A Globally Consistent Stress Testing Approach

This paper describes an approach for stress testing banks that is consistent across economies and geographies, in contrast to common “macro scenario” driven approaches. The latter would require economic scenarios to be both equally likely (in a probabilistic sense) and equally stressful (in a conditional loss sense) across countries in order to be comparable. The paper proposes a three-pronged approach for stressing bank solvency, which incorporates recalibrating pre-crisis Basel capital assumptions, adapting the BIS “expected shortfall” approach for securities, and using granular data for income haircuts. Loan losses are quantified using a simple “multiples” approach, starting from expected outcomes, which is derived from the pre-crisis Basel technical proposal. The approach is practical, can be more granular or conducted at a high level, depending on data availability, and offers a simple way for regulators, investors or risk assessors to compare and contrast stresses in different banking systems. Of the eight bank defaults recorded globally during 2017, this approach would have given a better “rank ordering” for seven of them, indicating the approach adds value to traditional solvency metrics.

Banking Authority (2014) have prescribed recent large-scale regulatory stress tests in the US and the European Union.
A key feature of these exercises was to shift away from previous "value at risk" (VaR) approaches. VaR analysis can offer some substantial advantages, including its practical viability and conceptual attractiveness (Kupiec, 1998) and the ability to contrast multiple models and calibrations (see for example Alexander and Sheedy, 2008). But with its decline, stress tests instead became increasingly reliant on a form of scenario analysis: taking unexpected (downside) macro scenarios and estimating how those impact, via loan and securities losses, on bank capital. Lopez (2005) was one of the first to note that this mechanism would link losses to "specific and concrete" events; Jokivuolle, Virolainen and Vähämaa (2008) is one early post-crisis example of macro-driven stress testing. The popularity of these macro-driven tests extended to regulators, with some policymakers arguing they should replace the previous "Internal Ratings Based" (IRB) approach to risk-weighted capital (Tarullo, 2014).
However, there is no consensus here; Borio, Drehmann and Tsatsaronis (2012) note that macro-driven stress tests are not suitable as early warning devices and would benefit from complementary information.
Concerns have also been raised about the appropriateness of the modelling framework that links macroeconomic data to bank loss rates. To start with, these frameworks were often similar to classic macro modelling and hence focused on the middle of the distribution of losses (see for example Bunn et al., 2015;and Miani et al., 2012), rather than the tail. In some instances, researchers recognized this by proposing adjustments to estimated models, for instance in Buncic and Melecky (2012). However, regulators may also have responded by picking unusually stressful scenarios in their macro-based stress tests (Ellis, 2017). More recently, there has been renewed focus on flexible models that allow these relationships to change as the analysis moves into the tail of the distribution (Covas, Rump, & Zakrajsek, 2013). These quantile models, as introduced by Koenker and Hallock (2001), may offer a better guide to stressed outcomes, but they are not yet widely employed.
However, macro-driven stress tests encounter further challenges when there is a need to compare and contrast banks in different regions or jurisdictions. Applying the same degree of stress across countries is far from simple for typical macro-driven stress tests. An assumed recession that decreases GDP by, say, 3% may not be as probable today in the US as in Indonesia; conversely, an equally probable scenario (say, with a 10% probability) may well entail a deep recession in one country and a more mild slowdown in another. Similarly, tying all countries to a single shock that is transmitted globally will not be equally stressful for every country that is affected, as the exposure to the shock and the nature of the transmission mechanism will vary from country to country.
In light of these challenges, the approach described in this paper is deliberately different: and, consequently, is not intended to be directly comparable with macro-driven stress tests. In large part, this reflects the aims and of and context in which the approach was developed, with comparability and consistency across different countries and regions being more important than country-specific risks or www.scholink.org/ojs/index.php/rem Research in Economics and Management Vol. 4, No. 3, 2019 155 Published by SCHOLINK INC. scenarios (which by their very nature will be heterogeneous). However, these discrepancies do not necessarily imply differences in judgments about the relative strength or viability of a bank under stressed conditions. Ultimately, as with other stress tests, this approach still aims to analyze banks' resilience under stressed conditions against a group of peers, in order to uncover potential weaknesses in the financial system. This paper therefore describes an approach to stress testing that does not rely on downside macroeconomic scenario and, unlike most macroeconomic-driven stress tests, allows consistency and comparability of results across banks within a jurisdiction and across different jurisdictions. The rest of this paper describes each of the components in this approach that determine the stressed capital ratio.
Section 2 deals sequentially with loan losses, stressing banks' income, and an approach for security losses. Section 3 then shows the results of this approach for over 70 banking systems, highlighting those more vulnerable to stressed conditions and those more resilient. Finally, the discussion in Section 4 concludes.

