Optimization models for interpretable analytics.

Companies are still far from capturing the whole potential of data analytics. According to a McKinsey & Company report launched on December 2016, the biggest barrier is the struggle to incorporate data-driven insights into day-to-day processes. This project aims to tackle this issue by developing new methodologies to estimate models that are easier to interpret than the current state-of-the-art methods.

Since the bankruptcy of Lehman Brothers almost one decade ago, the banking sector is facing a rise in regulatory requirements. Society is losing confidence in the sector due to a myriad of scandals and is becoming suspicious of the intensive use of methodologies that are seen as magician tricks by the general public. Even practicing statisticians at banks have more confidence in interpretable models and methods than in the new technologies even though they are often more accurate. Finally, when a loan is rejected, the applicant has the right for an explanation. Given the special importance of interpretability in the banking sector, our project will pay special attention to three clear applications: LGD, time to default and probability of default.

This project aims to mitigate one of the main barriers hindering full application of analytics in the day-to-day business environment: the lack of interpretability of the state-of-the-art predictive methods.

A clear industry impact is sought directly through the engagement of banks and credit companies. However, the developed methodologies have a much wider.

According to a recent study, 'a potential economic impact of $110 billion to $170 billion is estimated in the retail banking industry in developed markets and approximately $60 billion to $90 billion in emerging markets'. Our project will contribute to the achievement of this potential by creating methods that deal with complex data, but do not act as black boxes, which can be difficult to scrutinize or interpret. On the contrary, the proposed methods will enhance the interpretability of the final model, and will be accompanied by visualization tools that help the user to get business insight.