This article shows how to prevent and fix mistakes in scorecard development, in order to allow most precise borrower evaluation, rating and segmentation.
Loan application processing systems help lenders decrease operational costs and acquire more profitable accounts. These systems use data models called scorecards to provide a robust lending mechanism. Scorecards provide a set of weights assigned to characteristics that demonstrate customer’s credit worthiness. With scorecards, customers are rated according to the probability of timely payment.
Once customers are scored and rated, you can automatically make profitable decisions: accept/reject the loan application, set optimal limits and pricing etc.
Below we describe how to take full advantage of the available data and achieve the most precise borrower evaluation.
Inaccuracies in scorecard development can be made in each of the following stages – data sampling, statistical evaluation of borrowers’ characteristics, formation of training and validation datasets.
Mistakes during data sampling are caused by selecting data that does not meet the requirement of representativeness and randomness.
Representativeness means the maximum proximity of the data indicators in the sample to the actual borrower's characteristics in the loan portfolio. This requirement is natural and understandable, because the scorecard is expected to reflect specifics of the dataset used for its development.
Randomness means that loan application data should be included in the working sample independently.
There are two ways to prevent sampling mistakes – either by directly controlling data sampling procedures, or by evaluating the borrower’s statistical characteristics.
When evaluating borrower’s statistical characteristics you should pay attention to those indicators that are unnaturally distributed. For instance:
Once you have discovered unequally or unnaturally distributed characteristics, you should change or adjust the procedure of forming the working sample and set different rules for assigning values to each characteristic.
The main criteria for selecting a training data sample says: training data should provide enough examples of profitable and delinquent loans.
As a rule, a training dataset comprised of 3500-4000 records is enough to successfully train a scorecard; provided that this dataset offers a 3:1 proportion of “good” and “bad” loan cases. It is possible to train a scorecard with fewer records, but you should keep the 3:1 proportion of the loan case profitability.
This way, data mining techniques can be applied to improve borrower rating and segmentation in automated loan application processing. In upcoming articles, we will show more ways to apply data mining for decision automation.