Basel II and III allow banks to use own default statistics when estimating regulatory parameters (risk-weights) for the capital adequacy ratio purpose. Bank inputs own risk estimates into the Vasicek model. It yields a distribution of credit losses. Regulator then requires a bank to take 99.9% quantile of such a distribution as a risk-measure (a riskweight). When saying regulator we mean any Central Bank (including Bank of Russia, but no limited to it) that allow local banks to run the described approach. Having being criticized for excessive conservatism, we reveal that it still underestimates credit risk. This comes from the newly discovered fact that the default correlation may tend to co-depend with the systemic factor (for instance, with the GDP growth rate), albeit originally such co-dependence was not considered. We use the US statistics on total loans defaults for 1985-2019 to evidence the finding. The credit loss underestimation thus at least exceeds by 11% the loss estimates using the maximum (100%) correlation with the systemic factor. Basel Committee, IRB, credit risk, Vasicek, systemic factor, default correlation, crisis simply, loans).
The Basel Committee on Banking Supervision (BCBS) sets the standards for banks in order to assure their solvability. One of key metrics is the capital adequacy ratio (CAR). It benchmarks the bank capital to the amount of risk-weighted assets (RWA). The committee requires banks via local supervisors to hold the ratio no less than a predefined minimum. It varies depending on the definition of capital. For instance, for the total capital the ratio should be no less than 10.5% of RWA [3]. Generally speaking, RWA is the product of a risk-weight (RW) and the amount of assets (put it
2020 Copyright for this paper by its authors.
estimate. The world 13 mostly largest out of 30 global systemically important banks (G-SIBs, [3]) run
IRB [4, p. 28]. We proxy all G-SIBs’ assets to the amount of total exposure metrics used by the
BCBS.2 The G-SIBs’ total assets then equal to EUR 50 trln in 2018. 69% of the G-SIBs’ credit risk
assets depend upon IRB [4, p. 28]. Credit risk stands for an average of 84% of the total banking risk
(risk-weighted assets, RWA) ( [5, p. 49] , [6, p. 62], [7, pp. 8, 66], [8, p. 73], [9, p. 31]). Let us
roughly equate banking RWA to the total banking assets. Then this yields us that IRB coverage was:
EUR 50 trln of total assets * 84% of credit risk * 69% of IRB coverage = EUR 29 trln in 2018
worldwide at least. We should remember that 2k banks world-wide run IRB [
The amount of total banking assets can be proxied as the amount of the broad money (cash and credit money created from cash). Latter one was 123% of the world GDP for 2018, whereas the world GDP equaled to USD 86 trln. The average exchange rate in 2018 was USD 1.18 for EUR 1. Then we have that the total banking assets were USD 86 trln * 123% / USD 1.18 for EUR 1 = EUR 90 trln in 2018. So far, we may conclude that the IRB coverage stands for EUR 29 trln / EUR 90 trln = 32% of the total banking assets world-wide, or 32% * 123% = 40% of the world GDP for 2018.
Alternatively, such asset volume stands for 40% of the world GDP. When we discuss credit risk underestimation, we suggest the reader to apply those estimated to the mentioned asset base to feel its importance.
To prove this we briefly remind the baseline Vasicek [
The baseline Vasicek [1] model assumes that systemic factor Y and the idiosyncratic one εi determine the asset return rAi of a bank (corporate) borrower, i.e.
(1) rAi = Y√Ri + εi√1 − Ri, (1) where rAi ∼ N(0; 1); N(0; 1) – the cumulative distribution function for standard normal (Gaussian) distribution with mean zero and variance of one; and N−1( ) is it’s inverse;
Y ∼ N(0; 1) – the systemic (common) factor that co-depends with asset value, i.e., (2) Cor(Ai; Y) = Ri (2) where Ri ∈ [0; 1] is the value of ‘asset correlation’. Please, pay attention that Vasicek did not use lower subscript ‘i' for the asset correlation value as he considered homogenous loan pool with infinite number of loans. It is the BCBS who added it later. For comfort, we wish to keep the subscript from start.
A notable fact is that Vasicek never provided an example of such a systemic factor Y. Neither the
BCBS does this [
εi ∼ N(0; 1) – the idiosyncratic factor that co-depends with asset value, i.e.,
(3) Cor(Ai; εi) = 1 − Ri (3) εi is independent both to the systemic factor, i.e., Cor(Y; εi) = 0, and to the other idiosyncratic factors, i.e., Cor(εi; εj) = 0 given ∀j: i ≠ j.
A default case is called when the asset value breaches the debt held by the borrower. This results in the following probability of not paying back the loan where (4) pi(PDi; α) = N (N−1(PDi)+N−1(1−α)√Ri) (4) √1−Ri 2https://www.bis.org/bcbs/gsib/gsib_assessment_samples.htm; https://www.bis.org/bcbs/gsib/hl_ind_since_2013.xlsx where ∝ is the confidence level (quantile of the systemic risk factor realisation). BCBS sets it to 0.1%, or 1 − α = 99.9%. For formula derivation and already embedded shortcomings’ discussion, please, refer to [3]. Several researchers claimed that such a level yields excessively high portfolio credit risk losses estimates [4, p. 221], [5, p. 93], [2].
Here we would like to briefly explain the essence of the above formula (4). When there is no correlation of asset value and systemic factor, i.e. Cor(Ai; Y) = Ri = 0, the probability of not paying back the loan equals to the marginal default probability, pi(PDi; ∝) = PDi. Otherwise, for the case of positive asset correlation, there is a ‘bonus’ to the marginal default probability depending upon the tightness of codependence R .
i Originally Vasicek [6] does not define R . We may consider that he deemed it constant.
i Nevertheless, the regulator (it can be any Central Bank, including, but not limited to the Bank of Russia) adjusts it by introducing the following dependence and it restricts the weight (asset correlation) to belong to the range of [2%; 30%], see Annex 1 in [3].:
(5) Cor(PDi; Ri) < 0 or equivalently Cor(Ri; εi) > 0 (5)
The reader may argue that such a drastic violation of the model assumption should imply the decision to ultimately abandon the model for the prudential use. We do share such reader’s opinion. However, the BCBS proceeds with such an adjustment. Even more, by 2020 we know that one third of the world banking assets’ credit risk evaluation depends upon this model.
Same time the regulator does not modify any assumptions referring to the correlation with the systemic factor. In fact, this may turn out to be wrong as we are to show below. Such a deficiency may lead to at least 10% credit risk underestimation. To see this let us examine real-world data in the next section.
To do so we need to define the asset correlation. As Penikas [4] shows, we should focus on the default correlation , where = 2. Then the default correlation may be evaluated as follows: (6) ri = ̅D̅V̅R̅a∙(r1(−D̅DR̅̅R)̅) (6) where ( ) is the historical variance of the default rate for the asset (exposure) class that incorporates the i-th borrower; ̅̅̅̅ is the mean default rate for the respective time period for the same asset class.
Thus the objective is to check the hypothesis whether there is a dependency of default correlation and the systemic factor Y. To remind, Basel II and III assumes it to be nil.
Default rate is the basis to develop the probability of default model according to [5]. That is why its dynamics reflects the default correlations that we are interested in. We tried to find the longest time series of default rates in terms of data points to verify our hypothesis. Publicly available rating agencies data of Moody’s and Standard and Poor’s dates back to 1981, but has yearly frequency. The data on the US total loans delinquency starts from 1985, but has quarterly frequency. Thus we chose to proceed with the US data. The corresponding default rate historical dynamics is given below. The highest level was demonstrated in the world financial crisis of 2007-09, second in height were the times post savings and loan crisis of 1980s in the USA.
e t a ltr u a f e d
1/1/1990 1/1/2000 1/1/2010 1/1/2020
We proceed to analyzing codependence with the systemic factor. We proxy the systemic factor with the gross domestic product (GDP) quarterly growth rate. We take notional amount of GDP for the US as it has the corresponding depth of time series (the data for the real GDP is much shorter).3 First, we wish to check whether there is a codependence of the systemic factor and the probability of default, or the default rate. The figure below rejects such a hypothesis. This coincides with the assumptions from [1].
0 g_GDP_1Q default rate
Second, we extract default correlation using the above mentioned formula (6) and benchmark it against GDP growth as a systemic factor. We use a rolling window of two years (eight quarters, 8Q) for both variables to have negative GDP growth rates. In addition, we check that lower window size does not capture the dependency we discuss (we present regression calibration for such a case below 3 URL: https://fred.stlouisfed.org/series/NA000334Q; code - NA000334Q. . p p , e t a r . r g Q 8 P D G . n i m o n for comparison). Larger window size results in significant data sample shrinkage and we do not consider it. . p p , ) . d n i w l. o r Q 8 ( . r o c . f e d
Date 1/1/1990 1/1/2000 1/1/2010
1/1/2020 nomin. GDP 8Q gr.rate, pp.
def. cor. (8Q rol. wind.), pp.
We may take a look at the dependence in the scatter plot format in addition to the time perspective. The below figure shows that the negative dependence is mostly driven by the ‘outliers’ observed during economic crisis times.
0 2 5 1 0 1 5 0
To model the observed dependence we run the descriptive regression model and present its outcome in Table 1 below. The model adjusted R-squared is 46%. Using a shorter window (e.g. one quarter) does not reveal the dependence of default correlation and systemic factor, see Table 2.
The resulting residuals’ distribution from the model for the 8Q-window is close to the normal one, see Figure 5.
y it s n e D 0 1 8 6 4 2 0 -.2 -.1 0 Residuals .1 .2 .3
Added-value plots (avplots in Stata software) confirm strong factor significance, see Figure 6. ) X | R ( e
Our finding is particularly striking for the following reason. The default correlation by definition
implies that there is a concentration of two event types: defaults and non-defaults. Defaults occur
mostly in crisis and non-defaults take place in normal times. When the default correlation rises in
crises, this means that the defaults are observed even more often than expected against the presence of
the positive default correlation and even much more compared to the zero default correlation. Such a
situation of default correlation rise in crisis was coined by Longin and Solnik [6]. They seem to be the
first, to the best of author’s knowledge, to demonstrate that equity quotes tend to fall in a more
synchronous manner than they tend to rise. This gives the basis to apply Clayton copula for modeling
joint probability distributions [7]. Longin and Solnik studied the consequences of the Asian 1997
crisis. Comparable finding is available in Andrievskaya and Penikas [
As of now, we see that we need to incorporate two stylized facts into the [1] model that were unforeseen from start: 1. Positive correlation of the idiosyncratic factor εi and the default (asset) correlation. This is a requirement from the regulator, see formula (5). 2. Negative correlation of the systemic factor Y and default (asset) correlation, see formula (7).
