Description

J. of сorporate finance research ОБЗОРЫ ZETA methodology and variation in the systemic risk of default: accounting for the effects of Type II (false negative) errors variation on lending Stefano Olgiati,

Information

Category:
## Business & Economics

Publish on:

Views: 11 | Pages: 11

Extension: PDF | Download: 0

Share

Transcript

J. of сorporate finance research ОБЗОРЫ ZETA methodology and variation in the systemic risk of default: accounting for the effects of Type II (false negative) errors variation on lending Stefano Olgiati, Alessandro Danovi, Department of Management, Economics and Quantitative Methods, University of Bergamo, Italy. Abstract The loan manager dealing with one single borrower at a time and being responsible for that single decision to lend is exposed to the idiosyncratic risk of default of his customer just like the physician is exposed to the risk of a wrong diagnosis with our strep throat. At the same time if we do not expect the strep throat diagnostic test kit to change we would still expect that physician reading that test to become more careful or update his prior beliefs about his diagnoses when a flu epidemic is likely to kick in with a certain estimated probability (likelihood). However, this has not been the case with loan management there is in fact some consensus that before 2007 a reduction in the standards of idiosyncratic risk assessment by lenders prior to risks pooling coupled with a worsening of the systemic risk scenario, is partly to blame for the well known financial crisis, with some of the blame falling also on the incapacity of actuarial mathematical models (test kits) to update worst case scenarios or be calibrated continuously on the basis of variation in the likelihood of default of the underlying risks pool. The authors of this paper argue that, on the other hand, a standard Bayesian transformation of the ZETA bankruptcy prediction methodology introduced by Altman in allows for a continuous a posteriori update of conditional Type I and II errors due to variation in the systemic likelihood of default. The Bayesian transformation can be used both to condition the loan manager s prior decision (generally based on Basel II-compliant Internal Rating Based system or Credit Agency s Rating) and to update such decision on the basis of any posterior hypothesis (based on actuarial frequentist assumptions of conditional hazard rates) regarding the creditworthiness and the probability of default of an underlying pool of securities. At the same time under a Bayesian framework the ZETA diagnostic test can be conditioned on the new evidence introduced by other tests to increase the total sensitivity of the default prediction models (IRB ratings, TTC ratings, logit, probit, neural) to update the commercial bank s lending decisions. A ground-state, static meta-analysis of Altman s et al. ZETA original article (1977) reveals that the odds of the commercial bank detecting a default after the ZETA score has been introduced (posttest) is 13.2 times more effective than the a priori prediction. Under the same assumptions, the odds of the commercial bank detecting a survival after (post-test) the ZETA score has been introduced is 12.2 times more effective than the a priori. Integration of the ZETA model with other default prediction models reaches a credibility interval of CI 95% when the updated likelihood of default is equal to 60%. As expected, the Efficiency Comparison Test ECZETA= is invariant under the Bayesian transformation. Key words: Altman s Z Score, Risk of Default, Bankruptcy prediction, Bayesian Transformation JEL: G32 Introduction In general modeling correlation and dependence (Li, 2000; Embrechts et al., 2002; Embrechts et al., 2008; Embrechts, 2009) of pooled idiosyncratic risks should perform the task of increasing the efficiency of financial markets by transferring some of banks credit risks and allow them to free up capital (Donnelly and Embrechts, 2010). However, there is some consensus (Brunnermeier, 2009; 71 J. of сorporate finance research ОБЗОРЫ 72 Crouhy, 2008; Salmon, 2009) that before 2007 a reduction in the standards of idiosyncratic credit risk assessment by lenders prior to risk transfer (Donnelly 2010) coupled with a worsening of the systemic risk scenario (Stiglitz, 2008) has led to the well known financial crisis, with part of the blame falling on the incapacity of the mathematical models (Derman, 2013; Lohr, 2009; Donnelly, 2010) to update the worst case scenarios on the basis of variation in correlation in the underlying risks pool and the systemic probabilities of default. The authors of this paper argue that, on the other hand, the Bayesian transformation (Ross, 2002; Gelman et al., 2004) of the ZETA scoring methodology introduced by Altman in allows for a continuous a posteriori update of conditional Type I and II errors due to variation in the systemic likelihood of default (prevalence of disease in epidemiology ( Peacock, 2011). The Bayesian transformation can be used both to condition some prior decision generally based