Informal bankruptcy forecasting procedures

from Wikipedia, the free encyclopedia

With informal insolvency prediction failure prediction methods are referred to, in which human credit analysts failure scores based on their intuition and personal experience to create. If necessary, checklists or (more or less detailed and precise) guidelines or procedural regulations are available to them.

Empirical findings on the estimation quality

First of all, it is undisputed that borrowers can be valued cheaper, faster and more transparently using standardized methods than in the context of individual analyzes by human evaluators. For this reason, the valuations of small loans are largely standardized or even automated at most banks .

Beyond that, however, the prevailing opinion among banks and rating agencies is that if they are given sufficient time and resources, experienced analysts could beat any statistical model. Ultimately, their analyzes were not limited to evaluating a few key financial figures, but included all aspects of the company and its environment that were considered relevant. Last but not least, this is also demanded by the regulatory authorities. In the credit risk analysis of natural persons , the exclusive use of statistical models is even prohibited by law (even if this prohibition is often not observed).

A meta-analysis of over 100 studies from various scientific fields shows that individuals with little specialist knowledge can make considerably more precise forecasts than uninformed individuals, but that additional specialist knowledge does not lead to any increase in forecasting accuracy and that the quality of even knowledgeable human decision-makers is simply statistical Forecast method is inferior.

Comparable studies specifically on the subject of insolvency prognosis also show that human credit analysts, to whom the annual financial statements of the companies to be assessed were made available, can use the information contained therein to make insolvency prognoses that correlate with the subsequent insolvency events . However, the quality of the analyst forecasts is not only well below the quality of simple statistical methods (see for example discriminant analysis , logistic regression or decision tree methods ), but even below the univariate selectivity of individual financial indicators . Two basic theories describing human behavior can be used to explain these disappointing findings :

Both points are explained in the following sections. However, it can be assumed that at least in the studies cited, incentive problems did not play an essential role, but that the impairment of the predictive ability of human decision-makers caused by irrationalities is already sufficiently large to explain the inferiority to simple statistical methods.

Irrational information processing

Since there is no comprehensive, proven theory for explaining and forecasting corporate insolvencies and since human decision-makers do not have the storage capacity and processing capacity of modern computers to make at least empirically sound forecasts, credit analysts are always based on their own intuitions and experiences when making bankruptcy forecasts and reminders. However, it is known that the predictions made by human decision-makers under these circumstances are subject to systematic errors.

In the context of credit risk analysis of particular relevance here are the over-interpretation of random events (law of small numbers), ignoring general statistics in the presence of particularly memorable individual memories ( availability heuristic , availability heuristic) and persistence once preconceptions ( was perseverance , anchoring ) which is caused, among other things, by the selective perception of new information ( confirmatory bias ). Information that corresponds to the preconceived opinion of the analyst is perceived as particularly credible; contradicting information tends to be ignored or misinterpreted. Furthermore, the perceived probability of occurrence of events increases as soon as these events have actually occurred and become known to the individuals. Individuals therefore also believe that they have good these events with their knowledge can predict ( hindsight bias , hindsight bias ). Experts in particular overestimate the accuracy of their forecasts in an environment with little predictability. Theoretically, through a consistent comparison of past forecasts with insolvency events that have actually occurred, individual forecast errors can be identified and, if the individual forecasts have been adequately justified and documented, the causes of the errors made can be identified, so that learning processes can be triggered to improve the individual forecasting ability. Unfortunately, in the past, for example due to mergers or changes in their rating methodology, banks have mostly failed to consistently record and maintain the required data.