Loan Losses: Starting from the Expected Case
While point forecasts represent the average or most likely outcomes given a set of macroeconomic and industry conditions, stress tests literally represent unexpected developments. As such, it is possible to draw parallels between the two: and indeed to express stressed loss rates or stressed default rates as a "multiple" of expected rates. The higher the multiple, the bigger the increase from the expected to the stressed case.
In order to exploit this link -in the context of loan losses -the analysis needs to start from an expected case. Given data limitations in many countries, one simple approach is to focus on system-level trends in asset quality, as measured by the aggregate Non-Performing Loan (NPL) ratio. As shown in past work (see for instance Buncic & Melecky, 2012, and Moody's, 2014aand 2014b, it is possible to model system-level trends in asset quality and default rates based on expected developments in the economic and financial environment, where macroeconomic data -such as real GDP, unemployment, inflation, and the exchange rate -are used to obtain forecasts for the NPL ratio. However, econometric techniques and models differ. In general, NPL series tend to be relatively short for most banking systems (most of which are in emerging or developing economies); as a consequence, panel models may be needed to exploit cross-country patterns in the linkages between NPLs and macroeconomic variables. Wherever there is greater data availability, country-specific models can be estimated. But in either instance, the outcome from this approach is a set of projections for the aggregate or total NPL ratio in the banking system. In turn, this can be transformed into a Probability of Default (PD) given an assumption about the write-off rate of NPLs, using a simple "law of motion" approach (see Buncic & Melecky, op cit System-level profiles for NPL ratios can easily be transformed into NPL ratios -or PDs -for individual loan types, provided either disaggregated data on these categories are available, or assumptions are applied about the distribution of loans and asset quality. In some instances where data are more plentiful, it is possible to build specific forecasting models for individual loan types.
One important point here is that -by and large -input data are not adjusted for idiosyncratic factors above and beyond routine adjustments. Several private agencies collect arrays of balance sheet and accounting data from rated issuers, which are then adjusted to provide standardized metrics across different geographies and jurisdictions (see for instance . In principle, these input data should only be adjusted if there are clear and unambiguous grounds for doing so. For instance, an implicit assumption is often that underwriting standards are broadly consistent over time -or at least that any change is evident in metrics such as default rates or non-performing loans. This assumption is important because making sensible ad-hoc adjustments for as yet unobserved structural changes is difficult.
It is also important to note that the granularity of the data typically varies significantly from country to country and region to region. As such, simplifying assumptions are often needed. In the case of aggregate figures on, for instance, loan-to-value ratios, this implies that some banks in a system will be relatively "penalized" by using an average figure, while others will implicitly benefit. But provided the aggregates are broadly correct, these differences should average out across the system as a whole.
Where assumptions are needed for granular loan loss rates, these can be linear or non-linear provided they are consistent with the aggregate loan profile. It is also likely that individual banks may well see different PDs within the same financial system, as a result of bank-specific factors such as the quality of underwriting. To the extent to which these factors are evident in differential default data, they can also be incorporated.
This approach can then be used to calculate default rates on different loan types -and for different banks -that are consistent with a central macroeconomic scenario. However, in order to calculate loan losses estimates of loss given default (LGD) are also required.
There is a rich body of literature around LGD estimates, including many loan-specific estimates. Based on a survey of over 70 such estimates, there can be considerable variability in the appropriate LGD, depending on the type of loan (see Figure 1). A simple starting point is to use the median estimates shown; but wherever more country-specific data is available, that can also be incorporated. Similarly, differences in Loan-To-Value (LTV) ratios can also inform different LGDs for residential and commercial property.    In principle, these multiple curves can be used to generate stressed loss outcomes, obviating the need for a set of globally consistent macro scenarios for individual countries. However, they also suffer from obvious shortcomings, notably the poor performance of capital risk weights during the financial crisis.
This suggests that the original asset risk models proposed by Basel need recalibrating. Properly assessing this would require estimates of pre-crisis expectations for credit losses, which are not readily available. But two simple proxies are easy to construct: the first is based on the assumption that expected losses were equal to the pre-crisis series average; for the second, we can assume that expected losses followed a random walk, implying they would be the same as realized losses in the previous year.
Armed with these pseudo "real time" estimates of expected losses, it is possible to examine what the actual loss outturns during the crisis implied in terms of multiples, compared with the original Basel calibrations. Table 1 presents results for the US, where data are most granular and coverage is good. where the multiples observed during the crisis were much higher than the pre-crisis models suggested.
But this offers an obvious recalibration approach: parameters from the original Basel specification can be adjusted until the implied multiples match the observed outcomes at some given percentile. For instance, given the implied multiple observed during the crisis was as high as 8 for US real estate loans, but the 99th Basel-implied percentile was only around 6, the Basel approach can be recalibrated to generate in higher multiples for real estate loans.
This approach therefore offers a simple but consistent mechanism for constructing stressed loss rates on different lending types across banks in different systems.  Based on this assumption, the observed distribution can be used to generate income haircuts for banks, which can be applied in stress tests. In particular, depending on the desired degree of stress to applyand, consistent with the multiples approach to loan losses, this can be defined as a percentile of the distribution -an income haircut that is consistent with the observed distribution can be employed. For instance, in a "1 in 25" stress test, the income haircut would be informed by the 4th percentile of the distribution of changes in PPI.
In practice, this approach can be employed for different components of banks' income, such as net interest income and non-interest income, rather than focus on aggregate PPI. Importantly, however, income haircuts are not applied to trading income: this is covered under the securities stress approach, and hence would "double count" stresses if also applied here. Similarly, a simplifying assumption in several stress tests is that operating expenses are constant, and that the impact of management actions is limited to pre-announced measures. However, if these assumptions -or indeed the observed distribution of income changes -is not representative, then either the assumptions, or the distribution of income data, can be adjusted to inform different approaches. The main goal here is to demonstrate that the approach again offers a simple and consistent mechanism to consistently stress banks around the globe, based on the observed data.

