period delay)
• Information set: Ihi = xehi ∪ Ih = xehi ∪ {xh+1 , xh+2 , . . . , Fh+1 , Fh+2 , . . .}

12 / 27
Setup
• Using the law of iterated expectations, we rewrite using beliefs:

1
+ (fhi − E[π|Ihi ])2 + r (fhi − E[Fh |Ihi ])2 + κE[1{f i 6=f i } |Ihi ]
X
min Σh
{fhi }1h=12 |{z} | {z } | {z } h h+1
h=12 Unforcastable | {z }
accuracy strategic
stability
• Inflation beliefs: π|Ihi ∼ N (π̂hi , Σπh )
• Consensus beliefs: Fh |Ihi ∼ N (F̂h , σF2 )
• Total uncertainty: Σh ≡ Σπh + r σF2
13 / 27
Belief formation
12
1 − φhx

cx cx X
◦ Mean: π̂hi = h + x̂hi − + xj , h = 12, . . . , 1
1 − φx 1 − φx 1 − φx
| {z } j=h+1
| {z }
AR(1) projection
realized, j > h
(σx2 )−1
where x̂hi ≡ E[xh |Ihi ] = α[cx + φx xh+1 ] + (1 − α)x̃hi , w/weight α≡ (σx2 )−1 +(σζ 2 )−1
14 / 27
Belief formation
12
1 − φhx

cx cx X
◦ Mean: π̂hi = h + x̂hi − + xj , h = 12, . . . , 1
1 − φx 1 − φx 1 − φx
| {z } j=h+1
| {z }
AR(1) projection
realized, j > h
(σx2 )−1
◦ Mean: F̂h = cF + φF Fh+1
14 / 27
Belief formation
12
1 − φhx

cx cx X
◦ Mean: π̂hi = h + x̂hi − + xj , h = 12, . . . , 1
1 − φx 1 − φx 1 − φx
| {z } j=h+1
| {z }
AR(1) projection
realized, j > h
(σx2 )−1
◦ Mean: F̂h = cF + φF Fh+1
• Total uncertainty: Σh ≡ Σπh + r σF2
◦ Σπh falls deterministically with h and independent of i

14 / 27
Timeline
Observed at h
State at h
ℎ+1 Horizon ℎ ℎ−1
Private Signal
𝑥!"#
AR(1)
𝑥!$% 𝑥#!& 𝑥!
Actual
monthly
Monthly inflation
𝑥$!& inflation belief
History of monthly inflations
{𝑥#' , 𝑥## , … , 𝑥!"' } AR(1)
AR(1)
𝐹!"# 𝐹0! &&!
π 𝐹!
on
n
End-of-year Actual
ati
tio
Consensus
re g
ga
belief inflation belief consensus
re
g
gg
Ag
A
& Past forecast Revise or not
𝑓!"# State 𝑓!&
15 / 27
Recursive problem and optimal policy
Vh (π̂, F̂ , f ) = min{ VhI (π̂, F̂ , f ), VhA (π̂, F̂ ) }

| {z } | {z }
inaction action
VhI (π̂, F̂ , f ) = Σh + (f − π̂)2 + r (f − F̂ )2 + E[Vh−1 (π̂ 0 , F̂ 0 , f )|I]

n o
∗ ∗ 0 0 ∗
VhA (π̂, F̂ ) = κ + Σh + min
∗
(f − π̂) 2
+ r (f − F̂ ) 2
+ E[Vh−1 (π̂ , F̂ , f )|I]
f
16 / 27
Recursive problem and optimal policy
Vh (π̂, F̂ , f ) = min{ VhI (π̂, F̂ , f ), VhA (π̂, F̂ ) }

| {z } | {z }
inaction action
VhI (π̂, F̂ , f ) = Σh + (f − π̂)2 + r (f − F̂ )2 + E[Vh−1 (π̂ 0 , F̂ 0 , f )|I]

n o
∗ ∗ 0 0 ∗
VhA (π̂, F̂ ) = κ + Σh + min
∗
(f − π̂) 2
+ r (f − F̂ ) 2
+ E[Vh−1 (π̂ , F̂ , f )|I]
f
• Optimal policy is horizon-dependent:
I Inaction region: Rh ≡ {(π̂, F̂ , f ) : VhI (π̂, F̂ , f ) ≥ VhA (π̂, F̂ )}
I New forecast: fh ∗ (π̂, F̂ )

