Skip to Main content Skip to Navigation
Theses

Contributions de l’Apprentissage Statistique à l’Actuariat et la Gestion des Risques Financiers

Abstract : The continuous enhancement of computer performances during the last decades has favored a widespread application of statistical learning theory in multiple domains. Actuaries, long-standing statistical experts, notably turn more and more frequently to these new algorithms for the evaluation of risks they are facing. Thus, in this thesis, we examine how the integration of methods derived from statistical learning can contribute to the development of actuarial science and risk management through the study of three independent problematics, preliminarily presented in a general introduction. The first two chapters propose new mortality forecasting models within the longevity risk evaluation framework that insurance companies and pension funds encounter. Chapter 1 focuses on the singe population case, while Chapter 2 extends the study to the multi-population. In both situations, high-dimensionality appears to be a major concern. We tackle this issue thanks to a penalized vector autoregressive (VAR). This model is directly applied to the mortality improvement rates in the first chapter, and to the time series derived from Lee-Carter’s model fits in the second one. The elastic-net penalization preserves the large freedom in the spatio-temporal dependence structure offered by the VAR, while remaining sparse in terms of estimated parameters, thereby avoiding overfitting. In Chapter 3, we analyze lapse risk in life insurance policies through the use of supervised classification algorithms. We apply, among others, support vector machine (SVM) and extreme gradient boosting (XGBoost). In order to compare performances from one classifier to another, we adopt an economic point of view derived from the marketing literature and based on potential profits of a retention campaign. We insist on the importance of the loss function retained in the statistical learning algorithms according to the pursued objective: the use of a loss function linked to the performance measure leads to a significant enhancement in the application of the XGBoost in our study. In the last Chapter, in the financial risks’ management framework, we study agricultural commodity price dynamics throughout specific trading sessions where governmental reports, that contain valuable information for the agents, are released. We examine the potential of open data, in particular the satellite data on vegetation index made available thanks to the NASA, for market reactions forecast. We then suggest some improvements to be considered before an operational implementation of this forecasting method
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-02496251
Contributor : Abes Star :  Contact
Submitted on : Monday, March 2, 2020 - 7:15:08 PM
Last modification on : Tuesday, March 3, 2020 - 1:37:23 AM

File

TH2019PIETTEPIERRICK.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-02496251, version 1

Collections

Citation

Pierrick Piette. