Causality in AI: From Predicting What Happens to Understanding Why It Happens
In artificial intelligence, getting a prediction right or detecting complex patterns does not always mean understanding the underlying events that occurred to lead to that result. Understanding causal relationships is the core idea addressed by the doctoral thesis of Jesús Renero Quintero, who is part of the AI adoption team at the bank. The thesis was developed within the industrial PhD program in Data Science promoted by the University of Navarra and BBVA. The goal is to design AI that is more reliable and transparent.
A Machine Learning model can be very good at anticipating an outcome or identifying patterns in large volumes of data, but in many contexts, such as banking, predicting well is not enough. It is also necessary to understand how the relationship between different attributes or variables influences a prediction, or which hypotheses best explain an outcome. This allows us to audit the results of Machine Learning models and implement AI that is more reliable, interpretable, and transparent.
Jesús Renero’s thesis, titled Explainability-Guided Causal Structure Learning from Observational Data, starts from a basic but decisive distinction: predicting is not the same as understanding. Two variables can move together without one causing the other, and a model can learn useful patterns to anticipate an outcome without identifying the underlying mechanisms that produce it.
A simple example helps illustrate this. Even if we observe that a variable A and a variable B are related, we cannot automatically conclude that A causes B. There may be a third variable C that influences both, or it may be an indirect relationship. The model may make very good use of that association to predict, while still failing to tell us what the actual cause is. The thesis emphasizes precisely this distinction, because in many business contexts this boundary is what separates a useful explanation from a causal hypothesis that still needs to be verified.
To address this issue, the research proposes ReX, a method that seeks to leverage the signals generated by predictive models to construct more structured causal hypotheses. The idea is not to take a prediction and magically turn it into causality, but rather to use the model’s explanations as an initial clue to better organize the relationships between variables. These clues are then refined using rules that help form a plausible causal graph, that is, a representation showing how variables influence one another. In a large corporation such as BBVA, this type of representation helps technical teams, business experts, risk managers, and audit teams speak a common language.
The thesis approach is useful because it connects two worlds that often progress separately. On the one hand, there are advanced AI models capable of learning complex patterns. On the other, there is explainability, which seeks to make visible which variables carry the greatest weight in a prediction. By carefully combining both elements, it is possible to move beyond simple prediction toward a better understanding for analysis and decision-making.
That said, the study itself makes it clear that model explanations do not prove causality on their own. Any causal discovery method requires assumptions, context, and validation, where expert judgment remains essential. This caution is especially important because data is often noisy, incomplete, and nonlinear, and is frequently affected by omitted variables or weak signals. The purpose of the research is precisely to explore a complementary approach, using explainable predictive models to generate causal hypotheses that are more manageable and auditable.
This doctoral thesis is part of the collaboration between BBVA and the University of Navarra to promote industrial PhDs in Data Science, an initiative that seeks to bring advanced research closer to real-world application challenges in the financial sector.