(
0, if f ∈ Rh
I Revisions: ∆fh =
fh ∗ − f if f ∈
/ Rh
16 / 27
Calibration
Calibration
• Externally set
◦ Inflation process (cx , φx , σx2 ) = (0.013, 0.932, 0.0013) Estimation Inflation
◦ Initial disagreement σ02 matches forecast dispersion at h = 12
17 / 27
Calibration
• Externally set
• Calibration
Parameter Value Moment Data Model
κ adjustment cost 0.05 Pr[∆f 6= 0] 0.43 0.42
r strategic concerns 0.41 E[|∆f ||adjust] 0.25 0.22
σζ2 private noise 0.04 hazard slope −0.04 −0.04
◦ Consensus process (cF , φF , σF2 ): belief consistency Figure
17 / 27
Calibration
• Externally set
• Calibration
Parameter Value Moment Data Model
κ adjustment cost 0.05 Pr[∆f 6= 0] 0.43 0.42
r strategic concerns 0.41 E[|∆f ||adjust] 0.25 0.22
σζ2 private noise 0.04 hazard slope −0.04 −0.04
◦ Consensus process (cF , φF , σF2 ): belief consistency Figure
• Data implies:
? stability κ > 0 and strategic complementarity r > 0
? Large weight on private signal α = 0.56
17 / 27
Model in action
Simulation for one forecaster
• Beliefs, extensive and intensive margins, and uncertainty
1
3.7 0.5
0.8
3.6
0.4
0.6
3.5
0.4 0.3
3.4
0.2
3.3 0.2
0
3.2
0.1
-0.2
3.1
12 10 8 6 4 2 12 10 8 6 4 2 12 10 8 6 4 2
18 / 27
Simulation for a cross-section
• N = 100 forecasters, Y = 300 years
2.5 2.5 2.5
2 2 2
1.5 1.5 1.5
1 1 1
12 10 8 6 4 2 12 10 8 6 4 2 12 10 8 6 4 2
19 / 27
Untargeted term structures
(a) Frequency of revisions (b) Size of non-zero revisions (c) Hazard rate
0.7 0.35 0.5
0.6 0.3 0.45
0.4
0.5 0.25
0.35
0.4 0.2
0.3
0.3 0.15
0.25
0.2 0.1
0.2
0.1 0.05 0.15
0 0 0.1
12 10 8 6 4 2 12 10 8 6 4 2 2 4 6 8 10 12
20 / 27
Other moments
Mean squared error

0.6
0.5
Autocorrelations
0.4 Data Model
Forecast errors 0.88 0.70
0.3 Belief errors 0.60
All revisions −0.04 −0.06
0.2 Non-zero revisions −0.11 −0.15
0.1
0
12 10 8 6 4 2
21 / 27
Gap to consensus triggers adjustments
• Gap to consensus: chi ≡ fh+1

i
− Fh
(a) Prob. of Upward Adjustment (b) Prob. of Downward Adjustment
22 / 27
Role of each friction
Role of fixed cost κ and strategic concerns r
• We shut down each friction and reestimate parameters
Data (1) Baseline (2) No fixed (3) No strategic

Parameters revision costs concerns
κ 0.05 0.00 0.05
r 0.41 −0.38 0.00
σζ2 0.04 0.05 0.02
Moments
Pr[∆f 6= 0] 0.43 0.42∗ 1.00 0.59
E[|∆f ||∆f 6= 0] 0.25 0.22∗ 0.25∗ 0.22∗
Hazard Slope −0.04 −0.04∗ N/A −0.04∗
Figure
23 / 27
Roadmap
4 Applications
Rationality tests
Forecasts rationality tests
• Rationality test (predictability of individual forecast errors):
π−fi = γ0h + γ1h (fhi − fh+1

i
) + γ2h (Fh − fh+1
i
) + ih
| {z h} |{z} | {z } | {z }
forecast error bias revision consensus
• When forecasts = beliefs, then:

◦ Rational expectations: γ0h = γ1h = γ2h = 0
◦ Under/overreaction to private info: γ1h > 0, γ1h < 0
Bordalo et.al. (2020) for h = 9
◦ Under/overreaction to public info: γ2h > 0, γ2h < 0
Broer and Kohlhas (2022) and Gemmi and Valchev (2022) for h = 12
• We run test for each horizon

24 / 27
Rationality tests: Lumpy + Strategic
π − fhi = γ0h + γ1h (fhi − fh+1

i
) + γ2h (Fh − fh+1
i
) + ih
• Forecasts over-react to private info (γ1h < 0)

