Credit Scoring Models: Missing Information and the Use of Data from a Credit Register

We study the problem arising from the lack of information on some debtors’ behavior in the databases used to develop credit scoring models, and the use of the behavioral information stored at a Credit Register as a potential solution to the problem. To this purpose, we use yearly information provided by the “Central de Deudores del Sistema Financiero” (Credit Register) of the Argentine Central Bank. A limitation of this Register is the removal of a significant number of debtors, on a regular and widespread basis, without recording the reasons for such removal, which may be due to two opposite situations: debt cancellation or default (and the decision by the bank not to continue with the collection procedures). The goal of this paper is not to model the process of missing data, but to focus on: (i) estimating the risk of the removed debtors and (ii) taking advantage of the additional behavioral information of other creditor institutions –recorded in the Credit Register– so as to estimate the risk mentioned in (i) and improve the prediction of the score. In this way, we can also assess the impact of not considering this missing information when the portfolios’ risk of credit institutions is assessed. The main strategy consists in using the behavioral data of other entities and comparing the results provided by three methods: 1) ignoring the records with missing data (listwise deletion), 2) using direct imputation (i.e., in the case of missing information, imputing to behavior the value applicable to the worst behavior of that particular debtor in the system) and 3) using a fractionally weighted imputation method. The paper proves that the commonly-used procedure of removing from the sample the debtors that are no longer in the database, when the reasons behind their disappearance are unknown and cannot be modeled, is not innocuous. The bias that may be introduced is difficult to correct and, even when a correction is attempted, its accuracy cannot be known for certain. Therefore, we underline the importance of making sure that the design of credit risk databases rules out any “holes” that might hinder the follow-up of individuals. In addition, the comparison among the different methods being explored seems to indicate conclusively that the risk is overestimated when direct imputation of behavior is applied, using the worst status observed in other institutions. The model that appears to be more precise is the so-called “fractionally weighted multiple imputation” model, whereby the behavioral information of the Credit Register is used in an imputation model using a logit regression. This approach is innovative in scoring literature and apparently preferable to direct imputation. However, we cannot be conclusive regarding the convenience of using it in all cases. In particular, the analysis of the Argentine “Central de Deudores” suggests that a calibration adjustment to a simpler model using the listwise deletion method may solve an important part of these deficiencies. But this conclusion will depend on the case and the period under analysis. The specific data used in this study refer to an exceptionally good period in the domestic default rates. In addition, as in previous studies, there is evidence that the outcomes from scoring models developed on the basis of public credit information are very good, despite the limited selection of explanatory variables. The results of this study are of interest to the banking industry, supervisors and researchers, since it is common practice to develop models from a sample where a group of debtors has been removed because their information is incomplete, is of poor quality, or shows other deficiencies.