Securities Losses: A Differentiated Approach
The third key component in this stress testing approach is to impose losses on banks according to their securities holdings. For many banks around the world, securities represent a relatively small component of the total balance sheet, compared with loans. But despite this, securities can play a significant role in stress tests.
In principle, there are three broad categories of securities holdings (and a residual "other" category).
The first are securities that are Held To Maturity (HTM). In essence, banks will only realize losses on these securities if they default. The second are securities on the trading book (TRD). And the third are securities that are available for sale (AFS). Given the different nature of these three groups, a differentiated approach for stress testing is required.
For HTM securities, where a published credit rating exists, it is simple to apply published loss rates associated with that published rating (see Moody's, 2016a). However, in a stressed scenario there would likely be some deterioration in ratings from their pre-stress levels. This risk can be incorporated using published transition matrices for ratings. For instance, for a two-year stress, and again focusing on a "1 in 25" event, the 4th percentile of rating transitions can inform potential deteriorations in credit quality.
Based on published data, in this instance that would be broadly consistent with a three-notch rating downgrade; so a "1 in 25" stress for bank with bonds rated Baa2 (on Moody's scale) would imply a downgrade to Ba2 (Table 2). This in turn corresponds to a published two-year (idealized) loss rate of or assumptions; for instance, that the average rating of corporate bonds held by a bank matches the average rating in the region or country where the bank is domiciled. Note. Transition probabilities are cumulative from the left-hand column to the right-hand one. The rating category that nests a 4% cumulative downside outcome (consistent with a "1 in 25" stress) is shown in bold italics.
The treatment of securities on the trading book (TRD) should necessarily be different. In principle, these are not securities that banks will necessarily hold for long periods of time, so imposing large credit losses that may not crystallize for the bank may be inappropriate. At the same time, the trading book is affected by market risk in a much more immediate fashion than securities that are held to maturity. The process for stressing these securities is an adaptation of the "expected shortfall" approach outlined by the Basel Committee (BCBS, 2016). Essentially, this approach estimates loss rates on securities for a given holding period, which are calibrated using losses observed in a severe preceding year. Using past data for equity and bond indices that cover the global financial crisis in particularwhich tends to represent the most severe 12-month period in recent history -loss rates for different types of securities can be calculated. Where data limitations are prohibitive and do not allow country-specific loss rates to be calculated, regional loss rates can be constructed, or loss rates from comparable countries can be used. Illustrative examples of loss rates are presented in Figure 4; further details on the expected shortfall approach are provided in Appendix 2. www.

Sourc
The th  losses, informed by crisis-era data and new analysis from regulators, this offers a different approach to stress testing that more readily allows for global comparisons. That, in and of itself, means the stress testing approach described herein can add value to existing exercises conducted around the world, both by policymakers and private institutions. Furthermore -although the sample of ex-post outcomes is very small so far -this global stress testing approach appears to add value to rank orderings of credit risk. As such, it offers a useful tool to practitioners and policymakers alike.