Traditional credit scores can now be augmented with new information to build better credit models. Social media profiles, GPS location data, and transaction records can help to form a complete picture. The benefits of better credit models might include more accurate pricing of credit, fewer defaults, and broader access to credit.
How much does nontraditional data really help to improve credit models? In theory, additional information should always lead to models with equal or better predictive power. In actual practice, even the most easily obtained information has surprising power. Bertrand and Kamenica found that using an Apple iOS device was one of the best indicators of being in the top income quartile. Using iOS was also a good predictor of default rates. Most strikingly, Berg, Burg, Gombović, and Puri found that the discriminatory power of a model using only online data was 69.6%. That is actually higher than the 68.3% obtained by their model that used only credit bureau scores.
Using new sources of data can help many consumers access credit for the first time. According to the World Bank, 1.2 billion adults started using financial services between 2011 and 2017, but 1.7 billion remained unbanked. The reluctance to extend credit to new customers is understandable. Berg, Burg, Gombović, and Puri found that customers without credit scores had a default rate of 2.49%, compared to 0.94% for those with credit scores. However, the vast majority of customers without scores did not default. New sources of data, like social media profiles, can assist financial institutions in sorting these potential customers. Better models can also help consumers to improve their creditworthiness. Services like Credit Sesame already provide educational information about how to boost credit scores. Similar services incorporating online information could be very effective in helping the unbanked gain access to credit. The key is sharing information about new credit models with consumers.
The use of nontraditional information in granting credit can be even more effective when combined with traditional credit scores. Berg, Burg, Gombović, and Puri also created a model the used both the credit score and online data. The combined model had discriminatory power of 73.6%, higher than either credit scores or online data alone. This improvement over traditional credit scores is significant, and it is costly for most consumers to engage in manipulation. Buying an iPhone to increase a credit score may not be cost-effective. Spending time to build a strong profile on LinkedIn has a high opportunity cost.
It is becoming more important to use all available data to produce the best possible estimates of creditworthiness. Firms that have access to additional information are starting to arrive at different conclusions about credit. A Federal Reserve study by Jagtiani and Lemieux found a decreasing correlation between credit ratings from a fintech company and traditional credit scores. There was a correlation of approximately 80% in 2007, but it declined to below 35% by 2015. Additional data and new machine learning methods are increasingly useful in making credit decisions.
As the information age produces more information on all consumers, financial institutions must expand their horizons. Using nontraditional information is essential for minimizing risk when extending credit to consumers without formal credit histories. However, alternate data sources are becoming more useful for evaluating all potential customers. New data, new methods, and new partnerships are necessary to build better credit models for the future.