• Forecasts under-react to public info (γ2h > 0)
(a) No bias (b) Overreaction to private info (c) Underreaction to public info
25 / 27
Rationality tests: Frictionless (κ = r = 0)
π − π̂hi = γ0h + γ1h (π̂hi − π̂h+1

i
) + γ2h (Π̂h − π̂h+1
i
) + ih
PN
• Without frictions: fhi = π̂hi and Fh = Π̂h = i=1 π̂hi
• Beliefs are rational (γ0h = γ1h = γ2h = 0) ⇒ Proof
(a) No bias (b) No reaction to private info (c) No reaction to public info
26 / 27
More Applications
1 Heterogeneity
2 Response to higher volatility
Conclusion
• Forecasts are lumpy

I Adjustment frequency and size fall with the horizon
27 / 27
Conclusion

• Bayesian learning + fixed revision cost + strategic concerns

I Explain the microdata
27 / 27
Conclusion


• Lumpy forecasts help us understand...

I Heterogeneous effects by forecasts type
I Forecast rationality tests
I Responses to changes in volatility
27 / 27
Conclusion


• Lumpy forecasts help us understand...

I Heterogeneous effects by forecasts type
I Forecast rationality tests
I Responses to changes in volatility
• Companion project:
I Role of lumpy forecasts for transmission of monetary policy
27 / 27
Backup material
Appendix Index
A. Sample selection
B. Summary statistics
B1. Forecast revisions E. Extensive and Intensive Margins
B2. Forecast revisions, all years
B3. Forecast errors F. Forecast Efficiency
B3. Forecast errors, all years F1. Bordalo, et. al. (2020)
F2. Broer and Kolhas (2022)
C. Robustness F3. Gemmi and Valchev (2023)
C1. Weekly data
C2. Rounding G. Robustness
C3. Consensus Economics Survey G1. Weekly data
C4. Longer horizon G2. Rounding
G3. Consensus Economics Survey
D. Estimation
D1. Inflation process
D2. Consensus process
28 / 27
A. Sample selection Index
• Weekly observations, aggregated at the monthly level
• Keep forecaster with a minimum of 1 revision per year (at a monthly frequency)
• Eliminate extreme revisions
29 / 27
B1. Statistics of Forecast Revisions Back Index
Forecast revisions ∆fhi = fhi − fh+1

i
Average E[∆f ] −0.013

Size E[abs(∆f )|∆f 6= 0] 0.247
Variance Var [∆f ] 0.055
Number of revisions in a year count[∆f 6= 0] 5.059
Months of inaction E[τ ] 1.594
Adjustment frequency Pr[∆f 6= 0] 0.427
Upward Pr[∆f > 0] 0.196
Downward Pr[∆f < 0] 0.231
Spike rate Pr[abs(∆f ) > 0.2] 0.028
Serial correlation (all ∆f ) corr [∆f , ∆f−1 ] −0.043
Serial correlation (non-zero ∆f ) corr [∆f , ∆f−1 ] −0.107
Observations N 9,256
Notes: Bloomberg data for normal years 2010-2019.

Cross-sectional statistics are averaged across years and horizons.
30 / 27
B2. Statistics of Forecast Revisions: Normal vs. Turbulent Back Index
All Turbulent Normal

Average E(∆f ) −0.002 0.028 −0.013
Size E(abs(∆f )|∆f 6= 0) 0.307 0.453 0.247
Variance Var (∆f ) 0.104 0.227 0.054
Frequency Pr(∆f 6= 0) 0.444 0.492 0.427
Upward Pr(∆f ) > 0 0.228 0.318 0.196
Downward Pr(∆f ) < 0 0.216 0.173 0.231
Inaction rate Pr(∆f = 0) 0.556 0.508 0.573
Number of revisions count(∆f 6= 0) 5.204 5.602 5.059
Duration (months) E(τ ) 1.497 1.231 1.594
Spike rate abs(∆f /f ) > 1.2 0.081 0.231 0.028
Positive spikes ∆f /f > 1.2 0.076 0.227 0.023
Negative spikes ∆f /f < −1.2 0.005 0.004 0.005
Serial correlation (all) corr (∆f , ∆f−1 ) −0.035 −0.035 −0.043
Serial correlation (non-zero) corr (∆f , ∆f−1 ) −0.085 −0.078 −0.107
Annual Inflation π 1.896 2.175 1.795
Observations N 12,619 3,363 9,256
31 / 27
B3. Statistics of Forecast Errors Back Index
Forecast errors ehi = π − fhi

Average E[e] −0.055
Average of squares E[e 2 ] 0.235
Size E[abs(e)] 0.305
Positive Pr[e > 0] 0.345
Negative Pr[e < 0] 0.572
Variance Var [e] 0.252
Serial correlation corr [e, e−1 ] 0.877
Observations N 9,256
Notes: Bloomberg data for normal years 2010-2019.

