Last year, the rate of defaults on loans intermediated by financial technology platforms rose. But more recent numbers show some signs that the cycle of boom and bust in fintech lending may be nearing an end. I highlight how ignoring some basic economics exasperated this cycle and what lessons can be learned from it.
Let’s first understand the broad structure of a typical fintech algorithm. It uses three types of data. The first is ‘alternate data’, types that a brick-and-mortar bank generally does not use—ranging from shopping details to the type of phone or operating system used. The second source is the banking and investment details of the borrower.
Thanks to account aggregators, fintechs now have access to information such as savings account transactions, insurance and mutual fund investments, taxes, etc. These are data points that banks have access to, but may not be using optimally due to either technological or human resource constraints, or inertia. Finally, fintechs also use information about borrowers’ past loan performance from the credit bureau, which all formal lenders use.
All these data points are fed into a machine-learning algorithm that’s trained on past loan performance. In other words, the algorithm understands the relationship between all the input variables and past loan performance. Having understood the relationship, it can predict loan performance, given the input variables of new applications.
These algorithms are tested out-of-sample (sample of loans not used for training) to ensure they are robust. Once the fintech platform is satisfied with the robustness, it is deployed for new applications. Applications above a threshold on the probability of default, as predicted by the algorithm, are approved.
Although the structure appears robust, it ignores some basic economic principles. Let us focus only on the issues relating to the variable on which the algorithms are trained—past loan performance. There are at least three issues here. First, we know well that loan officers have an incentive to evergreen loans, the practice of issuing additional loans to a borrower on the verge of default to avoid recognition of default.
A recent paper, titled ‘Zombie lending: Theoretical, International and Historical Perspectives’, by professors Viral Acharya, Matteo Crosignani, Tim Eisert and Sascha Steffen, published in the Annual Review of Financial Economics, clearly shows that evergreening of loans is prevalent all over the world. My paper, titled ‘Identifying Evergreening: Evidence Using Loan Level Data’, published in the Journal of Banking and Finance, shows that loan officers in India routinely issue new loans to facilitate repayment of existing ones. One learning from this research is that it is tough to detect evergreening loan-by-loan using easily available data. The parties involved are likely to make every effort to hide it.
How does evergreening impact fintech algorithms? The issue here is that the outcome variable that fintechs use, based on credit bureau data, is past loan performance. It does not distinguish between evergreened loans and others. The algorithm treats loan repayment by a good borrower and a zombie borrower who uses evergreening to repay as equals. However, we know from research that evergreened loans eventually tend to default. Thus, by ignoring evergreening, the algorithms tend to underestimate the probability of default.
A reader may ask if evergreened loans eventually default, would the machine learning algorithm trained on loan performance also learn about this practice eventually and separate evergreened and other loans? This is where the second limitation comes in. Most fintech algorithms I am aware of do not cover the full business cycle.
Evergreened loans eventually default at the end of a business cycle, when the regulator intervenes by conducting audits or when the bank’s top management is replaced. Last year’s regulatory intervention in the form of increased capital adequacy requirements is one such action. Here again, the choice of the sample period should be dictated by economics—a complete business cycle—rather than a mechanical number of years. Unfortunately, most algorithms are not trained using the full business cycle data and, therefore, cannot detect evergreening.
The third issue comes from the ‘selective labels’ problem. After being trained, fintech algorithms are used to decide applications of new-to-credit borrowers. Such borrowers may be systematically different than those having bank loans. The problem becomes acute if the loan officers who issued past loans used ‘soft information’—a type that cannot be easily codified and communicated—during screening. Because this soft information is not an input into the machine learning algorithm, its predictions may not apply well to new-to-credit borrowers.
So fintechs will do well to take these basic insights from economics into account while designing algorithms. Using the entire business cycle for training algorithms should be straightforward. Although detecting evergreening is hard, some algorithms are available in the academic literature to help detect such lending. The papers I referred to can provide a good start.
To capture soft information, we are experimenting with training algorithms on what an experienced banker would have done. Finally, fintechs have to realise that research does not merely mean adding more data and obtaining higher explanatory power. There needs to be a more active attempt to understand the economics driving these algorithms.
Prasanna Tantri
Associate Professor, Finance, and Executive Director, Centre for Analytical Finance at ISB
(Views are personal)