Forecasting Inflation Across Euro Area Countries and Sectors: A Panel VAR Approach

• DOI: http://doi.org/10.1002/for.2444
• Date: 01 July 2017
• Published date: 01 July 2017
• Author: Stéphane Dées,Jochen Güntner

Journal of Forecasting,J. Forecast. 36, 431–453 (2017)

Published online 18 September 2016 in Wiley Online Library (wileyonlinelibrary.com)DOI: 10.1002/for.2444

Forecasting Inflation Across Euro Area Countries and Sectors:

A Panel VAR Approach

STÉPHANE DÉES1AND JOCHEN GÜNTNER2

1

European Central Bank, Frankfurt, Germany

2

Johannes Kepler University Linz, Austria

ABSTRACT

In this paper, we adopt a panel vector autoregressive(PVAR) approach to estimating and forecasting inflation dynamics

in four different sectors—industry, services, construction and agriculture—across the euro area and its four largest

member states: France, Germany, Italy and Spain. By modelling inflation together with real activity, employment

and wages at the sectoral level, we are able to disentangle the role of unit labour costs and profit margins as the

fundamental determinants of price dynamics on the supply side. In out-of-sample forecast comparisons, the PVAR

approach performs well against popular alternatives, especially at a short forecast horizon and relativeto standard VAR

forecasts based on aggregate economy-wide data. Over longer forecast horizons, the accuracy of the PVAR model

tends to decline relative to that of the univariate alternatives, while it remains high relative to the aggregate VAR

forecasts. We show that these findings are driven by the event of the Great Recession. Our qualitative results carry

over to a multi-country extension of the PVAR approach. Copyright © 2016 John Wiley & Sons, Ltd.

KEY WORDS economic fundamentals; inflation forecasting; panel VAR model; sector-level data

INTRODUCTION

While forecasting price inflation is fundamental for private-sector and political decision makers, it has always been

a challenging exercise. Faust and Wright (2013) review the state of the art in inflation forecasting and report ‘an

explosion in the number and variety of methods in recent years’, ranging from traditional time series to more structural

approaches, such as single Phillips curve equations or full-blown dynamic stochastic general equilibrium (DSGE)

models. Use of extensive sets of possible predictors—e.g. in a factor-augmented vector autoregression (FAVAR)

approach proposed by Bernanke et al. (2005)—has also gained popularity in recent years, as have methods based on

financial market indicators, which extract forward-looking information about expected future inflation.

Despite this variety, instability in inflation dynamics has often made it difficult to outperform even simple time

series models, such as a random walk (see, for example, Atkeson and Ohanian 2001). Moreover, the stability

of inflation dynamics during the Great Recession has called into question the usefulness of fundamentals-based

approaches, such as Phillips curve equations, in predicting inflation dynamics (see Ball and Mazumder, 2011;

Bassetto et al., 2013).

Rather than from continuously increasing the degree of complexity of forecasting techniques, forecast accuracy

might therefore benefit from the informational content of disaggregated data. Given that the disaggregated data con-

tain at least as much information as the aggregated time series, increasing the information set on which forecasts are

based could improve the accuracy of out-of-sample forecasts, at least theoretically (see, for example, Rose, 1977;

Tiao and Guttman, 1980). Since inflation and other macroeconomic variables represent contemporaneous aggregates,

it seems plausible that the use of disaggregated data facilitates an increase in forecast accuracy (compare Lütke-

pohl, 1984b). Disaggregated information can also be helpful in identifying common drivers of the aggregated series.

Finally, the forecast errors of disaggregated components might partially cancel out (see, for example, Theil, 1954).

In this paper, we use a fundamentals-based approach to forecasting inflation in the euro area and its four largest

member countries: France, Germany, Italy and Spain. We forecast inflation together with total real value added,

employment and wage inflation in order to account for the effect of labour costs and profit considerations on

price dynamics, using sector-level data for industry, services, construction and agriculture. Our fundamentals-based

‘bottom-up’ approach has several advantages over the standard practice of aggregating component forecasts of the

harmonized index of consumer prices (HICP). First, it allows us to test whether combining sector-level forecasts is

superior to methods based on economy-wide aggregate data.1Second, the use of sector-level data facilitates a more

Correspondence to: Jochen Güntner, Johannes Kepler UniversitatLinz Institut fur Volkswirtschaftslehre Department of Economics Altenberger

Strasse 69 Linz Austria 4040. E-mail: jochen.guentner@jku.at

1Similar bottom-up approaches in a data-rich environment havebeen used to forecast real GDP growth, for example by Hahn and Skudelny 2008)

for the euro area and by Barhoumi et al. 2012) for France.

Copyright © 2016 John Wiley & Sons, Ltd

432 S. Dées and J. Güntner

accurate monitoring of disaggregated prices and promotes our understanding of inflation dynamics in the euro area.