Cross-sectional statistics are averaged across years and
horizons.
32 / 27
B4. Statistics of Forecast Errors: Normal vs. Turbulent Back Index
Table: Summary Statistics of Forecast Errors
All Turbulent Normal

Average E(e) 0.023 0.237 -0.055
Size E(abs(e)) 0.434 0.789 0.305
Positive Pr(e > 0) 0.414 0.606 0.345
Negative Pr(e < 0) 0.510 0.340 0.572
Dispersion σ(e) 0.663 1.105 0.502
Serial correlation corr (e, e−1 ) 0.882 0.878 0.877
Observations N 12,619 3,363 9,256
33 / 27
Bahaj, Czech, Ding and Reis (2023) Back Index
• HIgh correlation between trading activities and inflation expectations in the data.
34 / 27
C1. Weekly data Back Index
• We repeat the analysis using the weekly data directly.
35 / 27
C2. Rounding Back Index
• Assess true lumpiness from rounding
• We repeat the analysis for revisions above threshold ϕ ∈ {0.01, 0.05, 0.1}
36 / 27
C3. Consensus Economics - Revision Frequency Back Index
• Monthly data, 40 forecasters, three decimal points
Rounding No Rounding
.9
.9
.8
.8
.7
Share of Updaters (%)
Share of Updaters (%)

.7
.6
.6
.5
.5
.4
.4
Turbulent Turbulent
.3
Normal Normal
Mean Mean
.2
.3
11 9 7 5 3 1 11 9 7 5 3 1
Forecast horizon Forecast horizon
37 / 27
C3. Consensus Economics - Revisions Size Back Index
• Monthly data, 40 forecasters, three decimal points
Rounding No Rounding
.5
.5
.4
.4
Mean revisions
Mean revisions
.3
.3
.2
.2
Turbulent Turbulent
Normal Normal
.1
Mean Mean
.1
11 9 7 5 3 1 11 9 7 5 3 1
Forecast horizon Forecast horizon
38 / 27
C4. Longer Horizon Back Index
• 18 to 13 months ahead (information about future end-of-year inflation)

• 12 to 1 months ahead (inflation is realized)
(a) Frequency of revisions (b) Size of non-zero revisions
39 / 27
C4. Longer Horizon Back Index
(a) Variance of revisions (b) Mean squared error
40 / 27
D1. Estimation of Inflation Process π back
• We estimate (cx , φx , σx2 ) with a rolling structure.
• Average estimates (across time): ĉx = 0.013, φ̂x = 0.932 and σ̂x2 = 0.0013.
Rolling Estimates AR(1) parameters
φ Constant Standard deviation of Errors
.03
.04
1
.0375
.02
.95
.035
.01
.9
.0325
.85
.03
0
2010m1 2012m1 2014m1 2016m1 2018m1 2020m1 2022m1 2010m1 2012m1 2014m1 2016m1 2018m1 2020m1 2022m1 2010m1 2012m1 2014m1 2016m1 2018m1 2020m1 2022m1
41 / 27
D2. Estimation of Consensus Process F back
• We estimate (cF , φF , σF2 ) with a rolling structure.
• Average estimates (across time): ĉF = 0.154, φ̂F = 0.913 and σ̂F2 = 0.188.
Rolling Estimates AR(1) Consensus Forecasts
φ Constant Standard deviation of Errors
.75
1.1
.6
.65
.4
1
.55
.2
.9
.45
0
.8
−.2
.35
−.4
.25
.7
2010m1 2012m1 2014m1 2016m1 2018m1 2020m1 2022m1 2010m1 2012m1 2014m1 2016m1 2018m1 2020m1 2022m1 2010m1 2012m1 2014m1 2016m1 2018m1 2020m1 2022m1
42 / 27
Relative losses from frictions
1
" #
X
L = E (fhi − π)2
+ r (fhi − Fh ) 2
+ κ1{fhi 6=fh+1
i
}
| {z } | {z }
h=12 accuracy strategic
| {z }
stability
(a) Losses across horizon (b) Relative losses by friction

