![Sample our Mathematics & Statistics journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5943/TOC-Subject-Sample-2021-Mathematics-Statistics.jpg"})
Abstract
We employ a machine learning approach to build a European sovereign risk stratification using macroeconomic fundamentals and contagion measures, proxied by copula-based credit default swap (CDS) dependencies over the period 2008–2017, for France, Germany, Greece, Ireland, Italy, Portugal, and Spain. By adopting a recursive partitioning strategy, we detect specific risk zones varying from safe to high risk based on key predictors, and we construct their specification by assigning specific risk thresholds. While key macroeconomic fundamentals such as Debt/GDP and the unemployment rate remained the same and maintained the same risk thresholds during the sub-periods 2008–2013 and 2013–2017, the CDS spreads contagion dropped significantly over the post-Quantitative Easing years, lowering the corresponding risk thresholds. We estimate an impact on CDS spreads approximately of basis points in the period 2013–2017 due to contagion mitigation.
Acknowledgements
The authors would like to thank Pierluigi Balduzzi, Andreas Pick and the participants of Consortium for Systemic Risk Meeting Analytics Semi-Annual Meeting (2014) held at the Massachusetts Institute of Technology in Cambridge, Massachusetts, the SYRTO Conference on Systemic Risk, organized by the Finance and Econometrics departments of the VU University Amsterdam and held in Amsterdam on 4–5 June 2015, the SYRTO Final International Conference, organized by the Université Paris 1 Panthéon, Sorbonne, and held in Paris on February 2016, the 2016 RiskLab/BoF/ESRB Conference on Systemic Risk Analytics held in Helsinki for their beneficiary comments.
Disclosure statement
No potential conflict of interest was reported by the authors.
Supplemental data
Supplemental data for this article can be accessed at https://doi.org/10.1080/14697688.2018.1562197.
Notes
1 Caporin et al. (Citation2018) study the same group of countries adding also a non-EMU country, UK. We stay focused to countries which are EMU members in order to ‘contain’ the currency risk.
2 A short description of the CDS and the euro area CDS market as of 2012 can be found at ECB (Citation2012).
3 German Federal Financial Supervisory Authority prohibited naked buying of CDS based on euro-denominated government bonds on 19 May 2010, and little after the European Parliament passed the ban on same naked CDS on 1 December 2011. The regulation was in effect on 1 November 2012.
4 See Oehmke and Zawadowski (Citation2017) for a comprehensive investigation on the motivations for trading in CDS markets and, more broadly, on the economic function played by these markets.
5 5-yr sovereign CDS are the most liquid contracts traded in the market, while others maturities, like 6 months, 1- 2- 3- 4- 5- 7- 10- 15- 20- 30- years, are also available.
6 For the Greek CDS spread only 1414 quotes out of 1505 are available over the crisis period. In fact, since the end of 2009 and the start of 2010, the Greek CDS spread traded at extraordinary levels. On 22 April 2010, the Greek CDS spread reached the 700 bps and afterwards the Greek CDS contracts have been converted to up front, as the protection seller had to pay up front a premium to the protection buyer, as the elevated Greek CDS basis mapped the sentiment of the market participants that the situation would be unsustainable in the long term. In the meantime, the volume of dealers quoting Greek CDS was not sufficient for the key providers of derivative pricing information to determine an official live price.
7 Their empirical study explains interesting contagion effects in the Asian and Mexican crises.
8 We develop a MATLAB code to implement the method. Description of the MATLAB codes used are in the Online Appendix https://www.researchgate.net/publication/322925040_Online_Supplementary_Material_for_Sovereign_Risk_Zones_in_Europe_During_and_After_the_Debt_Crisis.
9 For each predictor, the procedure starts by considering all the possible binary splits obtained from the starting node (the first node is the sample as a whole) considering each realization as possible splitting value (from min to max) and calculating the corresponding reduction in the MSE (node impurity) generated by the split from the starting node to the sub-nodes; the best split is the value for the predictor which attains the maximum reduction in node impurity. The procedure is run for each predictor at each split and the selected variables with corresponding threshold are those that most reduce the loss function in each partition. See Hastie et al. (Citation2009) for more technical details.
