Multi-Bureau Analysis for High-Level Tailoring of a Scorecard

Share on linkedin
Share on facebook
Share on twitter

Credit facilities are a strategic avenue for businesses to achieve growth, allowing one to “buy now, pay later” enables consumers to have a higher flexibility, and therefore makes merchants extending such facilities a more desirable outlet for the consumer. In early 2019, a multi-store credit-retail firm in South Africa noted a decline in performance from their existing scorecards. It was realised that the bureau’s variables upon which their scorecard was built had become unstable, which negatively impacted the scorecard’s performance. Thus, Incline was engaged to redevelop the firm’s scorecards.

Scorecards and Credit Bureaus

A scorecard is essentially a predictive modelling tool which quantifies the riskiness of an applicant for credit, based on known attributes of the applicant, and empirical performance of previous applicants with similar attributes. A retailer who lends negligently to every applicant will be subject to adverse selection, and quickly become bankrupt. A scorecard allows lending decisions to be made in an evidence-based manner, and protect the credit provider from excessive risk, whilst also protecting the credit-hungry consumer from taking on excessive debts (our previous article discusses the development of scorecards when providing credit). The key to a predictive model’s success is the data used to train the model, which needs to be representative of the population upon which the model is going to be used. As the population shifts over time, new behaviours can be captured by rebuilding a scorecard as opposed to realigning an existing model.

Incline was approached to reconstruct and tailor scorecards to best suit the retailer’s clientele.  We set about constructing a comparative review over three distinct data sources (credit bureaus). A credit bureau is a centralised authority which collects and distributes credit information shared from credit suppliers. South Africa has 4 main credit bureaus, which have access to the same credit information (L702 SACRRA layout), but there are also notable differences across them. Mainly the data they gather from alternative sources, as well as the meta-variables they construct internally. It is worth noting that gathering data from credit bureaus can be costly in time and data, however, there is a large potential return for this investment. By approaching three of the four large bureaus, we gathered a wealth of data, and directly tailored the credit-scorecard specifically for the client on a bureau selection-level, unlocking a notable 28.8% performance boost.

Comparing the Bureau’s Scorecard Performance

The Lorenz curve is a data visualising tool plotting the cumulative percentage of accounts with “bad” payment behaviour against the cumulative percentage of accounts with “good” payment behaviour. The larger the area below the curve, the better the discrimination ability by the score output. The Gini coefficient is derived as the area under the curve. Bureau A’s Gini coefficient showed a 28.8% improvement compared to Bureaus B and C. This performance was confirmed through holdout-data.

Bureau A's scorecard was the best at distinguishing between "bad" and "good" payment behaviour

The take-away from this exercise demonstrates that scorecard tailoring at a bureau-level is an avenue which may offer performance improvement. In this case, allowing additional resource investment into gathering data from multiple bureaus added more predictive power to the final scorecard model than any single variable.

See more of what Incline can offer on our website incline.co.za.