This week, Pierrick Piette defended his PhD thesis in actuarial sciences at the University of Lyon (France) under the supervision of Stephane Loisel and Olivier Lopez. It was a honor for me to be part of his PhD committee and discover how actuarial sciences use quantitative marketing methods to improve their prediction models. Congratulations Pierrick on a fine piece of work!

RETENTION MODELS IN ACTUARIAL SCIENCES

In his thesis, Pierrick shows how the profit-based loss function we developed together with Sunil Gupta in our “Managing Churn to Maximize Profits” paper can be used to improve lapse risk management. His article provides a very convincing replication of the benefits of a profit-based loss function in a very different setting. Loisel and colleagues find that our methodology provides a gain in profits of 26% (1.3 million USD) compared to a classic loss function. These results corroborate the gains we identified in our own research.

You can access their article via the following link: https://arxiv.org/abs/1906.05087

THE FUTURE OF LOSS FUNCTIONS

I am very excited to see that the profit-based loss function has potential in so many different areas and hope to find more replications in other fields soon. The choice of loss function is a feature that is often neglected in model estimation processes. In particular, this choice should match the objectives of the final user of a prediction model. Our approach fits many contexts, not only in marketing, where organizations seek to target a set of individuals with a specific intervention (e.g., catalog, mail, charitable giving, patient compliance, and personalized medicines). Estimating heterogeneous treatment effects is an exciting topic, featured in studies across economics and econometrics (Imbens and Rubin 2015), management (Godinho de Matos, et al. 2017), and computer science (Pearl and Mackenzie 2018). For each application, it is critical to carefully determine the appropriate loss function. When building their own “goal-oriented” loss functions, decision makers should (1) ensure that the margin specifies the true outcome of interest (i.e., goal of the intervention) and (2) use a weighting scheme that prioritizes customers who have the largest impact on the success of the intervention.

A CURE FOR  MANY AILMENTS: FROM PATIENT COMPLIANCE TO THE ALGORITHMIC BIAS

This is relevant even in non-profit contexts, such as for predicting patient compliance with medical treatments. In this case, the loss function could incorporate patient-specific health risks and benefits associated with complying with the medical treatment. In a totally different sphere, it could also be very useful to customize the loss functions of algorithms on the web as a remedy to the so-called “algorithmic bias.” Via a well-thought-out loss function design, one could make sure that the most sensitive minorities do not get discriminated by an algorithm.

Food for thought!