nowcasting euro area gdp growth using quantile regression · midas qr our focus is nowcasting...
TRANSCRIPT
Nowcasting Euro Area GDP Growth Using QuantileRegression
Gian Luigi Mazzi and James Mitchell
September 2019
Nowcasting GDP
Offi cial quarterly GDP data are published with a delay
In order to form a view about the current state of the economy, awide range of more timely and higher frequency indicator variables areconsidered
This paper uses an application to explore the utility of quantileregression (QR) methods in producing (density) nowcasts of EuroArea GDP growth from these indicators
Our QR strategy accommodates the mixed-frequency and“ragged-edge”nature of the increasingly big datasets that typicallycharacterise recent nowcasting applications
Quantile Regression (QR) and Density Nowcasts
Because of their flexibility, in modelling the entire density, QR mayprove a useful, and a relatively simple, method of producing reliabledensity nowcasts
Traditionally when seeking to model, explain and indeed nowcast orforecast offi cial statistics using regression methods, focus tends to beon the conditional mean of the variable of interest
QR models the conditional τ-quantile of the dependent variable – forexample the first decile (τ = 0.1) or the ninth decile (τ = 0.9). QRdescribes the relationship at different points in the conditionaldistribution of the variable of interest
In an era of fan charts there is increased awareness of the importanceof modelling and understanding the risks associated with the centralestimate or nowcast/forecast. QR is therefore attractive in modellingthe full distribution
MIDAS QR
Our focus is nowcasting current quarter GDP growth usingwithin-quarter (specifically monthly) known information/data onindicators such as industrial production, retail sales and qualitativebusiness surveysWe follow Ghysels (2014) and consider a Mixed data sampling (orMIDAS) QR:
Qyt (τ|Ωt ) = β0(τ) + β1(τ)[B(L1/m ; θ)xmt
], τ ∈ (0, 1)
where t (t = 1, ...,T ) is the quarter, yt is quarterly GDP growth,
B(L1/m ; θ) = ∑Kk=1 b(k; θ)L
(k−1)/m , where K is the order of the lagpolynomial, Ls/mxmt = x
mt−s−m , x
mt denotes the (in principle, vector
of) indicators, where m = 1, 2, 3 denotes the month in the quarterThe B(L1/m ; θ) function provides a means of parsimoniously modellingthe monthly indicators
The estimated parameters may differ by quantile, τ
Unrestricted MIDAS QR
Following Foroni, Marcellino and Schumacher (2015), we useUnrestricted or UMIDAS:
Qyt (τ|Ωt ) = β0(τ) + β1(τ)x1t + β2(τ)x
2t + β3(τ)x
3t , τ ∈ (0, 1)
i.e. when nowcasting quarterly GDP using monthly indicators, theUMIDAS approach estimates the model at the quarterly frequencywith month 1 of the monthly indicator data forming one variable, x1t ,month 2 another, x2t , and month 3 the third, x
3t
We therefore have 3x as many parameters to estimate compared to amodel specified at the quarterly frequency, that of course would nottherefore exploit the within-quarter data as they accrue
The downside is parameter proliferation
Bayesian UMIDAS QR
We propose a Bayesian approach to estimate the UMIDAS QR
This offers a way to shrink the parameter space in the face ofparameter proliferation and has advantages when density forecastingas it accommodates parameter estimation uncertainties
Yu and Moyeed (2001) show that likelihood-based inference usingindependently distributed asymmetric Laplace densities (ALD) isdirectly related to the traditional quantile regression minimisationproblem
Bayesian UMIDAS QR: Details I
A random variable U follows ALD if its probability density is
fτ(u) = στ(1− τ) exp−σρτ(u)
where σ is the scale parameter and ρτ(u) is the traditional lossfunction minimised in estimation of QR
min(β0,β1:3)
T
∑t=1
ρτ(yt+h − β0 − β1x1t − β2x
2t − β3x
3t )
where
ρτ(u) = u(τ − I (u < 0))= u(τI (u > 0)− (1− τ)I (u < 0))
=|u|+ (2τ − 1)u
2
is the check loss function, and I (.) denotes the indicator function
Bayesian UMIDAS QR: Details II
When τ = 0.5, fτ(u) = (1/4) exp− |u| /2 which is the pdf of astandard symmetric Laplace distribution. For all other τ, fτ(u) isasymmetric
So minimisation is equivalent to maximising a likelihood functionunder the ALD with σ = 1:
L(y|β) = τT (1− τ)T exp
−
T
∑t=1
ρτ(yt − β0 − β1x1t − β2x
2t − β3x
3t )
Priors can be placed on the vector β and Bayesian estimation canproceed
Standard conjugate prior distributions are not available for the QRformulation. So analytical solutions are not available. However,Markov chain Monte Carlo (MCMC) methods can still be used toextract the posterior distributions of the unknown parameters. This,in fact, allows for the use of basically any prior distribution
Bayesian UMIDAS QR with Lasso I
To overcome the curse of dimensionality, a Lasso QR is used. LassoQR involves the following adapted minimisation:
min(β0,β1:3)
T
∑t=1
ρτ(yt − β0 − β1x1t − β2x
2t − β3x
3t ) + λ ‖β‖1
where λ is a nonnegative regularisation parameter. The new secondterm is a penalty. As λ increases, the Lasso continuously shrinks QRcoeffi cients towards zero
Lasso, while originally developed for least squares, is a natural methodto consider in QR too - as it performs both variable selection andregularisation. By forcing the absolute value of the coeffi cients βj tobe less than a fixed value it forces certain coeffi cients to zero
Thereby Lasso identifies a simpler model that does not include all thecoeffi cients.