Until recently, a systematic comparison between individual prognoses and actual insolvency events was therefore not possible. It is only through the regulations of Basel II that banks are obliged to set up appropriate databases and regularly validate their default forecasts . Despite these regulatory requirements, the conditions for successful individual learning by human analysts will continue to be very poor for the following reasons:

  • Only in the case of extreme failure forecasts , i.e. H. In the case of a forecast failure probability of 0% or 100%, it is in principle possible, based on individual observations, to assess forecasts unequivocally as “right” or “wrong” or “good” or “bad”. As a rule, the quality of the forecasts can only be determined with a statistical uncertainty. Since bankruptcies are rare events, the predictive ability of a human analyst can only be determined with a very high degree of uncertainty on the basis of several hundred (!) Forecasts and associated failure realizations (bankruptcies vs. non-bankruptcies).
  • Depending on the intended forecast horizon, forecast errors can only be recognized with a relatively long time delay. Banks usually base their ratings on a validity horizon of one year, rating agencies on a longer, unspecified horizon.

Reliable and timely feedback on the quality of individual forecasts from individual analysts is therefore practically impossible. A statistically meaningful validation is ultimately only possible for automatable processes that can generate a sufficiently large number of forecasts. This argument also calls into question the right to exist of all those rating agencies that on the one hand only have a small customer base (track record) and on the other hand their rating judgments are essentially based on non-automatable evaluation procedures, in particular subjective analyst judgments . The numerous rating agencies established in Germany since 1998 have only issued around 300 ratings, of which only 30 have been published.

Institutional disincentives

In addition to the reasons explained in the previous section for the poor predictive ability of human decision-makers, which can be traced back to irrational learning behavior , rational behavior , in this case in the sense of opportunistic-self-interested behavior, can also be the cause of a poor forecast quality of human insolvency forecasts. On the one hand, opportunistic self-interested behavior can be expressed in the fact that the analyst, whose activities can only be incompletely monitored by the client, for example a bank, does not carry out the tasks assigned to him with the necessary care in order to save time - or that he will steer the rating result in a direction that is convenient for him through targeted manipulation . In the former case, the rating tends to be influenced unsystematically or with a “tendency to average” in order not to provoke unpleasant inquiries. In the second case, it is systematically, tending to be positive, distorted. Incentive problems of the last kind exist above all when ratings are created or can be influenced by relationship managers and not by special rating staff.

The worse the account manager rates the company assigned to him, the lower the profitability that the bank attaches to the commitment and the lower the loan volume that the bank is willing to grant the customer is typically. If the account manager's compensation is linked to the loan volume granted and / or to the profitability of the commitment expected by the bank, the account manager has strong incentives to distort the rating positively.

Critical appraisal

The previous studies show that the listed points of criticism have consequences for the desired role of human analysts in the credit assessment process. People who constantly evaluate each individual case according to new rules known only to them cannot make better insolvency prognoses than those who do so according to disclosed, relatively rigid and inflexible, but empirically checked and calibrated rules - typically with the help of a computer and under Use of statistical bankruptcy forecasting techniques. Once a bank has developed an empirically validated insolvency forecasting procedure, the further role of people in the credit assessment process does not have to be limited to the collection of data that can be determined largely without discretion. Numerous studies show that the inclusion of “soft factors” such as “market position” or “the quality of management ”, which must be subjectively assessed by human employees, improves the forecast quality compared to insolvency forecasts based solely on hard factors . In contrast to human insolvency prognoses, human value judgments of concrete facts seem to contain useful, additional information.

Empirical studies consistently show that human decision-makers when assessing qualitative characteristics of companies not only give significantly better credit ratings on average than statistical methods when assessing annual financial statements, but that these assessments are also significantly less differentiated. that is, their spread is smaller. The desired distribution properties of the data can be produced using suitable transformation and aggregation processes. The problem, however, is the low reliability of human value judgments.

literature

Individual evidence

  1. ↑ This article is based on Bemmann (2007, section 2.2).
  2. ↑ see Falkenstein, Boral, Carty (2000, p. 16)
  3. ↑ see Treacy, Carey (2000, p. 898), Basler Committee (2000, p. 18)
  4. ↑ Treacy, Carey (2000, p. 898): “Many banks use statistical models as an element of the rating process, but banks generally believe that the limitations of statistical models are such that properly managed judgmental rating systems deliver more accurate estimates of risk . ”
  5. ↑ Falkenstein, Boral, Carty (2000, p. 16): “The value of quantitative models over judgment is not purely a scale economy argument. […] Yet, many presume that given enough time most sufficiently intelligent and experienced analysts would outperform any model. "
  6. ↑ See Deutsche Bundesbank (2004, p. 6) for a description of the structure of one's own credit assessment procedure: “Since this 'standardized' procedure cannot always do justice to the specific circumstances of individual companies and more recent developments, there is scope for the assessor to deviate from the classification proposals. "
  7. ↑ see S&P (2003, p. 17): “There are no formulas for combining scores to arrive at a rating conclusion. Bear in mind that ratings represent an art as much as a science. A rating is, in the end, an opinion. "
  8. ↑ BaFin (2002, Item 70): “Relevant indicators for determining the counterparty default risk in the risk classification procedure must be qualitative as well as quantitative criteria, as far as possible. [...] ". According to the regulations of Basel II , the use of automatic models is permitted within the framework of the IRB approach , but “Sufficient human judgments and human monitoring are required to ensure that all essential information , including that which is outside the scope of the model, is taken into account Model is used in an appropriate manner. ”, See Basel Committee (2004, item 417).
  9. ↑ see ULD (2006, p. 18f. And p. 86f.)
  10. ↑ see Armstrong (1985, pp. 92f.)
  11. ↑ See Armstrong (1985, pp. 55f., P. 94 and the literature cited there). The studies examined the skills of experts and a. in the prognosis of length of stay in hospital, soccer results or study successes.
  12. ↑ Keasey, Watson (1991, p. 99) “For many well-specified and repetitive decisions, the classification accuracy of even relatively simple quantitative models have been shown to consistently outperform human decision-makers [...] This literature has also shown [...] that a statistical model is usually able to significantly outperform specialists. "
  13. ↑ see Falkenstein, Boral, Carty (2000, p. 16f. And the literature cited there)
  14. ↑ For example, the creditworthiness assessment by loan officers can be systematically influenced by easily identifiable and corrective circumstances such as sale-and-lease-back ( buy -back) or activation of R&D costs, see Blake et al (2000, pp. 139f. And the literature cited there).
  15. ↑ see Rabin (1998, pp. 24ff. And the literature cited there), Armstrong (1985, pp. 86f., 96f., 110f., Pp. 143f., P. 436f. And the literature cited there), Rosenkranz, Missler-Behr (2005, p. 108ff. And the literature cited there).
  16. ↑ see Rabin (1998, p. 24ff. With further references)
  17. ↑ see Rabin (1998, p. 32f.)
  18. ↑ See Fischer (2004, p. 381ff.) On the resulting low dynamics of human rating judgments.
  19. ↑ see Rabin (1998, p. 32f.)
  20. ↑ Basel Committee (2000, p. 25): “[M] ost of the surveyed banks did not have sufficient internal data for specifying loss characteristics for all borrowers based on their own default history […]. [D] ue to data constraints, the majority of banks surveyed relied either partially or fully on the data provided by the major rating agencies, public databanks such as national credit registries, or data from consulting companies. "
  21. ↑ Carey, Hrycay (2001, p 199) "Remarkably, very few financial institutions have maintained usable records of default and loss experience by internal grade for Their Own portfolio. Thus, the obvious actuarial approach, computing long-run average default rates from the historical experience of borrowers in each internal grade, is not feasible in most cases. […] Evidence presented below indicates that a relatively long time series of data is needed for good actuarial estimates. [...] Even in cases where banks have gathered such data, changes in the architecture or criteria of their internal rating system (which occur frequently) greatly reduce the utility of prechange data. "
