Optimizing Credit Limit Adjustments Under Adversarial Goals Using Reinforcement Learning

Credit cards are an essential part of modern financial life; according to the Consumer Financial Protection Bureau (2021), 175 million North Americans, more than half of its population, own credit card products. On the other hand, the same cannot be said for developing countries; according to the World Bank, an average of only 55% of Latin Americans had a bank account in January 2020, and only approximately 20% have a credit card (World Economic Forum, 2022). However, companies that use financial technology, known as fintechs, have enabled digital financial services that can help the unbanked population overcome difficulties such as costs, geographical impediments, long waiting times, and lack of financial history in accessing traditional banking products (Khera, Ng, Ogawa, & Sahay, 2022; Rojas-Torres, Kshetri, Hanafi, & Kouki, 2021). The number of fintech companies in Latin America has risen rapidly, and their appearance has altered the behavior of traditional banks, which are now seeking innovation and changes to customercentered approaches (Vives, 2019) and have decided in some cases to create alliances with these new companies (Bejar et al., 2022). Because the financial industry is primarily based on information, financial process reports have been more easily transitioned to the digitization stage; this situation is in contrast with the consumer goods industry, which includes a physical element (Puschmann, 2017). In addition, emerging "super-apps", which are mobile applications that offer different services and products in the same environment (e.g., goods deliveries, social networks, and financial services), collect a large amount of alternative data (Siddiqi, 2017) that are generated by the use of the given application and are supplementary to the traditional financial data. Several researchers have found that the use of alternative information is valuable in the financial sector because it allows for improvement in the performance of some models; for instance, Roa et al. (2021) showed that the inclusion of variables such as the number of payments with errors and orders paid with the superapp's own credit cards can add significant predictive value in the problem of default prediction.