An Extreme Value Model For the Analysis of the COVID-19 Pandemic and Its Impact, and the Mitigation of Future Related Crises

Inference and risk prediction are of critical importance to policymakers, financial and insurance companies, and society. They are especially important in environments that experience strongly disruptive and typically unprecedented events, such as the current twin shocks of the COVID-19 pandemic and the oil price crash. There are few statistical techniques for analysing extreme events in a non-stationary, time-dependent world available, and they are generally limited to data that either have a low number of features per sample or are not scalable to the complex data which could have hundreds or thousands of features for each sample of the kind that is currently routinely collected. Dr. Gilles Stupfler, AXA Research Fund Award recipient at ENSAI, seeks to push the boundaries of current knowledge by addressing this problem of modelling the impact of risks related to highly disruptive events such as the COVID-19 pandemic, in the absence of relevant data, through the development of adapted techniques. His project is made of two research strands.

The first strand of activity will involve inference on non-stationary extremes, with a view to modelling the large economic and financial consequences of the global spread of COVID-19 given auxiliary geographical, sanitary, and climate information. He and his team will develop a general methodology using various risk metrics, including recently introduced expectile and extremile-based risk measures that provide sensible alternatives to traditional quantile-based risk measures (e.g., Value-at-Risk and Expected Shortfall) when dealing with rare events that entail severe losses.

The second strand of work will harness the vast wealth of functional data at our disposal to construct estimators for extreme regression risk measures with functional covariates, with a view to analysing and predicting the aftermath of the COVID-19 pandemic. These estimators are so-called “nonparametric estimators”, in the sense that they progressively learn the structure of the data with minimal input from the user. The methods will be implemented on simulated and real data from climatology, epidemiology, insurance, and finance. This requires specialised technical skills and cutting-edge results from extreme value theory.

With this combination of novel risk measurement techniques and high-dimensional data methods, this research program will provide a diverse toolbox for the assessment and mitigation of extreme risk due to rare events, including a pandemic or a disruptive climate event.