System-wide risk and systemic importance: Incomplete review of metrics and data Nikola Tarashev, Bank for International Settlements Cambridge, 25 September 2014 Restricted This presentation does not necessarily reflect the views of the BIS, the BCBS or the BCBS Secretariat. Restricted Roadmap 1. Two competing metrics of aggregate risk: VaR and ES a. have different statistical properties b. are suited for different purposes 2. Systemic importance: what if data were not an issue? a. different measures for different purposes b. common misunderstandings 3. Data: an important driver of analyses of system-wide risk and systemic importance Restricted 3 Choice of metric: rarely motivated by objectives Two types of metrics: for portfolio risk & system-wide risk. Quantile-based: e.g. Value-at-Risk (VaR) - robust estimation - elicitable, if data sample is known Tail expectations: e.g. Expected Shortfall (ES) - coherent: well-defined capital optimization problems - limits arbitrage But these metrics serve very different purposes VaR: attain an acceptably low probability of distress ES: prepare for the fallout of distress Restricted 4 Limit the probability of failure: VaR Tail of loss distribution (a bank’s assets) VaR = default point Losses absorbed by capital Restricted 5 Prepare for / insure against costs of failure: ES Tail of loss distribution Losses absorbed by capital ES DI premium Restricted 6 Regulatory arbitrage VaR: incentives for banks to hide behind a quantile Tail of loss distribution VaR = default point Restricted 7 From system-wide risk to systemic importance Shapley value: an allocation methodology & disciplining device Satisfies appealing criteria. Captures how the interaction of players creates risk Tarashev, Borio and Tsatsaronis (2010) A popular alternative is a special case Aumann-Shapley value, applied to ES or VaR = marginal ES, popularised by Acharya et al (2009) Another popular alternative falls in a different category: CoVaR, popularised by Adrian and Brunnermeier (2008) Restricted 8 Contribution approach (CA) vs. participation approach (PA) 𝑆ℎ𝑉 𝑖; 𝑁, 𝜌 = 1 𝑛 1 𝑛 𝑛𝑠 =1 𝑐 𝑛 𝑠 𝑆∌𝑖 𝑆 =𝑛𝑠 𝜌 𝑆∪ 𝑖 −𝜌 𝑆 𝜌𝐶𝐴 𝑆; 𝑚 = 𝐸 𝐿 𝑆 𝑒 𝑆, 𝑞; 𝑚 𝜌𝑃𝐴 𝑆; 𝑚 = 𝐸 𝐿 𝑆 𝑒 𝑁, 𝑞; 𝑚 Restricted 9 Participation in tail events vs. contribution to systemic risk Participation approach charge premia for insuring against losses in tail events Contribution approach penalise banks for raising the risk in system Does it really matter which one you choose ? Restricted 10 Numerical setup Bank-level losses: data on non-equity liabilities 𝐿 𝑖 = 𝜑𝑖 𝐿 𝑖 Correlated defaults data on marginal PDs & asset correlations 𝐿 𝑖 = 1 𝑖𝑓 𝑟𝑖 𝑀 + 1 − 𝑟𝑖2 𝑍𝑖 < 𝛷−1 𝑃𝐷𝑖 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 Derive distribution of system-wide losses VaR or ES Restricted 11 Participation vs. contribution VaR example Low common-factor loading High common-factor loading CA PA CA PA Group A 34.34% 0.0% 28.15% 100% Group B 65.66% 100% 71.85% 0.0% Total VaR 26 26 28 28 (100%) (100%) (100%) (100%) Note: Each panel refers to a different banking system. VaR is measured in cents per dollar exposure to the system. The first two rows report the systemic importance of each group of banks, as a share in system-wide VaR and as measured by the approach specified in the column heading. For all banks, 𝑃𝐷 = 0.27%. The number of banks in each groups and their respective sizes (as a share in the total system size) are as follows: 𝑛𝐴 = 5, 𝜑𝐴 = 0.07, 𝑛𝐵 = 5, and 𝜑𝐵 = 0.13. The common-factor loadings are 𝑟𝐴 = 𝑟𝐵 = 0.60 (low) and 𝑟𝐴 = 𝑟𝐵 = 0.724 (high). Table 1 Similar message with ES Restricted 12 Need to remain mindful of causality A measure of system-wide impact (CoVaR idea): quite useful from a policy perspective in practice: E(systemic distress | individual distress) Think of the stylized banking system from above. To fix ideas: different PDs; identical exposures to common risk factor, etc. Which bank is designated as most systemically important? The bank with lowest PD Intuition: if the safest bank is in trouble because of common risk factor, other banks must also be in trouble. Spurious causality misleading message Restricted 13 Data availability: a factor in the metric design Data on interlinkages in the system: Interbank network is a key driver of system-wide risk. Different approaches to measuring systemic importance treat interbank borrowers and lenders differently. Drehmann and Tarashev (2013) Restricted 14 IB lender vs. IB borrower: the approach matters Restricted 15 Price data Rely on markets to convey information about interconnectedness, in reduced form. Data are rich: despite few direct observations in the tail of interest … EVT techniques possible Tarashev and Zhou (2013) Restricted 16 Empirical setup Sample of 50 largest banks with CDS data Data: Balance sheet data banks’ size CDS spreads LGD, tendency to default with others Moody’s KMV EDFs PDs Systemic event ≡ when losses exceed 15% of system size Restricted 17 Humble even with price data Restricted 18 Bank characteristics and systemic importance Restricted 19 Cited papers Tarashev, Borio and Tsatsaronis (2010): “Attributing systemic risk to individual institutions”, BIS Working Paper 308. Drehmann and Tarashev (2013): “Measuring the systemic importance of interconnected banks”, Journal of Financial Intermediation, v. 22, iss 4. Tarashev and Zhou (2013), “Looking at the tail: price-based measures of systemic importance”, BIS Quarterly Review, June Restricted 20