0.5 100
0.4 80
0.3 60
0.2 40
0.1 20
0 0
12 10 8 6 4 2 12 10 8 6 4 2
43 / 27
Rationality tests à la Bordalo, et.al, 2020
π − fhi = γ0 + γ1 (fhi − fh+1

i
) + ih
(a) No bias (b) Overreaction to private info
44 / 27
Rationality tests à la Broer and Kolhas, 2022
π − fhi = γ0 + γ1 (fhi − fh+1

i
) + γ2 Fh+1 + ih
45 / 27
Rationality tests in lumpy model
46 / 27
Heterogeneity
Heterogeneity across forecaster types
• Cross-sectional moments by forecaster type
Financial Inst. Banks Consulting Universities
Moment Data Model Data Model Data Model Data Model
Pr[∆f 6= 0] 0.45 0.46 0.38 0.37 0.47 0.45 0.34 0.34
Var [∆f ] 0.06 0.06 0.06 0.07 0.08 0.07 0.06 0.07
E[(π − f )2 ] 0.28 0.21 0.22 0.23 0.25 0.23 0.24 0.24
N 5,366 2,567 2,982 1,440
back
47 / 27
Heterogeneity across forecaster types
• Cross-sectional moments by forecaster type
Financial Inst. Banks Consulting Universities
Moment Data Model Data Model Data Model Data Model
Pr[∆f 6= 0] 0.45 0.46 0.38 0.37 0.47 0.45 0.34 0.34
Var [∆f ] 0.06 0.06 0.06 0.07 0.08 0.07 0.06 0.07
E[(π − f )2 ] 0.28 0.21 0.22 0.23 0.25 0.23 0.24 0.24
N 5,366 2,567 2,982 1,440
• Estimated parameters by forecaster type, within group consensus

Parameter Financial Inst. Banks Consulting Universities
κ 1 1.6 1.3 2.2
r 1 2.7 2.7 3.3
σζ2 1 1.5 1.1 1.2
Note: Parameters relative to financial institutions
? Financial institutions are the least stable and strategic. Also less noisy
back
47 / 27
Response to higher volatility
Time Series
• Normal years: 2010-19
• Turbulent years: 2008-09 and 2020-21
(a) Frequency by year × horizon (b) Size by year × horizon
1 1.5
short (h=1) med (h=6) long (h=12) short (h=1) med (h=6) long (h=12)
0.8
1
0.6
0.4
0.5
0.2
0 0
2008 2010 2012 2014 2016 2018 2020 2008 2010 2012 2014 2016 2018 2020
48 / 27
Changes in inflation volatility (in data)
• Turbulent years 2008-09 and 2020-21: σx2 ↑
• Frequency, variance, and errors increase at all horizons
0.6 0.5 4
0.55
0.4
3
0.5
0.45 0.3
2
0.4 0.2
0.35
1
0.1
0.3
0.25 0 0
12 10 8 6 4 2 12 10 8 6 4 2 12 10 8 6 4 2
back 49 / 27
Changes in inflation volatility (in model)
• Model – Increase in inflation volatility: 1.4 × σx2
• Captures qualitative changes in cross-sectional moments
0.7 0.5 4
0.6
0.4
3
0.5
0.3
0.4 2
0.2
0.3
1
0.1
0.2
0.1 0 0
12 10 8 6 4 2 12 10 8 6 4 2 12 10 8 6 4 2
back 50 / 27
Role of fixed cost κ and strategic concerns r
(a) Frequency of revisions (b) Size of non-zero revisions (c) Hazard Rate
1 0.45 1
0.4
0.8 0.8
0.35
0.6 0.3 0.6

0.25
0.4 0.4
0.2
0.15
0.2 0.2
0.1
0 0.05 0
12 10 8 6 4 2 12 10 8 6 4 2 2 4 6 8 10 12
Back 51 / 27
Consistency of perceived vs. actual consensus
50 100 150 200
Back
52 / 27
Extensive margin – Probability of revision
• Frequency of forecast revision increases with both gaps
Frequency and belief gap bhi Frequency and consensus gap chi
1.5
.9
.8
Forecast Change Frequency
1
Quadratic fit
.7
Distribution of Data
Forecast Revisions
.5
.6
Probability
.5
0
.4
−.5
.3
.2
−1
.1
−1.5
0
−1.5 −1 −.5 0 .5 1 1.5 −1.5 −1 −.5 0 .5 1 1.5
Percent deviation from Consensus Forecast Percent deviation from Consensus Forecast
53 / 27
Intensive margin – Size
• Positive gaps call for negative revisions, and vice versa

