eRese

Keywords:

appystebetcouepe

ment of specicities in terms of risk prole, i.e. heterogeneity of business models, risk exposures andinteraction with the rest of the macro environment.

2013 Elsevier B.V. All rights reserved.

crisisis faciere isderstanwith i

and by sector-specic risk drivers. The rst step in this paper con-sists in generating forward-looking sector-level risk indicatorsbased on Contingent Claims Analysis, a market-based indicator.Then, an analytic framework using the Common Correlated Effects(CCE) estimator from Pesaran (2006) is provided, allowing to studythe determinants and diffusion of risk across sectors and over time,in addition to those coming from other sector-specic determi-

misleading for policy if cross-section dependence and sectoral het-erogeneity is ignored.

A large amount of the emerging literature has focused mainlyon the effects of macroeconomic shocks on banking stability, whilesome work also addresses vulnerabilities in the corporate sector ataggregate level. These studies vary signicantly in terms of theempirical methods applied, the sectors and macroeconomic vari-ables of study, and the assumptions about the direction of shocks,but they all show this strong macro analytic focus. As an example,De Graeve et al. (2008) develop a model of shocks and feedback

Tel.: +351 213 128 139.

Journal of Banking & Finance 37 (2013) 45344555

Contents lists available at

k

w.E-mail address: martin.saldias@bportugal.pt(Financial Stability Board, 2010). The growing literature on nan-cial stability has been urged to expand the focus and to incorporatethe interaction between the nancial system and the rest of theeconomic agents and sectors.

This paper addresses the importance of heterogeneity in termsof risk determinants and risk transmission across corporate sectorsin the euro area. I propose a model where risk in the corporate sec-tor, comprising the nancial sector (banks and insurance compa-nies) and the non-nancial corporate sector (10 supersectors), isdetermined by general economic and nancial markets conditions

ronment have limited relevance on idiosyncratic sectoral riskwhen cross-section dependence is accounted for and the commonelement is ltered out. This result is partly driven by the market-base nature of the risk indicator under analysis and more impor-tantly because sectoral risk responds more signicantly to sector-specic shocks, including proximity-driven risk spill-overs. Resultsalso reveal a high degree of heterogeneity in terms of sensitivityand direction of the effects both from macro-nancial variablesand from sector-specic risk-drivers. These results show that amacro-only focus of the analysis of nancial stability would beMacro-prudential analysisPortfolio credit risk measurementCommon correlated effectsContingent claims analysis

1. Introduction

Due to the nancial and economic2007, research on nancial stabilityembarked on a growing agenda. Thnew and enhanced measures to unworks and to provide policy making0378-4266/$ - see front matter 2013 Elsevier B.V. Ahttp://dx.doi.org/10.1016/j.jbankn.2013.01.026that started in Summerng new challenges anda consensus to developd global nancial net-mproved analytic tools

nants and also from the macroeconomic environment and thenancial markets.

The results show rst that aggregate corporate default risk com-prises a stationary idiosyncratic factor and a non-stationary com-mon element that drives the deviations of the former from along-term equilibrium. Results of the econometric model showthat shocks originated in the macroeconomic and nancial envi-C31C33

nancial stability literature with a contingent claims approach to a sector-based analysis with a less dom-inant macro focus while being compatible with existing stress-testing methodologies in the literature. Adisaggregated analysis of the different corporate and nancial sectors allows for a more detailed assess-A market-based approach to sector risk din the euro area

Martn Saldas Banco de Portugal and Catlica Lisbon Research Unit, Banco de Portugal, Economic and

a r t i c l e i n f o

Article history:Available online 8 February 2013

JEL classication:G01G13

a b s t r a c t

In a panel data frameworkarea, this paper evaluates stress is transferred in andThis approach takes into acdegrees of cross-section d

Journal of Ban

journal homepage: wwll rights reserved.terminants and transmission

arch Department, Av. Almirante Reis 71, 1150-012 Lisbon, Portugal

lied to Portfolio Distance-to-Default series of corporate sectors in the euromic and idiosyncratic determinants of default risk and examines how dis-ween the nancial and corporate sectors since the early days of the euro.nt observed and unobserved common factors and the presence of differentndence in the form of economic proximity. This paper contributes to the

SciVerse ScienceDirect

ing & Finance

elsevier .com/locate / jbf

between the insurance and banking sectors and forecast their de-fault risk in presence of unobserved linkages and other commonshocks using the CCE estimator.1 The risk transmission betweenthe components of the nancial sector (banks and insurance) andadditional non-nancial corporate sectors and within the non-nan-cial sector is not directly tackled, leaving an important source of riskto be further analyzed.

2. Sectoral risk measure for the euro areas nancial and

& Finance 37 (2013) 45344555 4535effects between the real sector (through monetary policy shocks)and the nancial system with no prior assumptions about thedirection of shocks. On the same topic, Castrn et al. (2009) pro-pose a model to assess effects from macroeconomic variables, withno feedback, on credit risk measures of Large and Complex BankingGroups (LCBGs) in the euro area.

Focusing on the interdependence between macro variables andthe non-nancial corporate sector, sberg and Shahnazarian(2009) use an error correction model to assess sensitivity in theaggregate Swedish corporate sector to shocks in variables such asindustrial production, interest rates and consumer prices. Carlinget al. (2007) use a panel data model to assess empirically the im-pact of macroeconomic and rm-specic shocks on default proba-bilities also in the Swedish corporate sector. Bruneau et al. (2008)analyze links in both directions between non-nancial companiesand macroeconomic variables, including nancial shocks, for theFrench economy. Castrn et al. (2010) expand their previous workand study global macro and nancial shocks on the same creditrisk measures of the euro area nancial and corporate sectors sep-arately, using satellite-GVAR models. Castrn and Kavonius (2009)propose a different approach and include in their analysis the link-ages among the rest of economic sectors, e.g. households, govern-ment and rest of the world, using a network of balance sheetexposures and risk-based balance sheets.

Even though the assessment of the effects of general economicconditions on overall corporate risk is highly relevant for nancialstability, understanding also the credit risk relationships withinthe corporate sector with a less macro focus is certainly not negli-gible, yet it has not been extensively studied. As credit risk eventsat individual rm level are linked via sector-specic and generaleconomic conditions (Zhou, 2001), so is risk propagation acrosscorporate sectors through a number of complex channels. In addi-tion, sectoral risk features and responses to common shocks areheterogeneous, hence neglecting this heterogeneity may be mis-leading in terms of overall credit risk management (Hansonet al., 2008), nancial stability analysis and policy decisions.

During the Asian crisis in 1997, an over-leveraged and poorlyprotable corporate sector put the Asian nancial system on theverge of collapse and triggered a deep economic crisis (Pomerlea-no, 2007). The current crisis has highlighted the role of banks inheterogeneous risk transmission to the corporate sector in devel-oped economies either directly through credit constraints or indi-rectly through higher nancing costs, less investmentcounterparts or even second round effects on demand. Castrnand Kavonius (2009) show that the bilateral linkages betweenthe nancial system and the corporate sector in euro area (mea-sured by balance sheets gross exposures) are the most signicativeand take place through both the credit channel and the securitiesmarkets. In addition, the degree of correlation and default trans-mission between non-nancial corporate sectors is high due tocomplementary or similar business lines, e.g. Telecoms and Tech-nology, Utilities and Oil & Gas.

Sectoral risk relationships and their dynamics have previouslybeen analyzed using market-based indicators in Alves (2005) witha VAR approach and in Castrn and Kavonius (2009) using networkanalysis. Their results highlight important cross-dynamics acrosssectors in addition to the impacts viewed as systemic and gener-ated by macroeconomic variables. However, the high degree ofaggregation in these papers is likely to have neglected importantlinkages within the corporate sector and with the nancial system(Castrn and Kavonius, 2009) and may also have ignored sector-specic elements of default risk (Chava and Jarrow, 2004), whichprovide an additional motivation to this study.

M. Saldas / Journal of BankingAdditionally, the dimension limitations of a traditional VARmodel leaves some unobserved effects unaccounted for (Alves,2005). In a recent paper, Bernoth and Pick (2011) model linkagescorporate sectors

The risk measures chosen in this paper to analyze sector-levelstress in the euro area are Portfolio Distance-to-Default (DD) series,namely forward-looking DD series built using aggregated balancesheet information of individual companies by sector and marketinformation of their corresponding indices. DD series make partof the set of risk indicators based on Contingent Claims Analysis(CCA).2 DD series were initially developed and disseminated forcommercial purposes by Moodys KMV using market-based and bal-ance sheet information to assess credit risk in individual companies(Crosbie and Bohn, 2003).

They indicate the number of standard deviations at which themarket value of assets is away from a default barrier dened bya given liabilities structure. A decrease in DD reects a deteriorat-ing risk prole, as a result of the combination of the following fac-tors: lower expected protability, weakening capitalization and/orincreasing asset volatility. Variants of this indicator are increas-ingly used to analyze credit risk of aggregated corporate and macrosectors. Gray and Malone (2008) provide a comprehensive over-view of techniques and applications.

At aggregate corporate sector-level exclusively, DD signals theprobability of generalized distress or joint failure in a given sector

1 The authors use backward-looking Distance-to-Default series computed for a verylarge number of individual institutions and aggregate them into series of weightedaverages and lower quintiles to compute systemic wide forecasts.

2 Contingent Claims Analysis (CCA) is an analytic framework whereby a compre-hensive set of nancial risk indicators is obtained by combining balance sheet andmarket-based information including expected loss, probability of distress, expectedrecovery rate and credit spread over the benchmark risk-free interest rate. It is basedon the Black-Scholes-Merton model of option pricing and has three principles: (1) TheFor these reasons, this paper exploits recent techniques to dealwith panel data in presence of cross-section dependence (CD) andunobserved factors using the Common Correlated Effects (CCE)estimator introduced in Pesaran (2006). This study generates thefollowing contributions to the literature. First, it proposes a meth-odology to build sectoral risk indicators using balance sheet, mar-ket-based and, most notably, option prices information. Theseseries become forward-looking and allow for a wide range ofstress-testing exercises. Then, the paper provides an analyticframework to study risk determinants and transmission at sectorlevel in the euro area by taking into account both the cross-sectiondimension as well as the time series dimension of risk, which hasbeen long neglected in the literature due to lack of a suitable mul-tivariate methodology.