Bayesian UMIDAS QR with Lasso II
The idea is similar to in ridge regression, where again the inclusion ofextra coeffi cients is penalised. But in ridge regression whilecoeffi cients are shrunk, they are not set to zero as in Lasso
Lasso QR imposes a Laplace prior on the βj
p(βj |σ,λ) =(
λ
2σ
)exp(−λ
∣∣∣βj ∣∣∣ /σ)
which can be rewritten into a mixture of the following hierarchicalpriors integrating out γj
βj |(σ,γj
)∼ N(0, σ2γj ) (1)
γj |σ(τ) ∼ exp(λ2/2) (2)
The Laplace distribution is sharply peaked at zero, compared to theGaussian density, explaining how Lasso sets some coeffi cients to zero
Bayesian UMIDAS QR with Lasso III
Following Li et al. (2010), gamma priors are placed on σ2 and(
λσ
)2,
leading to a Bayesian hierarchical model
Posterior computation then follows Li et al. (2010): sequentiallysample from the posteriors of each unknown parameter conditional onall other parameters using Gibbs
Density Nowcasts
Based on in-sample (t = 1, ...,T ) estimation, quantile nowcasts canbe computed given xT+1
QyT+1(τ|ΩjT+1)
r=βr0(τ)+β
r1(τ)x
1T+1+β
r2(τ)x
2T+1+β
r3(τ)x
3T+1
τ ∈ (0, 1), where βrk (k = 0, ..., 3) denotes the r -th draw from the
posterior parameter distribution, and ΩjT+1 denotes the j-th available
information set or conditioning information
As we will explain, the information set, j , increases as within-quarterindicator data accrue
Recall the quarter T + 1 values of the indicator variables arepublished ahead of the quarter T + 1 values for yt and can thereforebe exploited when nowcasting
The nowcasts, QyT+1(τ|ΩjT+1), can be evaluated when yT+1 is
subsequently published
Density Nowcasts: Construction
We collect together r = 1, ...,R draws from the quantile forecastQyT+1(τ|Ω
jT+1) across τ ∈ [0.05, 0.01, ..., 0.90, 0.95] and then
construct the full posterior density nowcast - using a Gaussian kernelto smooth
Quantiles in the extreme tails, less than 0.05 and greater than 0.95,are not modelled directly
This delivers the density nowcast:
pQR (yT+1 | ΩjiT+1)
From this density nowcast for GDP growth, the user is free to extractany feature of concern to them. This might be the conditional mean.But interest often focuses on tail events
Users of GDP growth forecasts are concerned about the probability of(a one-quarter) recession i.e. of negative GDP growth: the IMF nowpublish GDP at risk measures
Real-time Data
We consider quantitative (“hard”) and qualitative (“soft”) indicatorvariables, with the soft indicators typically published ahead of harddata
And we consider both EA (aggregate) and country-level(disaggregate) indicators
Data are real-time vintages
We focus on nowcasting the EA12 GDP growth aggregate, given dataavailability constraints
Real-time Data
Table: Quarterly and monthly data used
Real-time EA12 Country-level Seas. Adj. Source From
Quarterly GDP Yes Yes All 12 Yes Eurostat 1991q1
Monthly IndicIndustrial Prod Yes Yes All 12 Yes Eurostat 1980m1
ESI n/a Yes exc. GRE, IRE Yes EC 1985m1
Business Clim n/a No Germany Yes Ifo 1991m1
Int rate spread n/a Yes No No ECB 1991m1
Comparison against density combination
To evaluate the relative performance of QR we compare accuracyagainst the nowcast combination methodology of Mazzi et al. (2014).This approach has been previously applied successully to nowcast EAGDP growth
They construct density nowcasts for EA GDP growth by taking densitycombinations across a large number of competing ‘component’models using the same set of indicator variables discussed above
The information set as within-quarter data accrues I
For both QR and the density combination method, yt is related to Ωjt as
follows:
1 j=1. t-30: 30 days before the end of the quarter.
Ω1t =
(xmsoft ,t
2m=1 , xhard ,t−l
p1l=1 , yt−l
p2l=1
)N1 = no of elements of Ω1
t . p1 and p2 denote the number of lags ofthe quarterly variables xk ,t (k = hard) and yt ; and
xmsoft ,t
2m=1
meansthe m=1 (first) and m=2 (second) month’s soft data for quarter t areused
The information set as within-quarter data accrues II2 j=2. t-15: 15 days before the end of the quarter.