• Larger gaps call for larger revisions
Revisions and belief gap bhi Revisions and consensus gap chi
1.5
1.5
1
1
Forecast Revisions
Forecast Revisions
.5
.5
0
0
−.5
−.5
−1
−1
−1.5
−1.5
−1.5 −1 −.5 0 .5 1 1.5 −1.5 −1 −.5 0 .5 1 1.5
Percent deviation from AR(1) Implied Forecast Percent deviation from Consensus Forecast
54 / 27
Extensive and intensive margins: Gap to AR(1)
• Gap to AR(1) projection: bhi ≡ fh+1

i − π̂h
0.6
0.4
0.2
-0.2
-0.4
-0.6
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
55 / 27
Extensive margin – Probability of revision
Pr(∆fhi 6= 0) = β0 + βb (bhi ×h) + βc (chi ×h) + εih

? Includes fixed effects (forecaster, year, horizon) + macro controls + forecast age + monetary policy shocks
• Positive consensus gap c lowers prob. of upward revisions, and vice versa
• Belief gap b does not trigger revisions

Prob (Rev > 0) − Gap Consensus Prob (Rev > 0) − Gap AR(1)
.3
.6
Effects on Linear Prediction

.15
.3
0
0
−.15
−.3
−.3
−.6
12 11 10 9 8 7 6 5 4 3 2 1 12 11 10 9 8 7 6 5 4 3 2 1
Horizon Horizon
Unconditional, Shocks Unconditional, News Negative revisions Back

56 / 27
Intensive margin – Revision size, conditional on adjustment
∆fhi = β0 + βf ∆fh+1
i
+ βb (bhi ×h) + βc (chi ×h) + εih
? Includes fixed effects (forecaster, year, horizon) + macro controls + forecast age + monetary policy shocks
• Closing gaps: Positive gaps call for negative revisions

• Consensus gap c becomes less relevant over time
Forecast Revision − Gap Consensus Forecast Rev − Gap AR(1)
.3
.4

.15
0 .2
−.15 0
−.2 −.4
−.3
12 11 10 9 8 7 6 5 4 3 2 1 12 11 10 9 8 7 6 5 4 3 2 1
Horizon Horizon
Unconditional, Shocks Unconditional, News Negative revisions Back

57 / 27
Determinants of Intensive Margin
∆fhi = β0 + βf ∆fh+1
i
+ βb bhi + βc chi + εih
Revih Revih Revih (w /zeros) Revih (w /zeros)
Revih+1 -0.214** -0.069**
(0.072) (0.027)
chi -0.258** -0.112**
(0.090) (0.042)
bhi 0.001 -0.022*
(0.025) (0.011)
+
chi
−
chi
+
bhi
−
bhi
Agehi 0.003 0.002 0.001 -0.001

(0.010) (0.009) (0.006) (0.006)
Constant -0.012 -0.049 0.024 0.010
58 / 27
(0.115) (0.153) (0.044) (0.058)
Determinants of Intensive Margin
+ + + −
∆fhi = γ0 + γf ∆fh+1
i
+ γb+ bhi + γb− bhi + γc+ chi + γc− chi + εih
Revih Revih Revih (w /zeros) Revih (w /zeros)
Revih+1 -0.214** -0.170*** -0.069** -0.051**
(0.072) (0.050) (0.027) (0.017)
chi -0.258** -0.112**
(0.090) (0.042)
bhi 0.001 -0.022*
(0.025) (0.011)
+
chi -0.174*** -0.075***
(0.039) (0.018)
−
chi 0.348* 0.359**
(0.170) (0.145)
+
bhi -0.010 -0.025**
(0.022) (0.009)
−
bhi 0.317 -0.020
(0.197) (0.055)
Agehi 0.003 0.002 0.001 -0.001
(0.010) (0.009) (0.006) (0.006)
Constant -0.012 -0.049 0.024 0.010
(0.115) (0.153) (0.044) (0.058) 59 / 27
Rationality Test - Frictionless κ = r = 0 Back
• The efficiency regression reads:
i
π − π̂h+1 = γ0 + γ1 (π̂hi − π̂h+1
i
) + γ2 (Π̂ih − π̂h+1
i
) + ηhi , E[ηhi ] = 0
= γ0 + γ̃(π̂hi − π̂h+1
i
) + η̃hi , E[η̃hi ] = 0
• Where γ̃ ≡ (γ1 + γ2 ) and η̃hi ≡ γ2 νhi + ηhi . We can show:
1 − φh+1
x 1 − φhx
π̂hi − π̂h+1
i
= [(1 − α)xh+1 + αζh+1
i
]+ α(xh + ζhi )
1 − φx 1 − φx
| {z } | {z }
A B
h−1
1− φhx X 1 − φj x x
bhi
π−π = ((1 − α)εxh − αζhi ) + ε
1 − φx 1 − φx j
| {z } j=1
C
| {z }
D
πhi − π
Cov (b i
bh+1 bhi )
,π − π Cov (A, C ) + Cov (A, D) + Cov (B, C ) + Cov (B, D)
γ̃ = i i
= =0
πh − π
Var (b bh+1 ) Var (π̂hi − π̂h+1
i )
60 / 27
Determinants of extensive margin
Pr(∆fhi 6= 0) = β0 + βb bhi + βc chi + εih

