2021/03/05 by Michel Denuit, Denuit, Michel, Arthur Charpentier +3 · 1 citation
Decision Sciences · Economics, Econometrics and Finance · Social Sciences · #Econometrics (econ.EM) #FOS: Computer and information sciences #FOS: Economics and business #Insurance, Mortality, Demography, Risk Management #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Probability and Risk Models #Stochastic processes and financial applications
paper · pdf · doi:10.48550/arxiv.2103.03635
openalex publication_date 2021/03/05 · openalex created_date 2022/09/20 · openalex updated_date 2026/07/28
Boosting techniques and neural networks are particularly effective machine\nlearning methods for insurance pricing. Often in practice, there are\nnevertheless endless debates about the choice of the right loss function to be\nused to train the machine learning model, as well as about the appropriate\nmetric to assess the performances of competing models. Also, the sum of fitted\nvalues can depart from the observed totals to a large extent and this often\nconfuses actuarial analysts. The lack of balance inherent to training models by\nminimizing deviance outside the familiar GLM with canonical link setting has\nbeen empirically documented in W "uthrich (2019, 2020) who attributes it to the\nearly stopping rule in gradient descent methods for model fitting. The present\npaper aims to further study this phenomenon when learning proceeds by\nminimizing Tweedie deviance. It is shown that minimizing deviance involves a\ntrade-off between the integral of weighted differences of lower partial moments\nand the bias measured on a specific scale. Autocalibration is then proposed as\na remedy. This new method to correct for bias adds an extra local GLM step to\nthe analysis. Theoretically, it is shown that it implements the autocalibration\nconcept in pure premium calculation and ensures that balance also holds on a\nlocal scale, not only at portfolio level as with existing bias-correction\ntechniques. The convex order appears to be the natural tool to compare\ncompeting models, putting a new light on the diagnostic graphs and associated\nmetrics proposed by Denuit et al. (2019).\n