on Basel II-compliant Internal Rating Based system or Credit Agency s Rating and to update any posterior hypothesis (based on actuarial frequentist assumptions of conditional hazard rates) regarding the creditworthiness and the probability of default of an underlying pool of securities. At the same time under a Bayesian framework the ZETA diagnostic test can be integrated with other tests (Ross Sheldon, 2002; Gelman et al., 2004) to increase the sensitivity of other default prediction models (IRB ratings, TTC ratings, logit, probit, neural) to condition the commercial bank s lending decisions. Background Bankruptcy prediction has always been a central theme for banks, which must decide whether or not to approve credit requests on the basis of some estimate of the likelihood of default of the borrower at the time of the request and for a reasonable future period. Since the 1930s scholars started to develop models using statistics and economic-financial indicators (e.g. Smith, 1930; FitzPatrick, 1931; FitzPatrick, 1932; Ramser and Foster, 1931; Smith and Winakor, 1935; Wall, 1936). From the 1960s (Tamari, 1966; Beaver, 1966; Altman, 1968, Deakin, 1972, 1977; Joy et al., 1975) to the 1980s discriminant analysis models became the principal approach 1. Among them, probably the most popular is Altman s Z-score model, formulated in 1968 (Altman, 1968) composed of a discriminant function based on five variables weighted by coefficients 2. The model has been revised several times by its author (Altman, 1983; 2002; etc.) who has constantly updated the parameters and adapted the indices for different populations of companies other than American manufacturers quoted on the Stock Market. The Z -Score (Altman, 1983) is a variation for private companies. The Z Score (Altman, Hartzell and Peck, 1995; Altman & Hotchkiss, 2006, p. 314) was introduced for the non-manufacturing sector or companies operating in developing countries (the original study investigated a sample of Mexican companies) and was recently tested for Italy (Altman, Danovi, Falini 2013). Another adaptation was the introduction of the Z-Metrics System (Altman, Rijken et. al., 2010) that refines the original model, including both market equity levels and volatility, as well as macro-economic variables. Research Questions In medical statistics the Bayesian approach to prior hypotheses testing and updating when new evidence has been introduced is uncontroversial, and represents the contemporary epidemiological standard for clinical and pharmacological research (Peackock, 2011). In particular, Bayesian 1. Univariate and multivariate statistical analysis are still in use, (Edmister, 1972; Blum, 1974; Elam, 1975; Libby, 1975; Taffler, 1976, 1982; Altman et. al., 1977, 1993; Wilcox, 1976; Argenti, 1976; Appetiti, 1984; Lawrence and Bear, 1986; Aziz, Emanuel and Lawson, 1988; Baldwin and Glezen, 1992; Flagg, Giroux and Wiggins, 1991; Bijnen and Wijn, 1994; Kern and Rudolph, 2001; Altman, Rijken, et. al., 2010 Kruchynenko, 2012). For a comprehensive overview see Bellovary, Giacomino e Akers (2007) that summarizes 165 works from the 1930s to The first application of the model involved a group of 66 American manufacturing companies (33 healthy and 33 bankrupt), listed on the Stock Exchange and showed that companies with a Z Score of less than 1.81 were highly risky and likely to go bankrupt; companies with a score more than 2.99 were healthy and scores between 1.81 and 2.99 were in a grey area with uncertain results). The model was extremely accurate since the percentage of correct predictions was about 95% and it received many positive reactions and only a few criticisms. J. of сorporate finance research ОБЗОРЫ transformations are widely used when there is a rapid variation in the prevalence and incidence in the underlying disease pool during an epidemic or a pandemic (Ashby, 2006; Bland, 1998) and the physician needs to update the probability of Type I and II diagnostic errors. The hypothesis of the authors of transferring the epidemiological experience to bankruptcy classification rests on the two following assumptions: 1. The credit manager will utilize the ZETA methodology to condition some a priori expectation (probability distribution) based on some Basel II-compliant Internal Rating Based system or Credit Agency s Rating regarding the probability of default of the corporation he is evaluating. However, the conditional probability that the credit manager will grant or withhold credit on the basis of the further evidence provided by the ZETA analysis is not the same as the underlying true probability of bankruptcy of the corporation (Ross Sheldon, 2002; Gelman et al., 2004), and neither it is the same as the probability of the corporation getting a negative ZETA if they are effectively going to go bankrupt (Ross Sheldon, 2002; Gelman et al., 2004). 