Revision > 0 Revision < 0 Revision > 0 Revision < 0
chi -0.101** 0.086*
(0.038) (0.039)
bhi 0.006 0.032**
(0.009) (0.012)
Agehi 0.024*** 0.025***

(0.005) (0.007)
Constant 0.112** 0.101
(0.049) (0.066)
Observations 7,619 7,619
? Includes fixed effects (forecaster, year, horizon) + macro controls 61 / 27
Determinants of extensive margin (cont...)
+ + + −
Pr(∆fhi 6= 0) = β0 + βb+ bhi + βb− bhi + βc+ chi + βc− chi + εith
Revision > 0 Revision < 0 Revision > 0 Revision < 0
chi -0.101** 0.086*
(0.038) (0.039)
bhi 0.006 0.032**
(0.009) (0.012)
+
chi -0.043** 0.034*
(0.015) (0.017)
−
chi 0.377** -0.493**
(0.142) (0.181)
+
bhi 0.006 0.039***
(0.006) (0.010)
−
bhi 0.046 0.120
(0.098) (0.073)
Agehi 0.024*** 0.025*** 0.022*** 0.026***
(0.005) (0.007) (0.005) (0.007)
Constant 0.112** 0.101 0.082 0.112
(0.049) (0.066) (0.050) (0.068)
Observations 7,619 7,619 7,619 7,619 62 / 27
F. Rationality Tests Back
All years 2010-2019 Turbulent Years
γ0 γ1 F-test γ0 γ1 F-test γ0 γ1 F-test
h=1 -0.0985*** -0.4556*** 0.0003 -0.0734** -0.2891** 0.0011 -0.0751 -0.5023** 0.0185
(.0358) (.1504) (.0302) (.1011) (.0656) (.1918)
h=2 -0.1083*** -0.4459*** 0.0000 -0.0759** -0.0906 0.0373 -0.1063 -0.5233*** 0.0001
( .0362) (.1108) (.03034) ( .1602) (.0653) ( .1207)
h=3 -0.1080*** -0.1671* 0.0033 -0.0715** -0.3099** 0.0006 -0.1182* 0.0348 0.2196
(.0371) (.0921) (.0304) ( .0946) (.0679) ( .1414)
h=4 -0.0987** -0.1705* 0.0062 -0.0768** -0.2315* 0.0142 -0.0611 -0.1396 0.3108
(.0376) (.0996) ( .0305) (.1362) ( .0728) (.1421)
h=5 -0.0886** -0.1873*** 0.0005 -0.0871** -0.3316*** 0.0000 -0.0015 -0.1579 0.2347
(.0385) (.0615) (.0307) (.0636) (.0742) (.0977)
h=6 -0.0802** -0.1204** 0.0194 -0.1036*** -0.2818** 0.0005 0.0713 -0.0801 0.5715
(.0391) (.0589) (.0314) (.1086) (.0783) (.1054)
h=7 -0.0642* -0.0926 0.1229 -0.1220*** -0.2468* 0.0001 0.2083*** -0.1251 0.0280
(.0383) (.09536) ( .0318) ( .1267) ( .0781) ( .1458)
h=8 -0.0547 0.0625 0.1487 -0.1451*** 0.0355 0.0000 0.3291*** 0.1139 0.0001
(.0393) (.05188) (.0320) (.0752) (.0767) (.0784)
h=9 -0.0396 -0.0944 0.1799 -0.1340*** -0.0306 0.0001 0.3672*** 0.0499 0.0000
(.0393) (.0627) (.0336) (.1078) (.0783) (.0853)
h = 10 -0.0473 0.0577 0.4126 -0.1221*** 0.2830*** 0.0000 0.2763*** -0.1033 0.0005
(.0384) (.1373) (.0322) (.0817) (.0732) (.2225)
h = 11 -0.0099 -0.0319 0.9162 -0.0869** 0.0185 0.0198 0.3336*** -0.0570 0.0000
(.0381) (.0845) (.0321) (.0851) (.0706) (.1304)
h = 12 0.0957*** 0.4634** 0.0001 -0.0313 0.2515** 0.0063 0.7777*** 1.3401 0.0000
( .0369) (.1144) (.0378) (.09313) (.0884) (1.1485)
63 / 27
Behavioral Biases
• Expand the baseline belief formation model to entertain two types of behavioral biases:
◦ Overconfidence: σ̃ζ2 = ωσζ2 , with ω ∈ (0, 1].