The rest of this paper is structured as follows: Section 2 intro-duces the sector-level risk indicator and the methodology to com-pute it for aggregate sectors. Section 3 describes the sample ofsectors and companies included in the analysis and the propertiesof the sectoral risk indicators. Section 4 describes the analyticframework of risk determinants and diffusion using the CCE esti-mator and other panel data methods applied in the empirical anal-ysis. The results of the former are explained in Sections 5 and 6concludes.economic value of liabilities is derived and equals the economic value of assets, whereliabilities equals debt plus equity; (2) liabilities in the balance sheet have differentpriorities and risk; and (3) the assets distribution follows a stochastic process.

or industry. Despite strong modeling assumptions,3 empirical re-search has shown that aggregate DD dynamics contains informa-tional signals of market valuation of distress and therefore DD is avaluable monitoring tool of the risk prole in the nancial andnon-nancial corporate sectors (Gropp et al., 2009; Vassalou andXing, 2004).

in sample composition. This approach is however likely to be af-fected by spurious variation due to classication changes affectinglarge companies (Alves, 2005) or due to relevant corporate events,including M& A, spin-offs or delistings.

This paper choice for sector classication is the Industry Classi-cation Benchmark (ICB) at Supersector level.7 ICB is a widely usedand comprehensive company classication system jointly developedby FTSE & Dow Jones Indexes to aggregate traded companies accord-ing to their main sources of revenue, as reported in audited accountsand directors reports. The grouping at Supersector level is wide en-ough to ensure a large degree of homogeneity in business models

Index options quotes and (2) stock market capitalization of the their

4536 M. Saldas / Journal of Banking & Finance 37 (2013) 45344555Since the same principles of CCA can be applied to aggregationof rms, the analysis of an entire corporate sector becomes theanalysis of a portfolio of companies. In empirical terms, individualcompany information needs to be aggregated together into a sin-gle, tractable and highly representative indicator by corporate sec-tor, where its composition must be clearly dened.

As for aggregation, most papers in the literature compute themedian or either the weighted or unweighed average of DD orEDF series4 for a large and changing sample of companies. Thismethodology produces an indicator that highlights the intensityand overall risk outlook in the sector but may overemphasize thelarge players or may partially neglect interdependencies amongportfolio constituents (Alves, 2005). Examples of this approach arefound in Alves (2005), Bernoth and Pick (2011), Carlson et al.(2008), Castrn and Kavonius (2009), Castrn et al. (2009, 2010),and sberg and Shahnazarian (2009).

By contrast, this papers aggregation approach are Portfolio DDseries, following research on nancial systemic risk in Cihk et al.(2009), De Nicol et al. (2007), Mhleisen et al. (2006), Echeverraet al. (2009), De Nicol et al. (2005), and Saldas (2012). This meth-odology treats the set of companies by sector as a single entity, itaggregates balance sheet and market data and incorporates the as-sumed portfolio volatility before computing the DD series. Appen-dix A contains a complete explanation of Portfolio DDs derivationand data requirements.

The Portfolio DD series obtained using this methodology haveseveral interesting features. Portfolio DD enhances the informa-tional properties of average DD series, since it does not only cap-ture company size but also interdependencies among theportfolio constituents. It may be considered as the upper boundof joint distance to distress (the lower bound in terms of jointprobabilities of distress) in normal times (De Nicol et al., 2007)but it tends to converge with the average DD in times of stress,when equity market volatility and correlation are higher. This fea-ture illustrates quick reaction of the indicator to market events andshows the generalized increase in returns covariance in a sectorduring distress times, even if fundamentals of portfolio constitu-ents may be solid. Aggregated company fundamentals embeddedin the indicator are informative of longer-term trends of sectoralrisk (see Saldas (2012) for an extensive discussion).

Finally, since aggregation of company information is conductedbefore computing the risk indicator, calibration of Portfolio DD alsoallows to add more easily forward-looking properties from optionmarkets via option implied volatilities from EURO STOXX indices,which also circumvent assumptions about constituents returnscorrelations. Portfolio DD acquires more responsiveness to earlysigns of sector-level distress and hence serves to stress scenarios.5

The second empirical issue deals with the sector classicationand the selection of constituent companies in the Portfolio DD. Re-search based on median and average DD series tackles only the for-mer issue6 and then picks the largest sample available with breaks

3 These assumptions are concerned mainly with those inherent in the Merton-based model (e.g. lognormal distribution of assets, constant asset volatility, etc.) andalso the liability structure.

4 Expected Default Frequency (EDF) is a credit measure based on CCA and adaptedby Moodys KMV to reect actual default distributions.

5 This paper does not include average DD series computed using option price

information as described in Saldas (2012) since there are not enough single equityoptions traded for all companies in this large sample.

6 In general, they adopt systems linked to those used for National Accounting.corresponding Supersector STOXX Indices at Deutsche Boerse.Table 1 briey lists them and provides relevant market information.

The dataset comprises monthly observations between Decem-ber 2001 and October 2009 (95 observations per Supersector). Thisperiod is characterized by an increasing degree of integration inEuropean nancial markets due to the introduction of the euroand a greater europeanization of corporate activities (Vron,2006). Recent trends and ndings suggest that equity markets inte-gration has lead to a reduction of home bias and to an increase ofsector-based equity allocation strategies at the expense of country-based strategies (European Central Bank, 2010; Cappiello et al.,2010). These developments give support to the aggregation ofcompany risk indicators into portfolios for the euro area as a wholeand they provide a rst tentative and equity-driven explanation tostrong comovement of the series over time, as can be seen in Fig. 1.

7 Even though Industries, Supersectors and Sectors are clearly differentiated as ICBCategories, the use of these terms in this paper will uniquely refer to Supersectors.

8 The remaining seven sectors are Construction & Materials, Travel & Leisure,Personal & Household Goods, Financial Services, Retail, Basic Resources and RealEstate. They were left out of the sample because of two reasons. First, their optionsseries start late in the sample and are comparatively less liquid, with several monthswithout reported trading. In addition, there are breaks in the data. For instance, theSTOXX indices shifted methodology from Dow Jones Global Classication Standard toIndustry Classication Benchmark (ICB)in September 2004, affecting the compositionof the Personal & Household Goods and Travel & Leisure Supersectors and makingtheir corresponding DD series not comparable. In addition, the Real Estate Supersectorwas elevated to Supersector in 2008, after having been part of the Financial ServicesSupersector, which constitutes another break in the data.and sectoral characteristics in each portfolio vis--vis grouping atIndustry level and it is narrow enough not to blur interactionsamong them, as it is the case at Sector level. An additional and veryimportant reason for this grouping criterion is the fact that PortfolioDD are built so they include option-based information and the mostliquid option market for sector indices are the EURO STOXX optionson ICB-based Supersectors traded at Eurex.

Constituent lists in each Supersector Index are revised everyquarter and reclassications take place whenever relevant corpo-rate events occur. In order to minimize possible spurious variationin the risk indicator, the portfolio constituents take into accountthese changes and make some assumptions when required. Appen-dices C and E describe in detail the company sample by portfolioand all additional assumptions made to ensure the portfolios accu-racy, including exclusions and ad-hoc reclassications.

3. Sample and preliminary analysis

The sample consists of 12 out of the 19 EURO STOXX Supersec-tors.8 These sectors are the most relevant by different measures ofsize, e.g. assets, market value, employment. They have been includedin the sample according to two main criteria in order to ensure thebest informative quality of their market-based indicators, namely:(1) availability and liquid trading volume of their associated Eurex

99 The DD series were initially computed on a daily basis and then averaged toobtain monthly data. Volatilities from a GARCH(1,1) model applied to the respectiveSupersector index were used to complete the volatilities series when unavailable.

ustry

anciacom& gaanciahnolsumlitiesustriic msumsumlth

3) 23

pen

& FTable 1Supersectors sample.

ICB Supersector ICBSupersector Code Ind

1 Banks1 BNK Fin2 Telecommunications1 TLS Tele3 Oil & gas2 ENE Oil4 Insurance2 INS Fin5 Technology1 TEC Tec6 Automobiles & parts2 ATO Con7 Utilities3 UTI Uti8 Industrial goods & services4 IGS Ind9 Chemicals4 CHM Bas10 Food & beverage4 FOB Con11 Media3 MDI Con12 Health care1 HCR Hea

Notes: Series of implied volatilities start dates:(1) 25-September-01, (2) 31-July-02, (methodology prior to September 2004.

a Portfolio size does not include companies predecessors, for more details, see Apb Average monthly volume over the whole timespan.c Year-end average over the whole time span in thousands of euros.

M. Saldas / Journal of BankingFig. 1 displays together the 12 sectoral DD series and the EUROSTOXX 50, the benchmark stock index in the euro area. Being amarket-based indicator, DD series move along together with thestock market benchmark. In fact, they visibly lead it. This featureserves to illustrate the forward-looking properties of the DD seriesfrom option prices as inputs (Saldas, 2012). The DD series antici-pate turning points along the entire period of analysis. Duringthe recent crisis, they reach their bottom at the end of 2008 whilethe EURO STOXX 50 only picks up after the end of the rst quarterof 2009.

The DD series do not show a clear linear trend but they sug-gest a high degree of comovement along the whole time spanand correlation among them is very high on average and statisti-cally signicant both in levels and in rst differences. Figs. 2 and3 show the median and quartile regions of bilateral correlationcoefcients across sectors using 24-month moving windows ofDD series levels and rst differences in order to illustrate thechanging pattern of cross-section sectoral risk correlation overtime. Median correlation is high in the entire sample but it showsgreater dispersion in tranquil times where idiosyncratic drivers ofsector risk dominate. However, median correlation increases and

0

1

2

3

4

5

6

7

8

9

10

11

12/0

1

06/0

2

12/0

2

06/0

3

12/0

3

06/0

4

12/0

4

06/0

5

12/0

5

BNK TLSINS TECUTI IGSFOB MDIEURO STOXX 50 (right)

Fig. 1. Sectoral Distance-to-Default series. December-2Portfolio Options MarketSizea Volumeb Capitalizationc

ls 40 24894.6 490278.1munications 17 5439.5 245011.1s 19 5130.3 272077.1ls 17 5406.9 233824.6ogy 21 2952.7 233154.2er goods 13 3161.0 117228.1

22 2536.2 216164.7als 56 412.6 108511.6aterials 14 162.1 147751.8er goods 13 677.4 94878.9er services 25 620.0 87118.6care 17 116.7 100830.1

-September-02, (4) 19-May-03. Supersector codes are assigned according to the ICB

dix C.

inance 37 (2013) 45344555 4537its dispersion across sectors narrows signicantly in episodes ofhigher stress in nancial markets, e.g. in the aftermath of thedot-com bubble burst in 2002; after the subprime crisis start inAugust 2007; and especially in the third quarter of 2008, afterLehman Brothers collapse. At the end of the sample, median riskcorrelation across sectors remains high, but there is greater dis-persion suggesting somehow a moderation in the role of sector-wide risk drivers.

