Reading between the lines - Using text analysis to estimate the loss function of the ECB
1. Reading between the lines -
Using text analysis to estimate the loss function of the ECB
Maritta Paloviita, Markus Haavio, Pirkka Jalasjoki, Juha Kilponen and Ilona Vänni
Juha Kilponen
24 November 2020
Bank of Finland Webinar: New Challenges to Monetary Policy Strategies
The views expressed do not necessarily reflect the views of the Bank of Finland or the Eurosystem
Kilponen Reading between the lines 24 November 2020 1 / 19
2. Introduction
Usual way to assess policy objectives of the central bank is to estimate
a linear reduced form reaction function (e.g. Taylor type rule) such as
it = ρit−1 + (1 − ρ)(α + βπ(πf
t+j|t − π∗
) + βy ∆yf
t+k|t + Drn
t ) (1)
Left hand side: policy instrument
Right hand side: deviation of inflation (forecast) from the target,
some measure of cyclical position of the economy, real natural rate
Parameters: Convolutions of CB preferences and structural
parameters of the economy, so it is not possible to infer directly what
the CB preferences are
Kilponen Reading between the lines 24 November 2020 2 / 19
3. Introduction
ECB price stability
In 1998, the GC defined price stability as a
’year-on-year increase in the Harmonised Index of Consumer Prices
(HICP) for the euro area of below 2%’.
In 2003, the GC clarified that
‘in the pursuit of price stability it aims to maintain inflation rates
below, but close to, 2% over the medium term’.
Kilponen Reading between the lines 24 November 2020 3 / 19
4. Related literature
Assessing “de facto” target and asymmetries by reaction function
based approach
Hartmann and Smets (2018), Rostagno et al. (2019), Paloviita et al. IJCB
forthcoming
De facto inflation target is well below 2 %
ECB has reacted to deviations of inflation from 2% in asymmetric way
Analysis of monetary policy communication using text mining
techniques
Shapiro and Wilson (2019), Baranowski et al. (2020), Picault and Renault
(2017), Berger et al. (2011), Ehrmann and Talmi (2020), Hansen and
McMahon’s (2016), Armelius et al. (2020), Kawamura et al. (2019), Jones
et al. IJF forthcoming, Kawamura et al. (2019), Bennani et al. (2020),
Fraccaroli et al. (2020)
Evaluation of real time data and central bank forecasting
Fujiwara (2005), Hubert (2014, 2015, 2017), Lyziak and Paloviita (2017,
2018), Potter (2011), Stockton (2012), Fawcett et al. (2015), Iversen et al.
(2016), Kontogeorgos and Lambrias (2019)
Kilponen Reading between the lines 24 November 2020 4 / 19
5. Reading between the lines
What does qualitative communication reveal about the ECB GC’s
preferences?
Apply text mining techniques (language processing) to introductory
statements in order to infer the ECB GC’s objectives directly
Construct net negativity index (tone) which measures the sentiment
(positive, negative) in the introductory statements
We relate this proxy for the loss to real time economic outlook for
inflation and output in order to estimate the loss function
Approach is similar to Shapiro and Wilson (2019) who study Fed
Kilponen Reading between the lines 24 November 2020 5 / 19
6. Net negativity index (tone)
Pre-process the text using standard steps in the text analysis
Use Loughran & McDonald (2011) finance-specific dictionary to
classify words into three classes (negative, positive and neutral) based
on their sentiment.
We modify dictionary to add some central bank specific terms and
expressions
Net negativity index: The difference of the number of negative and
positive words, normalized with the total number of words in the ECB
introductory statement:
Nt =
#Neg − #Pos
#Tot
(2)
Let us look how the tone have evolved in 1999-2019. . .Kilponen Reading between the lines 24 November 2020 6 / 19
7. Tone (net negativity)
Decreasing trend before the financial crisis
After the peak in the middle of the financial crisis, a gradual fall
(increasingly more positive) until the end of 2017
Pessimistic tone in 2019
Kilponen Reading between the lines 24 November 2020 7 / 19
8. Constructing loss function
Assume that CB’s loss is a function of deviation of inflation from the
target such that
Lt = |π − π∗
| (3)
and that we can relate the net negativity index (N) to this loss such
that
N = α + δL (4)
Then we can attempt to estimate a loss function such that
Nt = α + δ |πt − π∗
| + t (5)
d parameter is the preference parameter we are interested in.
We can also treat π∗ as latent variable and estimate it.
Kilponen Reading between the lines 24 November 2020 8 / 19
9. Extension to analyse asymmetry and other objectives
Split the right hand side into two separate segments (piecewise linear
loss function)
Nt = α + δB
˜π−
t (1 − D) + δA
˜π+
t D + εt (6)
add the output gap terms
Nt = α + δB
˜π−
t (1 − D) + δA
˜π+
t D + β1∆ ˜yt + β2∆ ˜yt
2
+ εt (7)
. . . and we can estimate (and test) directly the degree of asymmetry
in the CB’s preferences.
Kilponen Reading between the lines 24 November 2020 9 / 19
10. Extension with Linex (linear exponential) loss function
Piecewise linear loss function does not allow any convexity in the
preferences, so we consider also Linex loss function such that
Nt = α + γ
exp [θ (πt − π∗)] − θ (πt − π∗) − 1
θ2
+ εt (8)
Larger the parameter j, more averse the CB is for inflation rates
above the target (asymmetry is captured by one single parameter)
When parameter θ approaches zero, loss function becomes quadratic
(and symmetric) around some target π∗.
Kilponen Reading between the lines 24 November 2020 10 / 19
11. Right hand side variables
We use short run real time Eurosystem/ECB staff macroeconomic
projections for inflation and output gap (BMPE/MPE)
The average of the inflation nowcast and one-quarter-ahead projected
inflation rate
The average of the output gap nowcast and one-quarter-ahead projected
output gap
Note: The right hand side variables are similar to what are used in
the reaction function estimations.
Kilponen Reading between the lines 24 November 2020 11 / 19
12. Overview of the Results
The ECB has been either more averse to inflation above 2% or “de
facto” inflation aim has been considerably below 2%
Results are robust to
the tone measure (general/inflation specific)
the functional form of the loss function (piecewise linear/Linex)
inclusion/exclusion of output gap terms to the loss function
Kilponen Reading between the lines 24 November 2020 12 / 19
17. Estimated Linex loss function
Kilponen Reading between the lines 24 November 2020 17 / 19
18. Estimated Linex loss function
Kilponen Reading between the lines 24 November 2020 18 / 19
19. Conclusions
Analysis confirms the earlier findings based on reaction function
estimations
Somewhat stronger evidence in favor of asymmetry
Low de facto target and asymmetry:
Stabilizing during the period of high inflationary pressures
Possibly costly when inflationary pressures have been low
Lowered inflation expectations
Increased probability to hit the effective interest rate lower bound
Reduced monetary policy space
Kilponen Reading between the lines 24 November 2020 19 / 19