Parameter Moment Data Model
κ 0.039 adjustment cost Pr[∆f 6= 0] 0.429 0.421
r 0.250 strategic concerns E[abs(∆f )|∆f 6= 0] 0.248 0.336
ση2 0.050 public noise sd[∆fhi ] 0.055 0.074
σζ2 0.200 private noise E[(π − fhi )2 ] 0.235 0.213
π z

Σh Σh
◦ Stubbornness/Bias: π̃hi = σ2 +Σπ µo + 1 − σ2 +Σ π π̂hi , where prior beliefs π|Ioi ∼ N (µo , σo2 )
o h o h
Parameter Moment Data Model

κ 0.039 adjustment cost Pr[∆f 6= 0] 0.429 0.439
r 0.249 strategic concerns E[abs(∆f )|∆f 6= 0] 0.248 0.270
ση2 0.050 public noise sd[∆fhi ] 0.055 0.041
σζ2 0.199 private noise E[(π − fhi )2 ] 0.235 0.283
64 / 27

6A 1A JAVIER TUREN Lumpy Forecasters

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

6A 1A JAVIER TUREN Lumpy Forecasters

Uploaded by

Copyright:

Available Formats

Lumpy Forecasts

Isaac Baley Javier Turén

Banco Central do Brasil – Annual Conference – May, 2024

• Expectations and belief formation are central in macroeconomics

• Expectations and belief formation are central in macroeconomics

• Recently, exponential use of survey data to:

• Expectations and belief formation are central in macroeconomics

• Recently, exponential use of survey data to:

• Expectations and belief formation are central in macroeconomics

• Recently, exponential use of survey data to:

• Crucial implications for conducting monetary policy

• New facts: Survey of professionals that forecast end-of-year inflation

I Forecasts are lumpy: inaction + large revisions

• New facts: Survey of professionals that forecast end-of-year inflation

I Forecasts are lumpy: inaction + large revisions

• New model: Bayesian learning + Fixed revision costs + Strategic concerns

• New facts: Survey of professionals that forecast end-of-year inflation

I Forecasts are lumpy: inaction + large revisions

• New model: Bayesian learning + Fixed revision costs + Strategic concerns

? We provide direct evidence with high frequency data

? We provide direct evidence with high frequency data

? We develop Ss forecasting model and discipline it with microdata

? We provide direct evidence with high frequency data

? We develop Ss forecasting model and discipline it with microdata

? We quantify the strength of strategic concerns with microdata

? We provide direct evidence with high frequency data

? We develop Ss forecasting model and discipline it with microdata

? We quantify the strength of strategic concerns with microdata

? We show that lumpiness accounts for behavioral biases

1 Data and fixed-event forecasting

2 Term structure of forecast revisions and errors

3 A model of lumpy forecasts

• CPI inflation in the US

• CPI inflation in the US

• CPI inflation in the US

• Around 100 forecasters per year × 14 years ∼ 12,600 obs

• CPI inflation in the US

• Around 100 forecasters per year × 14 years ∼ 12,600 obs

• Four types of forecasters:

• CPI inflation in the US

• Around 100 forecasters per year × 14 years ∼ 12,600 obs

• Four types of forecasters:

• Fixed event is end-of-year inflation π ∼

I Forecast horizon: h ∈ {12, 11, 10, ..., 2, 1}

1 Data and fixed event forecasting

2 Term structure of forecast revisions and errors

3 A model of lumpy forecasts

• Revisions: ∆fhi ≡ fhi − fh+1

I Frequency: Pr[∆fhi 6= 0] = 0.43 (5 revisions/year, duration 1.6 months)

(a) Frequency of revisions (b) Size of non-zero revisions

• Errors: ehi ≡ π − fhi Stylized facts - Errors

I Mean squared error: E[(ehi )2 ] = 0.24

(a) Mean Squared Error (b) Hazard Rate

• Gap to consensus: chi ≡ fh+1

(a) Prob. of Upward Adjustment (b) Prob. of Downward Adjustment

A1. Weekly data

A3. Alternative survey – Consensus Economics

A4. Longer horizon (18 months)

1 Data and fixed event forecasting

2 Term structure of forecast revisions and errors

3 A model of lumpy forecasts

• N forecasters i choose inflation forecast fhi to minimize sum of monthly losses

• N forecasters i choose inflation forecast fhi to minimize sum of monthly losses