Table 2 reports preliminary cross-section dependence tests ap-plied to levels and rst differences of DD series regressed on sectorintercepts. High values of all these statistics reject the null hypoth-esis of cross-section independence and conrm the results ofgraphical inspection: DD series show a high degree of cross-sectiondependence even if the series are differentiated.

In addition to strong comovement and high correlation amongthe series, the results in Table 2 suggests that it is very likely tohave both observable and unobservable common factors at place.Variables from the macroeconomic environment and from nan-cial markets are strong candidates as common factors and inducestrong cross-section dependence across sectors (Alves, 2005; Hollyet al., 2010).

0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

06/0

6

12/0

6

06/0

7

12/0

7

06/0

8

12/0

8

06/0

9

ENEATOCHMHCR

001October-2009. Source: Authors calculations.

& Finance 37 (2013) 453445550.7

0.8

0.9

1.0

4538 M. Saldas / Journal of BankingAdditionally, this particular dynamics in DD series may also becaused by risk diffusion across sectors, which in turn may come inform of economic proximity and additional unobserved factors.Risk transmission is likely to be variant across sectors and changein time and the nature of sectoral economic proximity comes frommany sources. Similarity of business lines is a rst source of this typeof relationship and it includes common customer base and competi-tion relationships. Financial linkages are another source of shockspillovers and take place not only between the nancial sector and

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

12/0

3

06/0

4

12/0

4

06/0

5

12/0

5

06/0

6

Median Pairwise Correlation

Fig. 3. Sectoral DD Series Pairwise Correlation (series in rst differences). Source: Authorin rst differences.

0.0

0.1

0.2

0.3

0.4

0.5

0.6

11/0

3

05/0

4

11/0

4

05/0

5

11/0

5

05/0

6

Median Pairwise Correlation

Fig. 2. Sectoral DD series pairwise correlation. Source: Authors calcula

Table 2Preliminary cross-section dependence analysis DD series.

q CDP

DDi,t 0.843 66.7

DDDi,t 0.595 46.9

Notes: q and CDP are computed as detailed in Section B.1 using residuals ofregressions on a sector-specic intercept. Indicates the series show cross-section dependence at 5% level.the non-nancial corporates, but also between non-nancial com-panies through credit chains and counterparty risk relationships.See results in Couderc et al. (2008), Das et al. (2007), Jarrow andYu (2001), and Veldkamp and Wolfers (2007) for in depth discus-sions of these relationships.10 Finally, other complementarity rela-tionships are also relevant. They can take place throughtechnological linkages (Raddatz, 2010) or collateral channels of riskthrough the securities channel (Benmelech and Bergman, 2011).

4. The econometric model

This section describes in detail the econometric model to ana-lyze the risk determinants and transmission across the euro areas

12/0

6

06/0

7

12/0

7

06/0

8

12/0

8

06/0

9

First and Third Quartiles

s calculations. Correlation is calculated using a 24-month moving window of series

11/0

6

05/0

7

11/0

7

05/0

8

11/0

8

05/0

9

First and Third Quartiles

tions. Correlation is calculated using a 24-month moving window.

10 Bernoth and Pick (2011) also explore spatial effects in risk diffusion betweenbanking and insurance sectors using DD series of individual institutions from Asia,North America and Europe. In this paper, the spatial component is not relevant sinceportfolios are constructed bundling together only euro area companies.

nancial and corporate sectors using the Portfolio Distance-to-De-

M. Saldas / Journal of Banking & Finance 37 (2013) 45344555 4539fault series constructed following the methodology presented inSection 2 and Appendix A.

Under the potential presence of cross-section dependence in theDD series, a suitable econometric method is the Common Corre-lated Effects (CCE) estimator introduced in Pesaran (2006) andfurther extended in Pesaran and Tosetti (2011) and Chudik et al.(2011). CCE is a consistent econometric panel data method in pres-ence of different degrees of cross-section dependence coming fromcommon observed and single or multiple unobserved factors andfrom proximity-driven spillover effects (Pesaran and Tosetti,2011; Chudik et al., 2011). The CCE method also tackles methodo-logical limitations of other econometric models when modelinginterrelationships across sectors due to large N dimension, e.g.VAR (Pesaran et al., 2004).

This method is computationally simple and has satisfactorysmall sample properties even under a substantial degree of heter-ogeneity and dynamics, and for relatively small time-series andcross-section dimensions (N = 12 and T = 95 in this case). It is alsoconsistent in presence of stationary and non-stationary unob-served common factors (Kapetanios et al., 2011) and more suitablefor this dataset than a SURE model due to the possible presence oftime-variant correlation patterns, as suggested for this case inFigs. 2 and 3.

The general model specication is a dynamic panel and takesthe following form:

DDi;t aidt biXi;t ui;t; i 1; . . . ;N; t 1; . . . ; T 1where DDi,t is the Distance-to-Default of sector i at time t. The vec-tor dt includes the intercepts and a set of observed common factorsthat capture common macroeconomic and nancial systemic mar-ket shocks. Xi,t is the vector of sector-specic regressors, includinglags of is own Distance-to-Default, the direct risk spill-overs fromneighboring sectors and other sector-specic variables. All coef-cients are allowed to be heterogeneous across sectors11 and allremaining factors omitted in the specication and other idiosyn-cratic risk drivers are captured in the error term ui,t.

The CCE estimator can be computed by OLS applied to sector-individual regressions where the observed regressors are aug-mented with cross-sectional averages of the dependent variableand the individual-specic regressors. The CCE estimator providestwo versions, namely the CCE Pooled estimator (CCEP) and the CCEMean Group estimator, of which only the latter will be reportedbecause of slope heterogeneity and no need for CCEP efciencygains in this case.

4.1. Macroeconomic and nancial risk determinants

A set of ve exogenous variables is included in the model in or-der to control for determinants originated in the macroeconomicenvironment and to capture risk sensitivity to common shocks innancial markets. A number of papers quoted in Section 1 havedocumented the explanatory power of macroeconomic and nan-cial variables in corporate default risk, thus their omission couldbias the results of the parameter estimation in the model.

The model takes macro-nancial determinants as exogenousand chooses to ignore possible feedback effects to the macro-nan-cial environment. Examples of this approach and additional expla-nation for this modeling decision can be found in Castrn et al.(2010) and Castrn et al. (2009). Accordingly, the econometricspecication rst includes the annual rate of change of the Indus-

11 See for instance results in Castrn et al. (2010) for a more detailed, yet not strictly

comparable, discussion of heterogeneous impact of macro variables on distress ofcorporate sectors, which are dened using the European classication of economicactivities (NACE).trial Production Index (DPIt) and the Harmonised Index of Con-sumer Prices (DCPt) in the euro area, in order to capture theeffect of demand shocks. Brent Oil (1-Month Forward Contract)prices changes denominated in euro (OILt) detect supply shocks.The short-term benchmark interest rate is also included usingthe 3-Month Euribor Rate (R3Mt), which also reects developmentsin the money market affecting the nancial sector and serves as areference for corporate debt yields and borrowing. They also arelinked to corporate asset return growth. Finally, the Chicago BoardOptions Exchange Volatility Index (VIXt) is included to gauge globalequity market sentiment. The VIX index tends to be lowwhenmar-kets are on an upward trend and tends to increase with marketpessimism, therefore its relationship with DD series is expectedto be negative.

4.2. Sector-specic risk determinants

4.2.1. Sector-specic risk determinantsThe model includes four other sector-specic regressors com-

puted for each ICB Supersector Index,12 namely: (1) the annual rateof change of the Price-Earnings Ratio, DPEt; (2) the annual rate ofchange in Dividend Yields, DDYt, (3) the Return On Assets, ROAi,t;and (4) the monthly log excess return on each indexs daily price re-turn relative to the EURO STOXX 50 Index, EXRETi,t.

Earnings, as measured by Price-Earnings Ratio, and protability,as measured by ROA, are studied extensively in the corporatebankruptcy literature. Indeed, results in Shumway (2001), Beaveret al. (2005), and Chava and Jarrow (2004) show that higher earn-ings are traditionally associated with lower distress probabilities,in spite of a weaker informational ability detected in recent yearsdue to higher frequency in earnings restatements and the possibil-ity of data manipulation (Dechow and Schrand, 2004). Return OnAssets (ROA) incorporates further information about protabilityand the ability of the companies in a given sector to generate re-turns. Dividends traditionally serve to assess and infer corporateperformance. Recent work by Charitou et al. (2011) shows that div-idend payment initiations or increases tend to reduce corporatedefault and tend to raise the assets returns for several subsequentperiods. However, specially in the nancial sector, aggressive divi-dend policies may also encourage risk-taking and erode the capitalbase of a company or sector (Acharya et al., 2011). Excess returnsare a purely market-based measure of relative performance ataggregate level and is motivated by results in Campbell et al.(2008).

No additional rm-level information or sector specic indicatorare included in the model since the DD construction already in-cludes either directly or indirectly the most relevant variables ofsector risk, i.e. market-implied assets returns and volatility andaggregated leverage (Bernoth and Pick, 2011; Gropp et al., 2004).

4.2.2. Neighboring sectors risk spilloversThe risk spill-over across sectors is studied using DD series from

neighboring sectors. For a given sector i, the neighboring effect isdened by:

DDnii;t 1ni

Xnj1

DDj;t 2

DDni;t is a simple average of the DD series of the n neighbors (DDj,t)of sector i.

For each sector i, the number of neighbors and weighting oftheir corresponding DD series are determined by a contiguity ma-trix (see Table 3) derived from ad-hoc and predened neighbor-12 See Appendix D for details of these determinants and the rest of macro-nancialvariables.

5. Empirical results

5.1. Cross-section dependence and non-stationarity analysis

Preliminary analysis in Section 3 detected a high degree of com-ovement in DD series in levels and rst differences. This section

Table 3Contiguity matrix.