  22. ↑ Araten et al (2004, p. 93): “Many firms have evolved their rating scale and methodology over time, often as a result of mergers in which they have had to reconcile different rating systems employed by predecessor banks. Under these circumstances, it is often a challenge to develop a database of ratings history that fairly represents a consistent ratings philosophy. ” and analogously Treacy, Carey (2000, p. 912).
  23. ↑ see Basel Committee (2004, items 264ff. And 429ff.)
  24. ↑ see Basel Committee (2004, Item 500ff.)
  25. ↑ see Tversky, Kahneman (1986, p. 274) “The necessary feedback is often lacking for the decisions made by managers, entrepreneurs, and politicians because (i) outcomes are commonly delayed and not easily attributable to a particular action; (ii) variability in the environment degrades the reliability of the feedback, especially where outcomes of low probability are involved; (iii) there is often no information about what the outcome would have been if another decision had been taken; and (iv) most important decisions are unique and therefore provide little opportunity for learning. "
  26. ↑ For numerical examples see Engelmann, Hayden, Tasche (2003, p. 19).
  27. ^ See Basel Committee (2001, p. 12). See S&P (2003, p. 41): “ Standard & Poor's credit ratings are meant to be forward-looking; that is, their time horizon extends as far as is analytically foreseeable. " and Cantor, Mann (2003, p. 6f.): " Moody's primary objective is for its ratings to provide an accurate relative (i.e., ordinal) ranking of credit risk at each point in time, without reference to an explicit time horizon."
  28. ↑ see Wieben (2004, p. 10, p. 14f. And p. 85)
  29. ↑ See Fischer (2004, pp. 208ff.) Who creates a typology of corporate customer advisors based on his evaluations . The typifications range from the "system player", who generally answers questions with great care, but occasionally undertakes targeted interventions in order to achieve an overall grade from his point of view to the " ignoramus " who is fundamentally negative about the use of the computer-aided questionnaire and not the necessary one Spends time answering the questions carefully.
  30. ↑ see Totzek (1999, p. 321f.)
  31. ↑ see Treacy, Carey (2000, p. 904). In around 40% of the banks, customer advisors have the main responsibility for assigning ratings; in 20% of banks, ratings are always assigned in cooperation with customer advisors and special rating staff; in 30%, small commitments are handled by customer advisors and larger commitments by special rating staff, and in 15 % of the banks, the main responsibility for assigning ratings always lies with special rating staff (the sum of the names results in 105% due to rounding), see ibid.
  32. ↑ see Salomo, Kögel (2000, p. 236) and Hartmann-Wendels (2006, p. 209)
  33. ↑ In the case of credit evaluation by special rating staff, there is no reason why the bank should link the analyst's remuneration to the future credit volume or the expected return. In the case of credit rating by customer advisors, such a coupling would tend to make sense in order to provide incentives for acquisition efforts if their actual work effort can only be observed to a limited extent.
  34. ↑ Treacy, Carey (2000, p. 898): “At banks that use ratings in computing profitability measures, establishing pricing guidelines, or setting loan size limits, the staff may be tempted to assign ratings that are more favorable than warranted.” and ibid., p. 919: “Some institutions found that many loans were upgraded shortly after the introduction of profitability analysis, although the overall degree of the shift was small. One institution specifically mentioned an upward bias of about one-half grade relative to previous rating practice. Many noted that the number of disagreements in which relationship managers pressed for more favorable ratings increased once such systems were put into place. "
  35. ↑ see Lehmann (2003, p. 2), Fischer (2004, p. 229), Grunert, Norden, Weber (2005, p. 513f.)
  36. ↑ see Salomo, Kögel (2000, p. 234f.), Lehmann (2003, p. 9), Fischer (2004, p. 381ff.), Grunert, Norden, Weber (2005, p. 512, 515f. And those there cited literature), analogous to Bemmann, Blum, Leibbrand (2003, pp. 20ff.).
  37. ↑ See Blochwitz, Eigermann (2000) on the possibilities of integrating ordinal soft factor assessments into the various commonly used statistical forecasting methods, the technical requirements of which vary in part.
  38. ↑ see Fischer (2004)