2. Variation in the systemic risk of default P(D) of the credit pool should affect the credit manager s belief (Ross Sheldon, 2002; Gelman et al., 2004) in the likelihood of default of the corporation he is evaluating, even if the ZETA methodology discriminant coefficients are based on idiosyncratic risks and are not updated continuously. We performed a ground-state, static meta-analysis of Altman s ZETA original article (1977) to which we refer in full and transformed the variables under a standard Bayesian transformation. In this sense, these are the research questions: 1. We would expect the commercial bank s conditional estimates of default P(D) given that Z is negative (loan rejected) P(D Z 2.675) to be different from the discriminant s analysis probabilities of being negative given actual default P(Z D); 2. The pre-test effectiveness (odds ratio) P(D) of detecting a default increases by the post-test P(D c ) ratio Z D) P( Z D ). c 3. Reveal a commercial bank s lower tolerance for conditional Type I Errors and a higher tolerance for conditional Type II Errors. i.e. P(D Z c ) P(D c Z). 4. Verify that the Bayesian framework permits expansion and integration of the contribution P(D j Z) that the ZETA Analysis gives to any prior hypotheses D j (TTC ratings, neural networks, logit, probit) about the probability of default of the borrower. 5. Verify that ZETA Analysis can be continuously updated on the basis of data relative to the systemic hazard rate of the underlying risks pool P(D). 6. Confirm the postulate that the ZETA test causes, in epidemiological jargon and on a macroeconomic level, a selection of the fittest corporations (patients) amongst the fittest (cream skimming) (Levaggi, 2003) by the commercial bank. 7. confirm that the pre-test and post-test efficiency comparison EC ZETA = is invariant under the Bayesian transformation. All datasets, variables and parameters utilized in the Meta-Analysis are from Altman EI, Haldeman R.G., Narayanan P. (1977), ZETA Analysis, Journal of Banking and Finance; 1: The version utilized can be accessed and freely downloaded from E. Altman s website at the New York University Stern School of Business at (last accessed April 2013). The Bayesian notation utilized is coherent with Berkley University s textbook: Ross Sheldon (2002), Probability, 6th Ed., Prentice Hall (NJ), USA; 3: The epidemiological approach to Bayesian statistics is coherent with Oxford University s textbook: Peacock, JL, Peacock, P.J. (2011), Medical Statistics, Oxford University Press, UK; 7: and 14: Methods We partitioned Altman et al. ZETA Analysis (1977) sample space S=P(D)+P(Z)+P(Dc Zc)=1 on the basis of mixed probabilities P(Z D) = 0.018, P(D Zc) = 0.002, P(Dc Z) = and P(Dc Zc) = and we conditioned the a priori probabilities of default P(D) = 0.02 on the a posteriori evidence of the ZETA tests calculated by Altman (1977). We utilized Bayes Formula to condition the probability of default P(D ) and of survival P(Dc ) on P( Z) given the a priori probability of default P(D) = 0.02, of survival [P(Dc) = 1 - P(D)] = 0.98, the a posteriori experimental accuracy of the ZETA test P(Z D) = and P(Zc Dc) = -.930, and the experimental Type II P(Z Dc) = and Type I P(Zc D) = Errors. P(Z D)P(D) P(D Z) = P(Z D)P(D)+ P(Z D c ) 1 P(D) and P(Z c D c ) 1 P(D) P(D c Z c )= P(Z c D c ) 1 P(D) + P(Z c D)P(D). We then applied the results to the test utilized by Altman in his original paper (Altman et al. 1977, page 43) to compare the efficiency of the ZETA methodology with alternative strategies, which will be invariant after the Bayesian transformation if: q 1 ( M 12 N 1 )C I + q 2 ( M 21 N 2 )C II P( Z c )P ( D Z c )C I + P( Z)P ( D c Z)C II and therefore: EC ZETA EC ZETA BAYES = where CI (Type I Error) is the cost of an accepted loan that defaults, CII is the cost of a rejected loan that would have resulted in a successful payoff; M12,M21 are the observed Type I and II errors (misses) respectively; and N1,N2 are the number of observations in the bankrupt (N1) and nonbankrupt (N2) groups (Altman et al., 1977). Findings All findings are summarized in Exhibit A. The conditional probability of the commercial bank detecting a default when the corporation will actually default after the ZETA test has been introduced is P(D Z) = 21.2% (circa 2 out of 10), with a Type II Error P(D c Z) = 78.8% (circa 8 out of 10). The conditional probability of the commercial bank detecting a healthy corporation when the corporation will actually not go bankrupt (survive) after the ZETA test has been introduced is P(D c Z c ) = 99.8% (circa 10 out of 10), with a Type I Error P(D Z c ) = 0.2% (circa 0 out of 10). The odds of the commercial bank detecting a default after (post-test) the ZETA score has been introduced P ( D Z ) P( D c Z) =.269 is P ( Z D) P( Z D ) =13.2 times more effective than the a priori P(D) c P(D c ) =.02 The odds of the commercial bank detecting a default after (post-test) the ZETA score has been D Z) introduced P( D c = is P ( Z D ) P(D) =12.2 times more effective than the a priori P( Z D ) c P(D c ) = 49.0 The Efficiency Comparison Test EC ZETA = is invariant after the Bayesian transformation as was expected, in fact: P( Z c D)C I + P( ZD c )C II P( DZ c )C I + P( D c Z)C II = Exhibit A: Bayesian reduced sample space partitioning