BNK TLS ENE INS TEC ATO

BNK 0 0 0 1 0 0TLS 0 0 0 0 1 0ENE 0 0 0 0 0 0INS 1 0 0 0 0 0TEC 0 1 0 0 0 0ATO 0 0 0 0 0 0UTI 0 0 1 0 0 0IGS 0 0 1 0 1 1CHM 0 0 0 0 0 0

and

DDYi,t 0.383 0.350 0.352 0.348 0.346 0.342 0.322DPEi,t 0.030 0.023 0.024 0.021 0.040 0.036 0.034EXRETi,t 0.046 0.045 0.040 0.039 0.040 0.038 0.039ROAi,t 0.266 0.267 0.266 0.265 0.266 0.264 0.246DDni;t 0.716 0.715 0.719 0.718 0.719 0.720 0.710

Pesaran test (CDP)DDi,t 47.6 47.2 47.4 47.1 47.0 46.7 45.4

DDYi,t 27.6 25.1 25.3 25.0 24.9 24.6 23.2

DPEi,t 2.2 1.7 1.8 1.5 2.9 2.6 2.4

EXRETi,t 3.6 3.5 3.1 3.0 3.1 2.9 3.0ROAi,t 20.9 20.9 20.7 20.5 20.5 20.2 18.7

DDni;t 56.4 56.0 56.0 55.6 55.4 55.2 54.1

Notes: pth-order Augmented Dickey Fuller ADF(p) regressions are computed foreach Supersector i without cross-section augmentations and for lag ordersp = 0, . . . , 6 over the whole sample. Tests forDDYi,t andDPEi,t are based on a reducedsample N = 11, excluding the oil & gas supersector due to short series length. Nolinear trend is included. Indicates rejection of the the null hypothesis of no error cross-sectional depen-dence at 5% level.

4540 M. Saldas / Journal of Banking & Finance 37 (2013) 45344555hood linkages among sectors.13 Even though the denition of neigh-boring sectors and cross-sectional dependence in the literaturecomes largely from spatial proximity (Holly et al., 2011; Hollyet al., 2010; Pesaran and Tosetti, 2011; Chudik et al., 2011), othermeasures of proximity, from economic or social networks, are alsoused in recent research (Conley and Topa, 2002; Conley and Dupor,2003; Holly and Petrella, forthcoming).

In the case of corporate sectors, the literature does not provide adenite metric to determine neighborhood linkages, because sec-toral relationships depend both on the choice of sector classica-tion and on the sectoral characteristics to be linked.14 Pesaranet al. (2004) argue that the aggregation error in this type of exercisescan be minimized if the cross-section units, i.e. sectors in this case,are similar and the weights are chosen carefully.

As a result, the approach in this model is ad-hoc and market-based. It relies on similarity of business lines embedded in theICB methodology and covers important and overlapping dimen-sions of sectoral interdependencies, namely: balance-sheet expo-sures, nancial linkages, common accounting practices,technological linkages, etc.

Supersectors are rst assumed to be neighbors if they belong tothe same Industry, an upper level of aggregation to Supersectors inthe ICB methodology structure. For instance, the Industry of Con-sumer Goods links the Supersectors of Automobiles & Parts andFoods & Beverages while Banks and Insurance Supesectors are bun-dled together as Financials.

The second proximity criterion used to aggregate series intoneighbors is also based on the ICB methodology but it relies on themost frequent company reclassications across Supersectorswithinor outside a given Industry during the time span used in the paper.

FOB 0 0 0 0 0 1MDI 0 0 0 0 1 0HCR 0 0 0 0 0 0

Notes: Supersector codes are in Table 1. If element i, j = 1, the supersectors in row iUnder multiple business lines, company reclassications take placemainly due to changes in themain business line and also due to cor-porate actions such as spin-offs or M& A. Examples of this were fre-quent in supersectors such as Industrial Goods & Services, Oil & Gasand Utilities, which do not belong to the same ICB Industries.

13 The contiguity matrixW is an N N nonnegative matrix, whose wi,j element is 1 ifsectors i and j are considered neighbors and 0 otherwise. The number of neighbors forsector i is the sum of elements along row i. Although weighting criteria is not likely toaffect the properties of the econometric approach (Chudik et al., 2011) and a validalternative in this case could weigh DD series by implied assets from the calibration,this paper assumes equal weights in the neighborhood average (1/n) because thenature of the business in each sector affects considerably the asset sizes, hence, asset-based weights could introduce distortion. In addition, there is no only andunambiguous way to determine relative importance of sectors among each other.14 Most studies deal with manufacturing sectors data, excluding nancials. Forexample, Conley and Dupor (2003) study sectoral synchronization of output andproductivity growth using factor demand linkages as a metric for economic distancefor US corporates and dene the sectors of study using the SIC system. Holly andPetrella (forthcoming) use input-output linkages and analyze the shock propagationacross manufacturing sectors.UTI IGS CHM FOB MDI HCR

0 0 0 0 0 00 0 0 0 0 01 1 0 0 0 00 0 0 0 0 00 1 0 0 1 00 1 0 1 0 00 0 0 0 0 00 0 1 0 1 00 1 0 0 0 10 0 0 0 0 00 1 0 0 0 00 0 1 0 0 0

column j are considered neighbors. See Section 4.2.2 for more details.

Table 4Residual cross-section dependence of ADF(p) regressions.

ADF(0) ADF(1) ADF(2) ADF(3) ADF(4) ADF(5) ADF(6)

Average cross-correlation qDDi,t 0.605 0.603 0.608 0.608 0.610 0.610 0.596takes a step further and extends the CD tests to the rest of sec-tor-specic variables in the panel allowing for different degreesof serial correlation in the data. It also conducts stationarity anal-ysis of the data for correct model specication, taking into accountthe potential presence of CD.15

Table 4 reports CD statistics of residuals from ADF(p) regres-sions of the DD series and the sector-specic variables, includingthe neighboring sectors DD series (DDni;t). Results detect that DDseries and the DDni;t present very high and positive average correla-tion coefcients, above 60%, whereas correlation for dividendyields growth and Returns On Assets are also large but in the rangeof 2540%. Price-Earnings ratio growth show very low correlationacross sectors, with a coefcient of around 3%. Excess returns rela-tive to the benchmark index also shows a low, though negativeaverage correlation coefcient of 4% on average. CD test statistics,reported below, are in line with these results and are highest for

15 See Appendix B for technical details of the tests described in this subsection.

Table 5Panel unit root tests.

CADF(0) CADF(1) CADF(2)

CIPS panel unit root testsDDi,t 3.49 3.41 2.92

he o

M. Saldas / Journal of Banking & Finance 37 (2013) 45344555 4541DDYi,t 2.19 2.29 2.73DPEi,t 4.29 3.32 2.83EXRETi,t 6.18 6.06 5.00ROAi,t 1.49 1.75 1.90DDni;t 3.65

2.29 2.73

ADF(0) ADF(1) ADF(2)

IPS panel unit root testsDDi,t 1.34 1.68 0.80DDYi,t 0.43 2.29 3.91DPEi,t 9.97 6.39 4.8EXRETi,t 25.76 20.4 14.62ROAi,t 1.79 1.38 1.23DDni;t 1.05 1.39

0.47

Notes: Tests for DDYi,t and DPEi,t are based on a reduced sample N = 11, excluding t Indicate rejection of the the null hypothesis of unit root at 10% level. Indicate rejection of the the null hypothesis of unit root at 5% level. Indicate rejection of the the null hypothesis of unit root at 1%, level.

Table 6Unit root tests macroeconomic and nancial risk variables.

Variable Level First differenceDD series and neighboring effects, and smaller yet signicant forthe rest of the variables. These tests conrm the strong cross-sec-tion dependence in the data, with arguably the exception of Price-Earnings ratio growth and the relative indexes excess return.

In line of the results of CD tests, panel unit root tests for the DDseries and the sector-specic regressors need to take into accountcross-dependence. Accordingly, Table 5 summarizes the CIPS panelunit root tests described in Appendix B.2. IPS test statistics are alsoreported for robustness check and comparison. Both CIPS and IPStests reject unit roots in dividend yields growth, Price-Earnings ra-tio growth and excess returns and do not reject the null in the caseof ROA. Interestingly, the CIPS strongly reject unit roots in the caseof DD series and neighboring effects for all lag orders p, whereasIPS tests seem to suggest non-stationarity in most cases tested. Gi-ven the substantial degree of cross-section dependence detected inthese series, the CIPS tests provide a more reliable inference andthese variables are also taken as I(0). These tests also point outto the combination of non-stationary common factors and station-ary idiosyncratic components in the sectoral risk.16 ROA is taken asI(1) and enters the model in rst differences.

VIXt 2.07 7.78R3Mt 1.73 3.92OILt 2.11 6.09DPIt 2.79 2.60DCPt 6.00 4.77

Notes: Intercept included only in levels, lag length determined by AIC and HQcriteria. Indicate rejection of the the null hypothesis of unit root at 5% level.Results are robust to inclusion of trend and seasonal dummies; and also to struc-tural breaks in two cases (DPIt,DCPt). Indicate rejection of the the null hypothesis of unit root at 10% level. Indicate rejection of the the null hypothesis of unit root at 1% level.

16 The non-stationarity detected in DD series and DD-neighbors using IPS testscomes from the combination of non-stationary common factors and stationaryidiosyncratic components. This possibility has been veried by adopting the PanelAnalysis of Nonstationarity in the Idiosyncratic and Common components or PANICapproach advanced by Bai et al. (2004). This result is consistent with ndings in Alves(2005), and provides empirical support to the notion that aggregate sectoral riskevolves to a long-run equilibrium, which is in turn affected temporarily by the macro-nancial environment and the cross-sectoral dynamics captured by the CCE method.Finally, individual ADF(p) unit root tests were run for themacro-nancial variables described in Section 4.1 which enterthe model as exogenous regressors. Based on the results of thesetests reported in Table 6, the annual rates of change of the Indus-trial Production Index (DPIt) and the Harmonised Index of Con-sumer Prices (DCPt) enter the model as I(0) variables, while BrentOil prices (OILt), the 3-Month Euribor Rate (R3Mt) and the VIX Vol-atility Index (VIXt) are previously differentiated to enter the model.

5.2. Model estimation

The results from estimation of Eq. (1) are reported in Table 7.Columns [1][3] are estimates of nave OLS Mean Group (MG)models (Pesaran and Smith, 1995) that neglect cross-sectiondependence (CD). Columns [4][6] are Common Correlated Effects

CADF(3) CADF(4) CADF(5) CADF(6)

2.77 2.62 2.66 2.442.79 2.77 2.61 2.593.19 2.81 2.56 2.514.57 4.31 4.14 3.902.24 2.31 2.36 2.282.79 2.77 2.61 2.59

ADF(3) ADF(4) ADF(5) ADF(6)

0.66 0.40 0.55 0.454.39 3.82 4.11 5.945.96 4.84 4.26 4.8112.03 11.35 10.69 10.060.09 0.26 0.26 1.870.33 0.14 0.18 0.14

il & gas supersector due to short series length. No linear trend is included.(CCE) estimates of these same model specications, hence moresuitable to the CD properties analyzed in the previous section.Although MG estimates are likely to be biased, they serve as abenchmark for the CCE estimates and also put into context the rel-evance of CD in the model specication. They also serve to comparethese results with previous studies on determinants of aggregatesectoral risk. .

The most relevant nding from the estimation results is the lim-ited relevance of shocks originated in the macroeconomic andnancial environment on DD, especially when CD is accountedfor. This result has several interpretations and does not necessarilymean that sectoral risk is not affected by the macro-nancial envi-ronment. First, business cycle volatility is likely to be smoothed outin the construction of DD series or other CCA risk measures (espe-cially EDF). Indeed, as suggested in International Monetary Fund(2011), some high-frequency indicators of distress have the abilityto anticipate the cycle, which is very likely in the case of instru-ments derived from equity and option markets.17 Marked-basedindicators, such as DD, may also be less directly responsive due tonon-linearities in their interactions with macroeconomic and nan-cial variables (Sorge and Virolainen, 2006). In addition, macro-nan-cial effects may impact sectoral DD in a more indirect way, viamarket news already embedded in the DD inputs and/or through

17 To test this hypothesis, I conducted a simple Granger-causality test using theindustrial production growth (PIt) and the average of DD series and found that indeedthe DD series Granger-causes changes in activity up to one trimester.

CCEMG CCEMG CCEMG AMG

)5

)

)7

)2)1)

)

)

)

)

)

)

& Finance 37 (2013) 45344555Table 7Estimation results.

Dependent variable MG MG MGDDi,t [1] [2] [3]

Intercept 0.481 0.701 0.487

(0.068) (0.134) (0.206DVIXt 0.083 0.083 0.08

(0.005) (0.005) (0.006DR3Mt 0.670 0.691 0.678

(0.103) (0.101) (0.103DOILt 0.004 0.006 0.00

(0.004) (0.004) (0.004DPIt 0.000 0.003 0.00

(0.001) (0.004) (0.005DCPt 0.025 0.038 0.01

(0.011) (0.026) (0.032DDYi,t 0.001 0.001

(0.001) (0.001DPEi,t 0.001 0.002

(0.001) (0.001EXRETi,t 2.365 2.275

(0.486) (0.478DROAi,t 0.175 0.241

(0.072) (0.133DDi,t1 0.921 0.883 0.767

(0.008) (0.019) (0.035DDni;t1 0.141

(0.035Observations 1128 1072 1072

4542 M. Saldas / Journal of Bankingcross-dynamics transmitting risk across industries (Alves, 2005).Lastly, even though sector-specic coefcients may have individualstatistically signicant signs, they are allowed to be heterogeneousacross sectors and the effect across the panel members may be aver-aged out (Eberhardt and Teal, 2010).

In particular, the VIX Volatility Index (VIXt), a measure of inves-tors risk sentiment, is statistically signicant at 5% across all theMG estimates and shows a stable and expected negative sign, indi-cating an increase in sector-wide risk, i.e. a drop in DD, when equi-ty markets become more volatile. However, in all modelsestimated using the CCE method its effect on overall sectoral riskvanishes. This is not a surprising result, as Bernoth and Pick(2011) report that the VIX Index is absent in their CCE-based mod-els when forecasting DD at rm-level for banks and insurance com-panies. A very plausible explanation in this case is that optionimplied volatilities from index options endow the sectoral DDwiththe forward-looking information embedded in the VIX Index.

The same holds true for the 3-month Money Market (Euribor)Rate, R3Mt, which shows statistical signicance at ve percentlevel and a positive and stable coefcient only if CD is ignored.The effect of short-term interest rates on sectoral risk was ex-pected to be negative if we consider them as a proxy of borrowingcosts and risk premia. However, since short-term interest rates areclosely linked to the risk-free rate used to capture sectoral assetsreturn growth in the DD computation via the yield curve, this fea-ture is likely to be dominant in the estimates in this case. In addi-tion, several empirical studies link the short-term interest rates tohigher performance and make an empirical case for the positive

RMSE 0.562 0.507 0.496q 0.424 0.429 0.424CDP 33.40 37.78 31.50

AR(1) 7.396 5.482 7.044

Notes: MG, CCEMG and AUG stand for Mean Group (Pesaran and Smith, 1995), Common C(Eberhardt and Teal, 2010; Eberhardt and Bond, 2009) estimates respectively. Standard ecomputed on residuals of each equation. See Appendix B.1 for denitions of the cross-s Indicates rejection of the null of no rst-order autocorrelation test described in Drukk Indicates rejection of the null hypothesis of cross-section independence at 5% signic denotes signicance at 10%. denotes signicance at 5%.[4] [5] [6] [7]

0.058 0.024 0.128 5.403(0.136) (0.164) (0.156) (0.222)0.000 0.004 0.004 0.001(0.004) (0.004) (0.004) (0.007)0.010 0.038 0.066 0.024(0.095) (0.088) (0.088) (0.119)0.000 0.002 0.003 0.111(0.003) (0.003) (0.004) (0.003)0.000 0.001 0.001 0.094(0.003) (0.003) (0.004) (0.009)0.014 0.003 0.001 0.423(0.021) (0.017) (0.017) (0.031)

0.001 0.001 0.001(0.001) (0.001) (0.002)0.001 0.000 0.000(0.001) (0.001) (0.003)2.145 2.235 3.118

(0.455) (0.485) (0.626)0.079 0.079 0.062(0.064) (0.069) (0.045)

0.740 0.668 0.652 0.417

(0.042) (0.046) (0.047) (0.076)0.050 0.479

(0.037) (0.088)1128 1072 1072 1072sign. This positive effect becomes nil when CD is considered, prob-ably because the unobserved common factors capture it. This re-sult is at odds with ndings in sberg and Shahnazarian(2009),18 where the authors use a single risk indicator for the wholecorporate sector, but consistent with those from Castrn et al.(2010), where short-term interest rates are in general insignicantacross several corporate sectors studied individually.

Shocks from oil prices (OILt) do not exert any statistically signif-icant effect but in equation [3], when CD is omitted, and nonewhen CD is taken into account. The rst result is not entirely atodds with the literature, as Alves (2005) nds that oil prices donot affect but one of the seven sectors he includes in his study.Shocks from industrial production growth (PIt) are insignicanton DD even when CD is neglected, whereas growth in consumerprices (CPt) affects negatively, as expected, on overall sectoral riskin only one of the MG specications, equation [1]. This impact be-comes insignicant when sector-specic regressors are included inthe MG model in two of three cases when CD is controlled for. Itscorresponding coefcient equation [6] exerts a positive coefcient.Again, the changing statistical signicance in the MG models is asign of specication failure to account for unobserved common fac-tors appropriately. In turn, lack of statistical signicance of these

0.360 0.328 0.323 0.4380.082 0.080 0.078 0.0346.48 6.07 5.89 2.511.370 2.129 2.188 13.128

orrelated Effects Mean Group (Pesaran, 2006) and and the Augmented Mean Grouprrors are given in parenthesis. CD statistics (q and CDP) and panel unit root tests areection dependence tests.er (2003).ance level.

18 In this paper, the authors analyze effects of macroeconomic shocks on the themedian EDF of the whole corporate sector in Sweden in a VEC model. This series is aI(1) variable, in line with the ndings described in the stationarity analysis of thispaper, but this analysis does not take into account the heterogeneity across sectorsand the cross-section dependence is ignored.

TO UTI IGS CHM FOB MDI HCR

0.726 0.394 0.071 0.213 0.497 0.227 1.0350.390.00.01.0990.360.00.010.00.01.1400.100.00.00.0000.00.4881.330.00.19.3990.12.054

0.09

& Finance 37 (2013) 45344555 4543variables with CCE estimates show that the macro-nancial effectsare very likely to be captured either by the unobserved common ef-fects and/or the set of sector-specic variables more accurately.

Table 8CCE estimates of all cross section units.

DDi,t BNK TLS ENE(a) INS TEC A

ICB supersectorIntercept 1.007 0.175 0.206 0.444 0.095

(0.480) (0.346) (0.642) (0.367) (0.213) (DVIXt 0.010 0.005 0.01 0.003 0.006

(0.018) (0.013) (0.013) (0.012) (0.011) (DR3Mt 0.000 0.307 0.000 0.000 0.022 0

(0.000) (0.301) (0.000) (0.000) (0.250) (DOILt 0.006 0.01 0.000 0.015 0.011

(0.013) (0.01) (0.000) (0.010) (0.009) (DPIt 0.017 0.027 0.012 0.016 0.005

(0.014) (0.014) (0.020) (0.015) (0.009) (DCPt 0.000 0.000 0.038 0.018 0.002 0

(0.000) (0.000) (0.111) (0.067) (0.066) (DDYi,t 0.002 0.002 0.001 0.003 0.001

(0.004) (0.001) (0.003) (0.002) (0.003) (DPEi,t 0.003 0.001 0.004 0.000 0.000 0

(0.003) (0.001) (0.007) (0.000) (0.000) (EXRETi,t 0.812 4.341 3.090 0.793 1.732 3

(1.771) (1.205) (1.537) (1.238) (0.803) (DROAi,t 0.758 0.019 0.010 0.226 0.007

(2.002) (0.042) (0.123) (0.414) (0.044) (DDi,t1 0.519 0.484 0.728 0.640 0.837 0

(0.112) (0.103) (0.111) (0.088) (0.061) (DDni;t1 0.033 0.199

0.216 0.033 0.134 0(0.133) (0.116) (0.232) (0.077) (0.142) (

Notes: Individual estimates come from model [6] in Table 7. Signicance at 10%. Signicance at 5%.

