We have written several blog posts on credit scoring using both transaction data and telco data and in a 2012 blog series, MMU talked to Experian and Cignify, two third parties using telco data to develop credit scorecards. Since then, the industry has made significant progress in launching a range of credit products. Start-ups like Inventure, which develops credit-scoring models via mobile applications, and offerings like M-Shwari by Safaricom / CBA, show how a combination of telco and transaction data can be used to extend credit to customers on better terms.
Telenor Group is present in 14 markets across Asia and Europe, and are offering mobile financial services to customers in three of them. By utilising and monetising telco data, Telenor is introducing new types of mobile credit products, by building predictive credit-scoring models in-house. These new products are not only for mobile money loans, but also emergency airtime top ups and handset financing. In addition, these credit-scorecards can be used to increase credit limits for post-paid customers, or convert a pre-paid to post-paid customer without a financial credit history.
I recently spoke to Catrin Bekker Dahle, a Senior Analyst in Telenor Financial Services, to better understand the Group’s rationale for developing predictive credit scorecards in-house, the process for developing these scorecards, and the different use cases for credit scoring.
MMU: What are predictive credit scorecards?
Catrin Bekker Dahle (CBD): Credit scorecards are sophisticated statistical models that attempt to predict whether a customer is going to default on a loan. The main idea is that all defaulting customers behave in the same way. In Telenor, we develop predictive credit scoring modelling using machine learning and data mining software that can handle a large data set, thus detecting correlation that a human mind could never discover. Rule-based credit scoring models, which uses previous knowledge and expert judgement to assess the relationship between predictors and the likelihood of a customer defaulting on a loan. However, to increase the strength of our predictive credit scorecard, expert judgement is also a helpful tool and this is why we work closely with local experts in the different Telenor business units.
MMU: What is the predictive power of telecom data compared to transaction data?
CBD: In many of our markets, we have yet to launch financial services, and transaction data is not available. Â Customer data is limited to telco data, such as call data records (CDR), and top up and demographic data collected at sign up. However, the richness of telecom data provides insights that are far more accurate on predicting human behaviour than transaction data alone. Transaction data only tracks a customer when they transact, whereas telecom data can reveal a lot more on customer behaviour. For instance, the frequency of top ups can give indicators of income level, while location data could reveal job stability, how long a customer have had their phone, and how often they change, can provide insight to the likelihood of default. However, a combination of telco and transaction data is ideal, because it reveals a more complete picture of customer cash flow.
MMU: How do you develop scorecards from telecom data?
CBD: One of Telenor’s objectives is to build credit history for our unbanked customers. We plan to build credit history for customers gradually, by building customer credit profiles and more accurate credit scoring models with these following steps:
- Start to build a model for a low risk product, or a low risk customer segment. For example, this could be emergency top ups for pre-paid customers, or post-paid customers.
- Identify predictors for the product, which are the parameters most likely to predict customer default.
- Develop and test the statistical model/credit scorecard
- Pilot the model on a small set of customers to verify that the predictors are accurate in identifying customers more likely to default.
- Monitor how the pilot works, iterate the model, run the scoring again, and continuously monitor and adjust the model after launch.
- Finally, move to a high-risk product or segment and replicate the process again. This could be device loans for prepaid customers.
By continuously credit scoring customers and monitoring their behaviour, the credit worthiness of a customer becomes more precise over time. The characteristic of the risk associated with unsecured consumer loans, especially in emerging markets, is that risk changes rapidly.  This requires an analytic framework and a system that can very quickly respond to the dynamics of shifting risks. Monitoring the credit scorecard performance, evaluating the results and being able to quickly deploy new models is critical to find new opportunities and avoid new risks.
MMU: Why is it important for Telenor to develop the predictive scorecards in-house, and not just outsource it?
CBD: First, when we keep customer insights in-house and build experience, we can adjust the scorecards more dynamically than a third party could, which will provide more accurate and precise scorecards. Second, want to ensure that we protect our customer privacy by not sharing their data. The flexibility of building the capability in house allows us to score customers at our convenience, for instance every night or in custom batches. We manage the risk by building internal monitoring processes and our own black lists. Finally, by developing the scorecards in house, we keep the revenues that new credit products provide.
MMU: Finally, what advice do you have for other mobile money operators wanting to develop predictive score cards in-house?
CBD: Many MNOs already have the capability to develop telecom data-based scorecards in-house, but most likely not within the mobile money team. Often, this competence sits in an MNO’s business intelligence or marketing units, those developing and managing advanced churn prediction models. These data miners or analysts are familiar with the data sets used for other telco modelling, which is an excellent starting point for developing predictive credit scoring models.