On the correlation between fraud and default risk

The Bundeskriminalamt is the information office of the Police Offices of the “Länder” (the German states). It is also directly responsible for combating organized crime and terrorism. The official account fraud statistics for 2006 can be found in Bundeskriminalamt (2007).

Victim surveys seek to obtain information about identity theft from its victims—individuals who have the most limited view of the problem. For example, victims often do not know how their personal data were stolen or who stole the information. More importantly, victim surveys do not capture synthetic identity theft, a critical piece of the crime.

In an account takeover, an impostor uses one of the victim’s existing financial accounts. While credit card fraud is the most common example, account takeover is a much broader category. In particular, it includes “phishing”, the practice of tricking a victim into revealing passwords or other personal data that allow the thief to access or alter the victim’s existing accounts. In addition to credit cards, phishers target traditional checking and savings accounts, as well as payment systems and auction services such as PayPal and eBay.

Due to the sensitive nature of the data, I am required to keep the name of the bank confidential. Similarly, I am not allowed to report summary statistics for the bank due to the risk of making the bank identifiable.

Severe identity theft losses may undermine confidence in the security and fiscal soundness of a bank. Such losses are likely to trigger unwanted examinations and costly compliance duties by regulators. In addition, incentives to code default losses as fraud losses may stem from credit card asset-backed security deals where default losses are typically borne by the trust or the special purpose entity, while fraud losses are borne by the originating bank. Hence, banks could use the (mis)classification of losses as either default or fraud losses to manipulate the performance of credit card securitization trusts, thereby providing implicit recourse in securitization markets. Vermilyea et al. (2008) find evidence that among banks that securitize credit card receivables, those with poorly performing trusts report, on average, higher fraud losses than banks with healthy trusts.

The sub-sample of fraudsters includes 51 cases without a monetary loss to the bank. These are cases in which the attempted fraud was detected timely, thus any potential loss was successfully prevented.

Potential lenders normally check applicants’ credit history with the major German credit bureau, the SCHUFA Holding AG, providing external risk scores based on the behaviour of the account holder across all sources of credit. However, a credit history may not be available in the case of synthetic ID fraud, or when the applicant is young or recently immigrated to Germany.

White-collar worker refers to a salaried professional or a person whose job is clerical in nature, as opposed to a blue-collar worker whose job is more in line with manual labor. German law puts public employees into two classes, namely ordinary employees (“Angestellte”) and public servants (“Beamte”). The position of public servants is distinguished by the supposed advantages that it confers, such as the salary, a special health plan, an index-linked pension of (at most) 71.5% of the last salary, and most importantly, the virtual impossibility of losing one’s job (basically, the state may only terminate employment in cases of very serious misdemeanours).

A foreigner is a person without German nationality but with permanent residence in Germany. There are many reasons why foreigners are overrepresented in the group of suspects. First, foreigners who reside in Germany are, to a higher percentage, young men. Whereas the population share of foreigners was 8.8% in 2006, they accounted for 13.4% of men aged between 15 and 29 years. Second, some foreigners may be in Germany after fleeing in their homeland because they were offenders there. Third, most foreigners come to Germany because they had no economic success in their home country. The latter reason may be due to factors that foster crime, for example, a lack of education.

I thank an anonymous referee for suggesting this explanation. Furthermore, it is clear that fraudsters should have a low incentive to give away personal information that could lead to an easier identification in case the fraud is detected.

To stay informed about fraudster’s practices and behavioural patterns, my cooperating bank routinely collects police reports and newspaper articles. I investigated 16 fraud-related police reports from different regional police offices and around 25 newspaper articles published over the period 2004–2008. In sum, these sources do not reveal any evidence that fraudsters systematically manage their demographics.

However, any unequal treatment is permissible pursuant to § 3 (2) of the AGG if there exists a legitimate goal that factually justifies the treatment. Given the importance of establishing security systems to detect all kinds of fraudulent acts against a bank, as required by § 25a (1) No. 6 of KWG, identity fraud prevention appears to be such a legitimate goal, but whether this is indeed the case is still a matter of debate in German jurisdiction.

Note that such a selection effect would produce an aligned correlation of variables with fraud and default risk, contrary to what I actually observe. For example, if a significant fraction of foreign fraudsters in fact pretends they are German, fraud rates (in addition to default rates) of Germans should be higher compared to the ones of foreigners.

The population distribution represents the cross-tabulation of 10950 new account fraud suspects, reported in the annual crime statistics for the period 2005–2006. The term “population” is subject to two caveats. First, only new account identity fraudsters caught by the police are considered. Because the clear-up rate for new account fraud was around 75% in the period 2005–2006, for 25% of the population no information about the NATIONALITY/GENDER distribution is available. Second, it is important to mention that all empirical investigations of crime that are based on official statistics have at least one serious shortcoming: The statistics do not display the real intensity of crime but only the size of crime known to the police. How large the share of unreported crimes is depends heavily on the type of crime, with less serious crimes facing a lower probability of being reported to the police. Anecdotal evidence and my conversations with bankers indicate that new account frauds are not generally reported by all banks. Especially smaller banks appear to spare the additional efforts caused by recording and reporting frauds. It is a priori unclear whether and in which direction these two shortcomings bias the reported population distribution.

Similar analyses cannot be performed for the remaining variables (MARITAL STATUS, AGE, etc.) because corresponding information is either missing in the annual crime reports or, as in the case of AGE, a different categorization is used.

I use the default equation from Table 3 to estimate default probabilities for the whole sample, and the LGD regression from Table 6 to predict expected LGDs for each sample observation. The sample correlation between both quantities is ρ = 0.51. In the corporate credit risk literature, Altman et al. (2005) document the link between aggregate default and recovery rates (i.e., 1 minus the aggregate LGD) and show that high default rates tend historically to be associated with low recovery rates, implying that both variables are driven by the same common factor(s). Employing comprehensive data on defaulted bonds and loans in the U.S. over the period 1982–1999, Acharya et al. (2007) show that while balance-sheet based determinants of default risk and weighted average default rates are (negatively) correlated with creditor recoveries, this correlation is not perfect. Contract-specific characteristics (such as seniority or collateral) and, in particular, industry distress seem to affect creditor recoveries over and above factors that affect default risk.

Postal-code regional dummies (not shown) are included in both loss regressions, but they are only jointly significant in the LGD case.

See Efron and Tibshirani (1993) for an excellent exposition of bootstrap methods.

The percentage share of sample observations from the 10 almost equally populated one-digit postal areas is clustered at around 10%, with two outliers, postal area codes 7 (15.07%) and 9 (4.39%). Excluding these two regions, however, does not change any of the results in this paper.