M. Saldas / Journal of BankingSector-specic regressors on DD display better results in termsof stable and strong statistical signicance under CD, which chal-lenges the macro dominant focus in the existing literature of nan-cial stability and highlights the importance of market-based andsector-level information and interactions for policy analysis of sys-temic risk. Among the set of six sector-specic covariates, the mar-ket-based indicators show stronger relevance as distress driversthan those computed using balance-sheet information under alter-native econometric methods. In particular, dividend yields growth(DDYi,t) does poorly and shows no statistical signicance in allmodels. The Return On Assets (DROAi,t) and Price-Earnings Ratio,DPEt, show expected positive signs in MG models, equations [2]and [3], but become insignicant when the CCE method is applied,in line with ndings in Bernoth and Pick (2011).

Two sector-specic andmarket-based variables show strong andsignicant effects regardless the econometric method used, whichcan be interpreted as evidence of the relevance of market-basedinformation about distress beyond common observed risk drivers.The distress risk persistence, as proxied by the lag of the dependentvariableDDi,t1, shows a large and signicant positive sign across allmodels. The CCE estimates show however smaller coefcients asadditional regressors are included in the specications. These MGcoefcients are larger, close to one, probably because MG estimatescapture also the non-stationary common components. The strongsignicanceof this regressor conrms results in the literature (Alves,2005; Bernoth and Pick, 2011) and illustrates the persistence in idi-osyncratic sectoral risk even after controlling for CD.Withmore eco-nomic relevance, the sectoral indices excess return relative to theEURO STOXX 50 Index, EXRETi,t, exerts strong effects on sector-widedistress. This variable shows a positive and signicant coefcientsign across all model specications, illustrating that outperformingsectors relative to the corporate sector as a whole results in higherresilience and thus lower distress risk.Finally, the neighboring sectors risk lagged effect on DD,19

DDni;t1, is statistically insignicant in the CCE-estimated model,while MG estimates in Model [3] do exhibit a positive coefcient.

6) (0.466) (0.318) (0.381) (0.451) (0.274) (0.456)33 0.023 0.029 0.006 0.003 0.000 0.0046) (0.018) (0.014) (0.014) (0.018) (0.012) (0.014)

0.484 0.466 0.263 0.106 0.502 0.3056) (0.431) (0.313) (0.334) (0.43) (0.284) (0.324)10 0.020 0.005 0.021 0.022 0.005 0.0003) (0.015) (0.011) (0.012) (0.015) (0.010) (0.011)08 0.000 0.009 0.001 0.009 0.001 0.0045) (0.000) (0.014) (0.013) (0.018) (0.010) (0.012)

0.115 0.052 0.000 0.105 0.004 0.0003) (0.118) (0.076) (0.000) (0.115) (0.069) (0.000)01 0.001 0.003 0.002 0.000 0.003 0.0042) (0.005) (0.004) (0.007) (0.004) (0.002) (0.004)

0.009 0.001 0.000 0.008 0.000 0.0000) (0.007) (0.000) (0.003) (0.004) (0.001) (0.001) 4.813 2.777 1.136 0.907 1.029 3.486

6) (2.266) (1.779) (1.935) (1.895) (1.396) (1.369)69 0.126 0.104 0.048 0.231 0.003 0.0210) (0.245) (0.098) (0.056) (0.101) (0.038) (0.029) 0.711 0.843 0.423 0.736 0.843 0.664

3) (0.092) (0.122) (0.111) (0.09) (0.055) (0.104)0.108 0.005 0.025 0.023 0.296 0.040

1) (0.119) (0.371) (0.124) (0.136) (0.15) (0.098)This result implies that the risk impact in sectors with strong link-ages on other sectors does not work directly but is mainly capturedas unobserved common factor. It is also possible that the ad-hoc def-inition of neighboring sectors is not sufciently accurate and othersectoral dimensions than those described in Section 4.2.2 need tobe explored to obtain a more reliable contiguity matrix in terms ofdirect spillovers.20

Some of the overall results described so far are expected to varyacross sectors due to heterogeneous effects of the regressors andalso possibly because unobserved cross-sectoral and complexshocks alter the relationships with them. As the CCE modeling ap-proach allows to shed some light on this, Table 8 reports the resultsof model [6] at individual sector-level. To recap, this model is themost comprehensive and includes all variables described inSection 4.

At individual Supersector level, some macro-nancial variablesdo affect DD series but not necessarily with the same sign. As in theaggregate results, oil (OILt) and consumer prices (CPt) are the onlymacroeconomic variables that fail to show also at individual leveany effect on sectoral risk. Interest rates (R3Mt) do play a signi-cant role as proxy of borrowing costs and risk premia for the Media(MDI) supersector (0.502), while shocks from industrial produc-tion growth (PIt) exert a surprisingly negative effect on the idiosyn-cratic component of risk in the Telecommunications (TLS)supersector (0.027). The signicance of the VIX Index (VIXt) on

19 Contemporary effects were not gauged do the risk of dealing with potentiallystrong endogeneity and limited possibilities to nd valid instruments for thisregressor.20 Robustness checks have been conducted using matrices of return and volatilityspillovers according to the methodology described in Diebold and Yilmaz (2012) toconstruct the contiguity matrices. The results did not change for all modelsspecications.l

distress in the Automobiles & Parts and Industrial Goods & Services

This paper presents a framework to analyze risk in the corpo-

4544 M. Saldas / Journal of Banking & Fsectors with alternating signs, 0.133 and 0.029, respectively,illustrate the possibility of heterogeneous responses at individuallevel and nil effect on the average.

As for the sector-specic variables, dividend yields growth, DDYt, and Price-Earnings Ratio growth, DPEt, and Return On Assets(DROAi,t) affect also heterogeneously across Supersectors. For in-stance, the coefcients associated to dividend yields are signicantand surprisingly positive in the Telecommunications and Mediasector. Price-Earnings Ratio and Return On Assets growth affectsonly the Food & Beverages sector with expected positive signswhile being irrelevant for the rest of Supersectors.

Mirroring aggregated results, the lag of the dependent variable,DDi,t1, is highly signicant also at individual level, for all supersec-tors with the no exceptions, while the sectoral indices excess re-turns, EXRETi,t, are signicant for half of the sectors in the sample,with a positive sign in all cases. Finally, the lagged neighboring riskeffect, non-signicant overall, is a risk driver in two Supersectors,Telecommunications and Media, with positive signs in both cases.

Finally, the econometric estimates worth mentioning is thehigher goodness-of-t of CCE estimates, as measured by the lowerRoot Mean Squared Errors across all model specications. As itmight be expected, the cross-section dependence test statistics re-ported below display a remarkable decline when the CCE estimatoris applied and there is no signicant evidence of remaining CD inthe estimation residuals.21 It is however noticeable the negativesign in all q and CDP statistics for CCE estimates. Since these indica-tors are based on the sum of pairwise correlation coefcients, thesign indicates that negative correlation coefcients are more fre-quent and sizable after controlling for CD.22 Finally, the serial corre-lation tests show that residuals from all estimated models arestationary both individually and jointly.

5.3. Robustness check

Column [7] in Table 7 reports results of a robustness check ofmodel [6] using the Augmented Mean Group (AMG) estimator,which shows very similar properties to CCE and account forcross-section dependence by inclusion of a preassumed singlecommon dynamic process in the sector regressions imposedwith unit coefcient. AMG estimates provide an alternative esti-mator under CD and obtain an explicit estimate for the unobservedfactors. This estimator has been developed in Eberhardt and Bond(2009) and Eberhardt and Teal (2010) to deal with macro panelsand been applied to estimate Total Factor Productivity (TFP) incross-country production functions (Eberhardt and Teal, 2010)where TFP is the single common dynamic factor.

In our case, there are no priors to assume there is only one com-mon factor driving the idiosyncratic sectoral risk but this exerciseserves to stress the robustness of the results in the previous sec-tion. The estimates provide support to the previous results in termsof the statistical signicance of the lag of the dependent variableDDi,t1, and the sectoral indices excess return, EXRETi,t. In contrastto CCE estimates, the neighboring sectors risk lagged effect alsoappears relevant, with a negative coefcient, pointing out to acompeting nature of the relationships between sectors. Interest-ingly, three macroeconomic variable also show statistical signi-cance, namely the shocks from oil prices, industrial productionand consumer prices. Overall, these estimates conrm the CCE re-sults, although their goodness-of-t is relatively lower and theresiduals present some degree of serial correlation.

21 The CDP remains high enough not to reject the null of CD at 5%. This is due to the

large time series dimension compared to the number of cross-section units.22 Although not reported, a closer look at bilateral residual correlations conrms thisfeature and that this sum drives the value of the statistic, given that

2T

NN1q

1.rate sector that takes into account their strong sectoral linkagesand comovement. In a rst part, the paper describes a methodologyto compute comprehensive forward-looking risk indicators at sec-tor-level based on Contingent Claims Analysis with informationfrom balance sheets, equity markets and, more importantly, indexoption prices. The second part of the paper analyzes the propertiesof the resulting Distance-to-Default series and sets up an econo-metric model that incorporates the cross-section dependence ofsectoral risk. This model allows to study the determinants and dif-fusion of risk across sectors, including sector-specic drivers, themacroeconomic and nancial markets environment and proxim-ity-driven risk spill-overs.

In particular, the paper computes forward-looking Distance-to-Default DD series, a market-based indicator, for 12 of the 19 nan-cial and corporate sectors in the euro area as dened by the EUROSTOXX indices between December 2001 and October 2009. Theseseries show very good properties in terms of capturing cyclesand episodes of distress. The econometric analysis relies on theCommon Correlated Effects estimator of Pesaran (2006) in orderto stress the importance of cross-section dependence (CD) in therisk series over time, which is driven by common observed andunobserved factors.

Controlling for cross-section dependence among the Distance-to-Default series, the rst result of this analysis shows that sectoralrisk comprises a stationary idiosyncratic component and a non-sta-tionary common factor. This result provides empirical support tothe notion that aggregate sectoral risk evolves to a long-run equi-librium, with temporary deviations caused by the macro-nancialenvironment, sector-specic shocks and the cross-sectoraldynamics.

Results of the econometric model estimation using the CommonCorrelated Effects (CCE) method nd evidence supporting a morerelevant role of sector-specic variables as sectoral risk determi-nants in the corporate sector overall at the expense of the impactfrom macro-nancial variables. The sector-specic drivers includerisk persistence, measures of overall sectoral performance and alsodirect risk spill-overs from risk in related sectors. The macroeco-nomic and nancial common variables are either captured asunobserved common effects, averaged out by heterogeneous af-fects across sectors or smoothed out by construction of the Dis-tance-to-Default series. This empirical nding challenges much ofthe literature that focuses mainly on macroeconomic risk driversand tends to ignore sector-specic characteristics and speciallyinteractions either explicitly or implicitly through an aggregateanalysis of the whole corporate sector.