Ω2t =
(xmsoft ,t
2m=1 , x
1hard ,t , xhard ,t−l
p1l=1 , yt−l
p2l=1
)N2 = no of elements of Ω2
t . This means Ω2t now includes the first
month of within-quarter hard data, as well as Ω1t
3 j=3. t+0: 0 days after the end of the quarter.
Ω3t =
(xmsoft ,t
3m=1 , x
1hard ,t , xhard ,t−l
p1l=1 , yt−l
p2l=1
)N3 = no of elements of Ω3
t . Ω3t now includes the final month of
within-quarter soft data, as well as Ω2t
4 j=4. t+15: 15 days after the end of the quarter.
Ω4t =
(xmsoft ,t
3m=1 ,
xmhard ,t
2m=1 , xhard ,t−l
p1l=1 , yt−l
p2l=1
)N4 = no of elements of Ω4
t . Ω4t now includes the second month of
within-quarter hard data, as well as Ω3t
Results: out-of-sample
The evaluation period is fixed to match Mazzi et al. (2014):2003q2-2010q4 - to ensure replicability and comparisonThe nowcasts are evaluated by defining the ‘outturn’, yτ, as the first(Flash) GDP growth estimate from EurostatEvaluation using the logarithmic scores, log S , of the density nowcastsfrom QR and the combination approach, i.e. pQR (yT+1 | Ωj
T+1) and
plop(yT+1 | ΩjT+1), against the subsequent GDP growth outturn
yT+1The average log S over the evaluation period running from t = τ tot = τ are:
log SQRj =1
τ − τ ∑t=τ
t=τln pQR (yτ | Ωj
τ)
log S lopj =1
τ − τ ∑t=τ
t=τln plop(yτ | Ωj
τ)
for j = 1, ..., 4
N1 = 214; N2 = 293;N3 = 351;N4 = 430
Average log scores
Table: Average logarithmic score: 2003q2-2010q4
log SQRj log S lopj log SARjt − 30 : j = 1 −1.48 −0.85 −0.84t − 15 : j = 2 −1.44 −0.80 −0.87t + 0 : j = 3 −1.37 −0.79 −0.87t + 15 : j = 4 −1.41 −0.50 −0.84
Average log scores
2004 2005 2006 2007 2008 2009 2010 20112.5
2.0
1.5
1.0
0.5
0.0
0.5
1.0
j=1j=3Outturn (EA GDP Growth)
j=2j=4
GDP-at-Risk: Probability of Negative GDP Growth
2005 2010
0.5
1.0Prob RW_(30_days) QR t30 days
2005 2010
0.5
1.0 RW_(15_days) QR t15 days
2005 2010
0.5
1.0Prob
RW_(0_days) QR t+0 days
2005 2010
0.5
1.0 RW_(15_days) QR t+15 days
2005 2010
2101%
GDP_growth
2005 2010
2101
GDP_growth
PCA
These mixed results, and evidence that QR is overstating uncertainty,suggest the need to impose additional parameter or data parsimony inestimation
One option is to explore the use of different priors, that allow fordifferent types of dependencies between the indicators
Here we take up an alternative but pragmatic strategy, also designedto acknowledge that the Lasso prior may not be working well for oursample of indicators given that there is considerable dependencebetween them
From Nj>200 indicators recursively we select the first 30 principalcomponents (based on the size of the associated eigenvalue) and usethese in Bayesian QR
A large number of principal components is chosen deliberately, so theyexplain most of the variation in the underlying data
We also impose Gaussianity on QR
Average log scores with PCA
Table: Average logarithmic score from QR with PCA used for data shrinkage
log SQR ,PCAj log SQR ,PCA,Nj log SQRj log S lopj log SARjt − 30 : j = 1 -0.83 -0.61 −1.48 −0.85 −0.84t − 15 : j = 2 -0.85 -0.60 −1.44 −0.80 −0.87t + 0 : j = 3 -0.97 -0.80 −1.37 −0.79 −0.87t + 15 : j = 4 -0.90 -0.84 −1.41 −0.50 −0.84
Conclusion
We propose a UMIDAS strategy within a Bayesian QR toaccommodate a large mixed frequency dataset when nowcastingIn a real-time application to Euro Area GDP growth, using over 400mixed frequency indicators, we find that the QR approach does notproduce as accurate density nowcasts as the density combinationapproach unless indicators are orthogonalised and shrunk to a smallernumber prior to QRQR overstates the uncertainties associated with GDP growthHowever, when the nowcasts are formed t-30 or t-15 days before theend of the quarter, QR is better able to detect the ensuing recessionthan the density combination approachBut later in the quarter the density combination approach dominatesFuture work should consider the use of alternative priors, within theproposed Bayesian QR framework, to assess whether these helpimprove the performance of QR when producing density nowcasts
Lasso is in effect a variable selection method. Priors that allow for sets(or groups) of variables to be selected may help