To amalgamate the two requirements we suggest the following modification of the asset distribution formula: (8) Ai = Y ∙ i + εi ∙ (1 − i), (8) where i = (εi − Y) is the normal distribution cumulative density function. For the lowest values of Y, the weight is the largest reflecting positive co-dependence. The lowest values of εi (the higher marginal default probabilities) are assigned close to zero weight all other things being equal.
The figure below presents the distributions for the asset value from the baseline [1] model (case A) and modified version from formula (8) (case B). The key finding is that the distribution goes to the left, i.e. lower values of the asset value are observed more often implying more frequent cases of not paying back the loan. Part II of the below figure compares the modified model to the worst case as it seemed before, i.e. to asset correlation of R=100%. In fact, it implies normal distribution of asset values reflecting the Gaussian distribution of the systemic factor.
However, the modified model from formula (8) incorporates the observed negative dependence of default correlation and systemic factor. This means that the asset value goes down more often either when systemic factor has negative values with higher weight, or when the marginal default probability is high and it is also assigned higher weight compared to the baseline.
(I) R=20% (II) R=100% 400 y300 c n eu200 q e r F100
More precisely, the table below compares the quantiles of the loss distributions. It shows that for the mean values of asset correlation imposed by the regulator (e.g. R=20%), the loss underestimation is around 30% where as it is at least 10% when the most severe case of R=100% is considered (please, check,
Table 3).
Thus, the finding of the paper suggests that 100% asset correlation is not the worst case for any longer.
Basel IRB credit risk regulation uses a model that combines the impacts of systemic and idiosyncratic factors. The model is often blamed for over-conservative credit risk treatment as it requires systemic risk realization at the 99.9% confidence level [2]. However, no one before us has shown that there is statistically significant negative dependence of systemic risk factor realization and default correlation. By accounting for this finding we arrive at the credit loss distribution that is shifted to the left, i.e. to the more severe losses domain. This implies that actual losses might occur even larger than those prescribed by the 99.9% confidence level. It means such a model de facto is not as conservative as claimed before. Path to model adjustment is suggested together with the impact assessment. We find that the credit risk underestimation is at least 10% if we consider the worst case of 100% asset correlation and is up to 30% when benchmarked to the current asset correlation levels prescribed by the BCBS [8].
The author acknowledges two anonymous reviewers for their useful comments that helped improve the paper.
The paper was prepared within the framework of the Basic Research Program at the National Research University Higher School of Economics (HSE) and supported within the framework of a subsidy by the Russian Academic Excellence Project `5-100'.
Opinions expressed here are solely those of the authors and may not reflect those of the affiliated institutions.
[1] O. Vasicek, "The Distribution of Loan Portfolio Value," Risk, pp. 160-162, December 2002. [2] Zimper, "The Minimal Confidence Levels of Basel Capital Regulation," Journal of Banking
Regulation, vol. 15, no. 2, pp. 129-43, 2014. [3] BCBS, "Basel III transitional arrangements, 2017-2027," 07 December 2017. [Online].
Available: https://www.bis.org/bcbs/basel3/b3_trans_arr_1727.pdf. [Accessed 10 February 2020]. [4] FSB, "2019 list of global systemically important banks (G-SIBs)," 22 November 2019. [Online].
Available: https://www.fsb.org/wp-content/uploads/P221119-1.pdf. [5] Bank of Russia, "Abouth systemically important banks definitions and approaches to their regulation," 23 January 2020. [Online]. Available: http://www.cbr.ru/content/document/file/98619/consultation_paper_200123.pdf. [6] BCBS, "Regulatory Consistency Assessment Programme (RCAP): Assessment of Basel III riskbased capital regulations - Russia," 15 March 2016c. [Online]. Available: http://www.bis.org/bcbs/publ/d357.pdf. [7] BCBS, "Regulatory Consistency Assessment Programme (RCAP): Assessment of Basel III regulations – European Union," December 2014c. [Online]. Available: https://www.bis.org/bcbs/publ/d300.pdf. [8] BCBS, "Regulatory Consistency Assessment Programme (RCAP): Assessment of Basel III regulations – China," September 2013d. [Online]. Available: https://www.bis.org/bcbs/implementation/l2_cn.pdf. [9] BCBS, "Regulatory Consistency Assessment Programme (RCAP): Assessment of Basel III regulations – United States of America," December 2014d. [Online]. Available: https://www.bis.org/bcbs/publ/d301.pdf.