With regard to the role of relative pricesin the build-up of macroeconomic imbalances, similar bottom-up approaches

have received increasing attention since the start of the Great Recession. Furthermore, modelling inflation together

with total real value added, employment and wage inflation allows disentangling the roles of unit labour costs and

profit margins as the main fundamental factors affecting price dynamics from the supply side.2

Figure 1 reveals substantial heterogeneity in the sources of price pressures across countries. In the decade preceding

the crisis, for example, unit labour costs rose strongly in Spain, whereas they remained comparatively subdued in

Germany.Hence forecasting inflation based on supply-side fundamentals, s uch asunit labour costs and profit margins,

is crucial for assessing future developments in the relative competitiveness of different sectors or countries and thus

from the perspective of fiscal and monetary policy makers alike.

Our approach relies on estimating multivariate models that allow for the possibility of dynamic interdependencies

between the variables and sectors of interest as well as for the influence of exogenous driving forces, such as fluc-

tuations in world demand or oil prices. Given the size and complexity of this system, we face two main issues. On

the one hand, estimating separate sector-specific models is relatively parsimonious in terms of the number of coef-

ficients, while it ignores any interdependencies between sectors. On the other hand, a large-scale VAR model of the

entire economy quickly runs into the curse of dimensionality. With ND4sectors, KD4variables, pD2lags

of the endogenous variables and an intercept, for example, we would have to estimate NKpC1D33 parame-

ters per equation, even if we ignored the likely influence of exogenous variables. Given that our sector-level data is

available from 1995:Q1 only, there is little hope of obtaining precise coefficient estimates – especially in conducting

out-of-sample forecast exercises, which further reduces the estimation window.

As a consequence, a suitable shrinkage method is required in order to reduce the parameter space of the model,

while preserving the possibility of interdependencies between sectors. Due to the fact that the number of observation

units (ND4sectors) is small relative to the number of observation periods (T75), a panel vector autoregression

(PVAR) approach seems ideal for the task at hand. In light of limited data availability, we abstain from estimating a

time-varying parameter model, as proposed by Canova and Ciccarelli (2009), even though we might thus miss some

of the variation in dynamic interdependencies between variables, sectors and countries, e.g. due to structural change.3

Forecasting economy-wide variables based on the PVAR approach requires contemporaneous aggregation of the

respective sector-level forecasts. Kohn (1982) and Lütkepohl (1984a) and others show that aggregating forecasts is

generally preferable to forecasting the aggregates directly, if the data-generating process (DGP) is known in terms

of its order and coefficients.4In practice, however, the DGP is rarely known and a trade-off arises with respect to

forecast accuracy. Lütkepohl (1984b) shows that, even if the process is consistently estimated, the information gain

from using the disaggregated time series might be more than offset by higher specification and estimation error of

a less parsimoniously parametrized process, especially at long forecast horizons. As the sample size increases, the

MSPE component due to specification and estimation uncertainty becomes sufficiently small, and the forecast based

on the disaggregated multivariate process is again more accurate than directly forecasting the aggregate. The previous

conclusions are based on asymptotic theory. Given that both asymptotic and small-sample simulation results depend

on the DGP of the multiple time series (see, for example, Hendry and Hubrich, 2011), however, relative forecast

accuracy ultimately remains an empirical question.

Forecasting macroeconomic variables for the euro area entails at least two dimensions of contemporaneous aggre-

gation: countries and subcomponents. Regarding the first dimension, Marcellino et al. (2003) find that forecasting

inflation at the country level and aggregating the forecasts increases the accuracy of euro area forecasts relative to

direct forecasts at the euro area level. Regarding the second dimension, country-specific studies by Bruneau et al.

(2007) and Duarte and Rua (2007) and Moser et al. (2007) all find that aggregating HICP component forecasts yields

more accurate predictions of inflation in France, Austria and Portugal, respectively.

On the contrary, using monthly data for 1992:1–2001:12, Hubrich (2005) finds that direct forecasts of euro area

HICP inflation are often more accurate than reaggregated component forecasts, indicating higher estimation and spec-

ification error at horizons above 6 months. In particular, contemporaneous aggregation seems to increase rather than

reduce bias if unexpected events, such as the surge in unprocessed food and energy prices in 2000, affect components

in the same direction.5

2Based on the same set of data and a vector autoregressive(VAR) approach, Maurin et al. 2011) model profit dynamics in the four largest euro area

countries (Germany, France, Italy and Spain) and the euro area as a whole, considering three sectors (manufacturing, construction and services)

in each economy.

3See, for example, Canova et al. (2012) for recent evidence of variations in European and national real business cycles over time, based on

aggregate macroeconomic time series.

4If the disaggregated time series are approximately uncorrelated and have similar stochastic structures, there is no information gain from using a

multivariate model of the disaggregated variables (see, for example, Lütkepohl, 2006), and the mean squared prediction errors (MSPEs) willbe

identical.

5Bermingham and D’Agostino (2014) argue that the empirical evidence is generally supportive of reaggregating forecasts of individual price

components and that negative results, such as in Hubrich (2005), tend to suffer from either short spans of data and thus largeestimation error or a

relatively small number of disaggregated series.

Copyright © 2016 John Wiley & Sons, Ltd J. Forecast. 36, 431–453 (2017)