10 It was heuristically shown that the accuracy of random forest converges around 3000 trees.
11 Predictor values are standardized to avoid different scale orders, and depicted by a rectangular tiling of different colors within the data matrix. Heatmaps also compute two hierarchical cluster analyses: one is being implemented on the contagion-based variables and country-specific fundamentals and the other one on the observations (more precisely, on the countries in correspondence to different years), thereby realizing two dendrograms appended on the x- and y-axes, respectively (Ling Citation1973). In doing this, the columns and the rows of the data matrix are permuted based on column and row means. In this way, similar values are placed near each other according to the clustering algorithm used in the analysis (Sneath Citation1957).
12 While deflation complicates debt sustainability and is related to large contractions, which in turn are both related to sovereign risk, it is more appropriate to examine the shocks driving down inflation, as the risk of deflation assessment demands the identification of the nature and the persistence of the determining factors and, in particular, the degree to which inflation developments can be attributed to supply-side or demand-side forces. As discussed in ECB (Citation2014), the overall price index may turn negative for a short period on the back of transitory supply-side shocks, such as commodity price movements, as it occurred in the euro area and in other countries in 2009. However, a period of negative annual inflation does not in itself imply deflation, in a meaningful economic sense, unless the price declines become generalised and entrenched in inflation expectations. For instance, if longer-term inflation expectations remain stable, the ebbs and flows in commodity prices are bound to exert only transitory effects on inflation. Furthermore, it is crucial to disentangle the impact of supply-side shocks resulting from structural reforms, which may have implications for inflation developments over the policy-relevant horizon. While structural reforms may initially lead to downward pressures on inflation rates, reflecting also supply-side improvements in the economy, inflation can be expected to pick up over time as aggregate demand gradually recovers. In the case of the euro area, one should not confuse relative price adjustments with overall changes in the price level: to speak meaningfully of deflation, the generalised and prolonged fall in the price level should be broadly based across countries. There is no risk of outright deflation as long as euro area HICP (Harmonised Index of Consumer Prices) inflation is in line with price stability. Negative inflation rates in individual countries may, on occasion, be consistent with the normal functioning of a monetary union, as they help to restore competitiveness, i.e. they may be symptomatic of supply-side induced relative price adjustments. Under these circumstances, a period of deflationary pressure may have been seen beneficial for Greek growth, thereby alleviating the sovereign risk, as it seems to prove the splitting rule on inflation of our regression tree.
13 The first country moved to other regimes was Spain in September 2008, followed by Greece in April 2009, Ireland in May 2009, Portugal in May 2010, and Italy in September 2010. Starting from 2011, only France and Germany remained in the safe zone until the end of the crisis period. These findings confirm the ‘wake-up call’ phenomenon in the Eurozone (Goldstein Citation1998), since markets ignored deteriorating fundamentals during times of non-crisis and became highly sensitive upon the onset of crisis.
14 Portugal applied for financial support in April 2011.
15 Ireland is already in financial support program since November 2010.
16 For e.g. when we comment the high risk zone for the Euro sovereign risk area and observe the time varying composition of the nodes.
17 By confronting the five trees with the original one, we observe a little change between the Kendall's τ with Euro Other Financials and the Kendall's τ with Euro Banks, which is selected by the trees excluding Ireland and excluding Spain.
18 Council Regulation (EU) No 1024/2013.
19 Regulation (EU) No 1022/2013 of the European Parliament and of the Council on October 22, 2013.
21 To confirm this conjecture, for the observations falling within the final nodes with expected CDS levels of 105, 303, 168 bps, we regressed the daily CDS onto exports/GDP, , while controlling for the GDP growth and with D a dummy, which takes the value of 1 when exports/GDP is greater than
and 0 otherwise. With an
of 0.484, the coefficient estimates (all statistically significant at 0.001 level) were: (1)
for exports/GDP; (2) 315.578 for
; (3) 1267.441 for GDP growth. Hence, the positive sign for
confirms the conjectured about the non-linear relationship.
22 See for example, ECB's Executive Board Members Benoît Coeuré speech on 2 September 2013, http://www.ecb.europa.eu/press/key/date/2013/html/sp130902.en.html