Clara Higuera, lead data scientist and project manager for BBVA’s new predictive debt mitigation models, talks to Global Finance about key steps toward AI innovation.
Global Finance: BBVA recently introduced a machine learning (ML) pipeline for early debt repayment. What problem does this innovation solve?
Clara Higuera: When clients begin to have difficulties repaying loans, this is a double problem – for the client and for the bank. Just two years ago, the debt collection process was entirely manual, which delayed offering solutions to clients.
We currently have five machine learning models that are either in development or in production. For example, one model predicts the likelihood that a customer will experience difficulties (such as being unable to repay a loan) even if they are up to date on their payments. Another model is a model for predicting long-term default, that is, a default lasting two years or more for a client that is already in default.
These models allow the bank to offer solutions sooner—to many customers within a month, whereas previously it might have been a year before a solution was offered to those customers.
GF: What solutions or interventions can a bank take if a loan turns out to be problematic?
Higuera: Varies. This may include a financial advisor calling the client or suggesting a refinancing solution. ML models also rank clients based on their criticality, so financial advisors can prioritize which clients to contact first.
GF: Your proposal uses the XG-Boost non-linear boosting algorithm. Nonlinear machine learning algorithms have been criticized for their lack of interpretability. That is, it may be difficult for a bank to explain in everyday business language how it came to such a decision – internally or to a client. Is this worrying BBVA?
Higuera: Currently, we mainly use tree models such as XG-Boost, which are less interpretable than logistic regression. [which the bank used previously]. But for this reason, we have an assessment and interpretation module in development that will help us visualize and understand the most important variables at the global and local level.
Additionally, these models will not be used directly for the client, but rather are intended for our financial managers to manage debt. We don’t use the most uninterpretable models, such as neural networks, because we need some interpretability.
Finally, we meet with business units to review case studies with them, which helps [explain decisions and] identify potential errors.
GF: Does the machine learning pipeline require a “culture change” within BBVA before an innovation can be developed?
Higuera: The bank has been using analytics for a long time and already has an established operating system. However, this new model took some time to create and we had to build trust to achieve our goal. This was probably the hardest part because we had to live with a lot of uncertainty.
We had to approach this by integrating traditional methodology. [i.e., logistic regression] with the new [XG-Boost] and report performance comparisons so we can reassure everyone that it works.
It was also important that the formed team included data specialists from the bank’s risk analysis department and our AI Factory division. As new modules were added to the pipeline, progress was constantly discussed.
GF: Is there any one of the five models that has proven particularly useful?
Higuera: We view each of them as a funnel that helps prevent customers from moving from a critical state to an even more critical one. Each model provides information that can slow progress toward default. The project is conceived as a holistic one – to help clients earlier.
GF: What organizational and innovation lessons have you learned?
Higuera: The value that data and AI can bring to your organization is significant. Communication between the technical and non-technical sides of the organization is also critical. Teamwork, a fundamental value of BBVA, is also important.