This study also provides empirical evidence of the high degree ofheterogeneity as concerns the relevance and responsiveness to therisk drivers used in themodel, both inmacro-terms as in sector-spe-cic terms. These results show that a macro-only focus of the anal-ysis of nancial stability would be misleading for policy if cross-section dependence and sectoral heterogeneity is ignored. These re-sults make a case for a more disaggregated analysis of risk acrosssectors without neglecting the inherent interactions that take placeamong them. Subjects for further research include the inclusion ofnon-linearities in the interaction of risk across sectors and exploringmore accuratemetrics to assess the direct risk intersectoral linkagesin order to extend the model to conduct stress tests.

Acknowledgments6. Concluding remarks

inance 37 (2013) 45344555The author is grateful to Antnio Antunes, Diana Bonm, JrgBreitung, Ben Craig, Frank De Jonghe, Olivier De Jonghe, Jonas

r is the instantaneous rate of growth of the portfolios assets

of Eqs. (A.2) and (A.5).The traditional approach aggregates individual estimates of im-

A j1 k1covariance of companies j and k.

B.1. Cross-section dependence tests

i1 ji1

& Finance 37 (2013) 45344555 4545Dovern, Salvatore DellErba, Dries Heyman, Peter Pedroni, Ivan Pet-rella, Andreas Pick, Nuno Ribeiro, Paulo Rodrigues, Elisa Tosetti,Rudi Vander Vennet and participants at the XX International TorVergata Conference on Money, Banking and Finance, the 24th Aus-tralasian Finance & Banking Conference and at seminars at Bancode Portugal, Bank of England and Banco Central de Chile for helpfulcomments and suggestions. The views expressed in this paper asthose of the author and do not necessarily reect those of Bancode Portugal or the Eurosystem. All remaining errors those of theauthor.

Appendix A. Derivation of Portfolio Distance-to-Default

The Portfolio Distance-to-Default treats the portfolio of compa-nies in the sample of each Supersector as a single entity, thus theMerton model assumptions still apply and the calculation methodis the same as in the case of a single company with some practicalconsiderations, especially about the difference between the ap-proach in this paper and other applications in the literature, suchas De Nicol et al. (2005), De Nicol et al. (2007), Echeverraet al. (2006), Echeverra et al. (2009), and Gray and Malone(2008). As a result, given the three principles in Contingent ClaimsAnalysis (CCA) mentioned in Section 2, the economic value of theportfolio (represented by its assets, A) is the sum of its risky debt(D) and equity (E). Since equity is a junior claim to debt, the formercan be expressed as a standard call option on the assets with strikeprice equal to the value of risky debt (also known in the literatureas distress barrier or default barrier).

E maxf0;A Dg A:1Given the assumption of portfolio assets distributed as a Gener-

alized Brownian Motion, the application of the standard BlackSholes option pricing formula yields the closed-form expressionof equity E as a European call option on the portfolios assets Aat maturity T:

E ANd1 erTDNd2 A:2where r is the instantaneous rate of growth of the portfolio assets,generally approximated by the risk-free rate, and N() is the cumu-lative normal distribution. The values of d1 and d2 are expressed as:

d1 ln AD r 12r2A T

rAT

p A:3

d2 d1 rAT

pA:4

where rA is the is portfolios asset volatility. The Merton model usesan additional equation that links the former to the volatility of theportfolios equity rE by applying Its Lemma:

ErE ArANd1 A:5The Merton model uses Eqs. (A.2) and (A.5) to obtain the

implied portfolios asset value A and volatility rA, which are notobservable and must be estimated by numerical methods. Theportfolio equity volatility rE enters as initial value of market valueof rA in the iteration. The growth rate of the assets in the portfoliois proxied by risk-free interest rate r as in Gropp et al. (2006) andmost papers in the literature. Once a numerical solutions for A andrA are found, the Portfolio Distance-to-Default DDP T periods aheadis calculated as:

DDP lnAD

r 12r2A TrA

T

p A:6

M. Saldas / Journal of BankingD is the total value of the portfolios risky debt or distress barrierand is obtained by adding up the individual distress barriers acrossthe P constituents in each Supersector, i.e. D PPj1Dj.The Pesaran CD statistic, CDP, was developed in Pesaran (2004)and is used for panels where series may be either stationary or con-tain unit roots. This CD statistic shows good properties with dy-namic panels but has also a caveat. Since it involves the sum ofpairwise correlation coefcients instead of the sum of squared cor-relations, the CDP statistic might miss out CD where there are alter-nating signs of correlations in the residuals. This statistic takes thefollowing form and distribution.

CDP

2TNN 1

s XN1i1

XNji1

qij!dN0;1 B:2

B.2. Unit root tests

In addition to the IPS test, cross-sectionally augmented IPStests (CIPS) (Pesaran, 2007) are applied. This test allows for indi-vidual unit root processes and for different serial correlation prop-erties across units. It is more suitable in the presence of cross-section dependence in the series, since IPS may lead to spuriousinference. The CIPS test statistic is computed using the average ofthe individual pth order cross-sectionally Augmented DickeyFuller

PTThe two statistics of cross-section dependence (CD) in paneldata used in the paper are based on pairwise correlation coef-cients, qij, of regressions residuals.23 The average of cross-correla-tion coefcients, q, is applied to provide a rst assessment at adescriptive level.

q 1NN 1

XN1XN1qij B:1In this paper, the calibration of Portfolio Distance-to-Default dosolve Eqs. (A.2) and (A.5) to obtain rA and A in each Supersector,hence the equity market value of the portfolio, E PPi1Ei, is ob-tained directly from the reference Supersector index on a daily ba-sis, and the equity volatility rE is obtained from index optionimplied volatilities. As a result, Portfolio Distance-to-Default donot only capture covariances in rA. See Saldas (2012) for more de-tails in the case of banks in Europe.

Appendix B. Cross-section dependence and panel unit root testsplied assets Aj, thus A PP

j1Aj and it aggregates the individualestimates of asset volatilities using a asset return based covariancestructure, r2 PP Pp wjwkrjk, where rjk is the asset returnand in general is proxied by a weighted average of individual rjfrom government bond yields of each companys home market,i.e. r PPj1wjrj. The individual weights wj are obtained fromestimates of implied assets A, thus wj AjA . In this paper, r isproxied by the Eurozone synthetic 10-year government bondyield.

The remaining terms in (A.6), namely the portfolio asset volatil-ity rA and the value of the portfolio assets A, should be in principleobtained as in the case of individual companies, solving the system23 qij qji t1u^it u^jtPT

t1 u^2it

q PTt1 u^

2jt

q , where u^it and u^jt are residuals from Eq. (1) orindividual series ADF(p) or cross-sectionally augmented ADF(p) regressions, CADF(p).

Equant (NL0000200889).

& F(ADF) regressions statistics (CADF). It assumes a single unobservedcommon factor, but is robust to other potential sources of CD, suchas spill-over effects (Baltagi et al., 2007). This assumed factor struc-ture is accounted for by adding the averages of lagged levels andrst-differences of the dependent variable to each standard ADFregression.

Dyi;t ai biyi;t1 Xpil1

ci;lDyi;tl d0izt mi;t;

i 1; . . . ;N; t 1; . . . ; T B:3

where zt yt1;Dyt;Dyt1; . . . ;Dytp 0. The joint asymptotic limit of

the CIPS statistic is nonstandard and critical values can be found inPesaran (2007) for various numbers of cross-section units N andtime series lengths T. Under the null hypothesis of non-stationarityagainst the possibly heterogeneous presence of unit roots across i,the CIPS statistic takes the following form.

CIPS 1N

XNi1

~ti B:4

where ~ti is the t-statistic associated to b^i in CADF equations.

Appendix C. Sample selection methodological notes

The analysis in the paper covers 12 out of 19 Supersectors, asclassied by STOXX. The list of Supersectors is found in Table 1.The companies included in a given Supersector Index are part ofthe STOXX Europe 600, which represents large, mid and small cap-italization companies across 18 European countries. Since the com-position list of the STOXX Europe 600 is revised periodically,mostly according to changes in market capitalization or relevantcorporate actions, the list of companies in each Supersector portfo-lio is revised accordingly and updated.

Since the most relevant changes take place at the bottom of theranking, some companies do not stay long in the Supersector Indi-ces and may only add noise to the series. Therefore, some smallcompanies were excluded from the sample under the assumptionthat their low weight in their respective index would not affectthe aggregation of company information by Supersector duringthe calibration of DD series. In addition, some companies arereclassied and should therefore be assigned to one Supersectoronly according to the time listed in a given supersector. See TablesE.1E.14 for individual cases. The list of exclusions from the sampleby Supersector is below.

Banks: Banque Nationale de Belgique (BE0003008019), BancaAntonveneta (IT0003270102), IKB (DE0008063306), Rolo Banca1473 (IT0001070405), Crdit Agricole le-de-France(FR0000045528), Emporiki Bank Of Greece (GRS006013007),Banco Pastor (ES0113770434), Marn Financial Group(GRS314003005), Depfa Bank (IE0072559994), Banca Fideuram(IT0000082963), Finecogroup Spa (IT0001464921), First Active(IE0004321422), KBC Ancora (BE0003867844).Oil & Gas: Fortum (FI0009007132), Royal Dutch Petroleum(NL0000009470, excluded due to incorporation in the UK witha primary listing on the London Stock Exchange), Enags(ES0130960018).Insurance: Fortis (BE0003801181), Nrnberger Beteiligungs(DE0008435967), Irish Life & Permanent (IE00B59NXW72).Utilities: SolarWorld (DE0005108401).Technology: SAFRAN (FR0000073272), Eutelsat Communica-tion (FR0010221234), Amadeus Global Travel Distribution

4546 M. Saldas / Journal of Banking(ES0109169013), Terra Networks (ES0178174019), InfogramesEntertainment (FR0000052573), Wanadoo (FR0000124158),Industrial Goods & Services: Linde (DE0006483001), Pirelli &Co. (IT0000072725), Gamesa (ES0143416115), Wendel Inves-tissement (FR0000121204), Q-Cells (DE0005558662), IndraSistemas (ES0118594417), Ackermans & Van Haaren(BE0003764785), Altran Technologies (FR0000034639), Aixtron(DE0005066203), CGIP (FR0000121022), Eurotunnel(FR0000125379), Snecma (FR0005328747), Rexel(FR0000125957), ASF (FR0005512555), Aurea (ES0111847036).Chemicals: Altana (DE0007600801), Degussa (DE0005421903),Celanese (DE0005753008).Food & Beverage: Parmalat Finanziaria (IT0003121644), IAWSGroup (IE0004554287).Media: RTL Group (LU0061462528), Premiere(DE000PREM111), Gestevisin Telecinco (ES0152503035), TeleAtlas (NL0000233948), Fox Kids Europe (NL0000352524).Healthcare: Fresenius Medical Care (DE0005785802), Alapis(GRS322003013), Altana (DE0007600801), Schwarz Pharma(DE0007221905), Omega Pharma (BE0003785020), Instrumen-tarium(FI0009000509).

Appendix D. Data sources

The structure of balance sheets varies by sector. Companies areclassied into the following sectors: Banks, Insurance Companiesand Industrials.

Balance-sheet Information, Obtained at quarterly/half-yearlyfrequence from Annual and Interim Reports. Total Assets. For banks, Bankscope (code 2025); for Insur-

ance Companies and Industrials, ThomsonWorldscope (codeWC02999A).

Short-term Liabilities: For banks, Bankscope (Deposits andShort Term Funding, code 2030); for Insurance Companies,Thomson Worldscope (code WC03051A); for Industrials,Thomson Worldscope (code WC03101A).

Total Equity: For banks, Bankscope (code 2055); for Insur-ance Companies and Industrials, Difference between TotalAssets (Thomson Worldscope, code WC02999A) and TotalLiabilities (Thomson Worldscope, code WC03351A).

Market Information. Sector Index Tickers. Thomson Datastream (codes DJESBNK,

DJESTEL, DJESEGY, DJESINS, DJESTEC, DJESAUT, DJESUSP,DJESIGS, DJESCHM, DJESFBV, DJESMED, DJESHTC).

Market Capitalization. Thomson Datastream (code MV). Price Indices. Thomson Datastream (code PI). Index Options Implied Volatilities: Thomson Datastream

(codes DJBXC.SERIESC, DJCXC.SERIESC, DJEXC.SERIESC, DJI-XC.SERIESC, DJTXC.SERIESC, DJAXC.SERIESC, DJUXC.SERIESC,DJIGC.SERIESC, DJCMC.SERIESC, DJFBC.SERIESC, DJMXC.SERI-ESC, DJHXC.SERIESC, DJBXC.SERIESP, DJCXC.SERIESP,DJEXC.SERIESP, DJIXC.SERIESP, DJTXC.SERIESP, DJAXC.SER-IESP, DJUXC.SERIESP, DJIGC.SERIESP, DJCMC.SERIESP,DJFBC.SERIESP, DJMXC.SERIESP, DJHXC.SERIESP).

Interest rates. Thomson Datastream (code EMBRYLD).

Macro-Financial Variables and Sector-specic Variables.Riverdeep Group (IE0001521057), Tiscali (IT0001453924),

inance 37 (2013) 45344555 VIX Volatility Index, VIXt: Chicago Board Options Exchange. Money Market Rate, R3Mt: Three-month Euribor Rate, ECB. Oil Price, OILt, Brent Crude 1-Month-Forward Price, ECB, level.

Euro Area Industrial Production Index, DPIt: ECB, Annual rate ofchange, working day and seasonally adjusted.

Euro Area Ination Rate, DCPt: ECB, HICP Overall index, Annualrate of change, Neither seasonally nor working day adjusted.

Price-Earnings Ratio, DPEt: Thomson Datastream (PE).Weighted average of PERs of index constituents, Annual rateof change.

Dividend Yield, DDYt: Thomson Datastream, Market-valueweighted average of individual DYs of index constituents,Annual rate of change.

Excess Returns, EXRETt: Monthly log excess return on eachindexs daily price return relative to the EURO STOXX 50 Index,as constructed in Campbell et al. (2008).

Return On Assets, ROAt: Bloomberg (ROA), Computed as theratio of net income to total assets and averaged for all membersof a given index. Monthly frequency.

Appendix E. Constituents by supersector

See Tables E.1E.14.

Table E.1Supersector constituents list Banks (BNK).

Name ISIN code Country Portfolio constituent

From: To:

1 Deutsche Bank DE0005140008 DE 31-December-01 31-October-092 BNP Paribasa FR0000131104 FR 31-December-01 31-October-09? Fortisb,c BE0003801181 BE 31-December-01 21-September-09? Banca Nazionale IT0001254884 IT 31-December-01 22-May-06

del Lavorob

3 Crdit Agricole FR0000045072 FR 18-March-02 31-October-09? Crdit Lyonnaisb FR0000184202 FR 31-December-01 19-June-034 Socit Gnrale FR0000130809 FR 31-December-01 31-October-095 UniCredit IT0000064854 IT 31-December-01 31-October-09? Capitaliab,d IT0003121495 IT 31-December-01 1-October-07? HypoVereinsbankb,e DE0008022005 DE 31-December-01 19-June-06? Bank Austriab AT0000995006 AT 24-October-03 5-December-056 Santanderf ES0113900J37 ES 31-December-01 31-October-09? ABN Amrog NL0000301109 NL 31-December-01 2-November-077 Dexiah BE0003796134 BE 31-December-01 31-October-098 Commerzbanki DE0008032004 DE 10-August-07 31-October-099 Intesa Sanpaoloj IT0000072618 IT 31-December-01 31-October-09? San Paolo IMIb IT0001269361 IT 31-December-01 2-January-0710 Natixis FR0000120685 FR 19-April-05 31-October-0911 BBVA ES0113211835 ES 31-December-01 31-October-0912 KBC BE0003565737 BE 31-December-01 31-October-09? Almanijb BE0003703171 BE 31-December-01 3-March-0513 Deutsche Postbank DE0008001009 DE 20-September-04 31-October-0914 Erste Group Bankk AT0000652011 AT 31-December-01 31-October-0915 Bank Of Ireland IE0030606259 IE 31-December-01 31-October-0916 Banca Monte dei IT0001334587 IT 31-December-01 31-October-09

Paschi di Sienal

17 Allied Irish Banks IE0000197834 IE 31-December-01 31-October-09

a Increase in share capital and free oat change on 19-May-09.b Takeover.c Also constituent prior to 21-June-04.d Formerly Banca di Roma.e Also constituent between 24-November-05 and 19-June-06 after takeover.f Increase in share capital due to takeover of Abbey on 16-November-04.g Takeover by Royal Bank of Scotland, Fortis and Santander.h Increase in share capital on 8-January-09.i Increase in share capital on 23-July-09.j Banca Intesa is the predecessor company. Increase in free oat on 19-April-04.k Increase in share capital on 31-January-06.l Increase in share capital due to takeover of Banca Agricola Mantovana and Banca Toscana on 31-March-03.

M. Saldas / Journal of Banking & Finance 37 (2013) 45344555 4547

Table E.2Supersector constituents list Banks (BNK) (cont.)

Name ISIN code Country Portfolio constituent

From: To:

18 Banco Popolarea IT0004231566 IT 2-July-07 31-October-09? Banca Popolare Italiana IT0000064300 IT 31-December-01 2-July-07? BP di Verona e Novara IT0003262513 IT 4-June-02 2-July-07? BP di Novara IT0000064508 IT 31-December-01 4-June-02? BP di Verona IT0001065215 IT 31-December-01 4-June-0219 UBI Bancab IT0003487029 IT 2-April-07 31-October-09? Banca Lombarda IT0000062197 IT 31-December-01 2-April-07

e Piemontese? BP di Bergamo IT0000064409 IT 31-December-01 1-July-03? BP Commercio e Industria IT0000064193 IT 31-December-01 1-July-0320 Banco Popular Espaol ES0113790531 ES 31-December-01 31-October-0921 Anglo Irish Bankc IE00B06H8J93 IE 31-December-01 26-January-0922 National Bank Of Greece GRS003013000 GR 31-December-01 31-October-0923 BCP PTBCP0AM0007 PT 31-December-01 31-October-0924 Raiffeisen International AT0000606306 AT 20-June-05 31-October-0925 Banco Sabadelld ES0113860A34 ES 31-December-01 31-October-0926 EFG Eurobank Ergasias GRS323013003 GR 31-December-01 31-October-0927 Banco Espirito Santo PTBES0AM0007 PT 31-December-01 31-October-0928 Mediobancae IT0000062957 IT 31-December-01 31-October-0929 Alpha Bank GRS015013006 GR 31-December-01 31-October-0930 Bank Of Greece GRS004013009 GR 14-August-03 31-October-0931 Bankinter ES0113679I37 ES 31-December-01 31-October-0932 BP dellEmilia IT0000066123 IT 31-December-01 31-October-09

Romagnaf

33 Piraeus Bankg GRS014013007 GR 31-December-01 31-October-0934 BP di Milano IT0000064482 IT 31-December-01 31-October-0935 Banco BPIh PTBPI0AM0004 PT 31-December-01 31-October-0936 Banca Carige IT0003211601 IT 20-June-05 31-October-0937 Pohjola Bank FI0009003222 FI 18-September-06 31-October-0938 Banco de Valencia ES0113980F34 ES 23-June-03 31-October-0939 BP di Sondrioi IT0000784196 IT 31-December-01 31-October-0940 Credito Valtellinese IT0000064516 IT 22-December-08 31-October-09

a Merger on 2-July-07 between BP Italiana (IT0000064300, formerly BP di Lodi) and BP di Verona e Novara (IT0003262513, merger between BP di Novara and BP di Veronain June 2002).

b Merger on 2-April-07 between Banche Popolare Unite (predecessor) and Banca Lombarda e Piamontese. The former was formed by the merger between BP di Bergamo, BPCommercio e Industria and BP di Ruino e di Varese (no data) on 1-July-03.

c Previous ISIN IE0001987894, nationalized.d Increase in share capital on 15-March-04.e Also constituent prior to 23-December-02.f Temporary deletion between 22-December-03 and 10-September-09.g Increase in share capital on 2-January-04.h Also constituent prior to 24-March-03.i Temporary deletion between 22-December-03 and 21-September-09.

Table E.3Supersector constituents list Telecommunications (TLS).

Name ISIN code Country Portfolio constituent

From: To:

1 Deutsche Telekom