of Altman s ZETA Analysis (1977) P(D) Prior hypothesis - Probability of default P(D) % P(D c ) Prior hypothesis - Probability of survival [1 - P(D)] % P(Z) Probability of negative Z % 74 P(Z c ) Probability of positive Z % P(Z D) Probability of negative Z given default % P(Z c D c ) Probability of positive Z given survival % P(Z c D) Type I - Probability of positive Z given default % P(Z D c ) Type II - Probability of negative Z given survival % корпоративные финансы P(D Z) Probability of default given that Z is negative % P(D c Z c ) Probability of survival given that Z is positive % P(D Z c ) Type I - Probability of default given Z is positive % P(D c Z) Type II - Probability of survival given Z is negative % P(ZD) negative Z and default % P(DZ) negative Z and default % P(ZD c ) negative Z and survival % P(D c Z) negative Z and survival % P(Z c D c ) positive Z and survival % P(D c Z c ) positive Z and survival % P(DZ c ) positive Z and default % P(Z c D) positive Z and default % S Space of events (Summation) % P(D) / P(D c ) Odds ratio of default PZ D) / P(Z D c ) Likelihood 13.2 P(D Z) / P(D c Z) New odds ratio of default after the Z-test has been introduced P(D c ) / P(D) Odds ratio of survival P(Z c D c ) / P(Z c D) Likelihood 12.2 P(D c Z c ) / P(D Z c ) New odds ratio of survival after the Z-test has been introduced Source: Authors elaborations on Altman s ZETA Analysis (1977) Medicine students and practitioners are usually very surprised by the fact that any reliable diagnostic test, with an accuracy of 92.4%, Type I errors of 7.6% and Type II Errors of 7.0%, is capable of diagnosing a patient affected by a disease when the patient actually has that disease only 2 out of 10 times (Ross, 2002). They are normally subject to the instinctive but sometimes severely misleading - a priori belief in the perfect correlation between the outcome of the diagnosis (Z) and the disease itself (D), and in the a posteriori belief that the probability of diagnosing the disease D given a positive test Z is a linear function of utilizing a test Z which is positive given that the disease is present (Peacock, 2011). The reason is that Bayes rule states that the probability of the event default D is a weighted average of the conditional probability of P(D ) given that Z ZETA has occurred and the conditional probability of D given that Z has not occurred (Zc ZETA 2.675), each conditional probability is given as much weight as the event P(D) or P(Dc) on which it is conditioned to occur (Bayes, 1763). In addition, the Bayesian transformation of Type II errors from P(Z Dc) = 7% to P(Dc Z)=78.8% is so steep at low P(D)s (2% in the 1977 ZETA Model) that, at such low P(D)s, increases in Type II Errors of the ZETA test P(Z Dc) ( 7%) (Exhibit B) implies a growing number of creditworthy corporations ( 8 out of 10) whose requests for loans are unexplainably rejected. ZETA Analysis, after the Bayesian transformation consciously or unconsciously performed by the commercial bank s credit manager selects, from a macroeconomic point of view and according to epidemiological jargon, the fittest amongst the fittest (cream skimming). In this sense it appears to be also a very effective managerial (clinical) tool to diagnose the conditional health of a corporation, with a P(Dc Zc) = 99.8% (almost 10 out of 10). 75 Bayesian model averaging (BMA) in forecasting has been analyzed and described in detail by Stock and Watson (2004). When idiosyncratic credit default prediction models are combined, the forecast is a weighted average of the individual model outcomes, where the weights depend on the accuracy of the individual forecasts. In BMA the weights are computed as posterior probabilities that the model is correct. In addition, the individual forecasts in BMA are model-based and the posterior means of the variable are forecast, conditioned on the selected model. Thus BMA extends forecast combining to a fully Bayesian setting, where the forecasts themselves are Bayes forecasts, given the model (and some parametric priors) (Stock and Watson, 2004). In other words, if the commercial bank utilizes several scoring and rating systems and is e.g. P(Dj) = 60% confident that the borrower will default, the ZETA Analysis will condition the a priori probabilities of default P(Dj ) of several other tests on Z so that: P( D j Z)= P ( Z D j)p ( D j ) 95% n P( Z D i )P ( D i ) i=1 Exhibit B: Relationship between decisor s conditional P(D Z) and ZETA Analysis P(Z D) at P(D) = 0.02 and Type II Error P(Z D c ) = 0.07 Robustness E. Altman s ZETA (1977) bankruptcy classification model is based on the sound and empirically tested scientific principles of discriminant analysis which comply with the ROBUST (Reporting Of Bayes Used in clinical STudies) criteria of conditional Bayesian analysis (Sung 2005; Spiegelhalter et al., 1999a; Spiegelhalter et al., 1999b; Spiegelhalter et al., 2000). In particular: the prior distribution is specified normal dis

Related Search

Similar documents

We Need Your Support

Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks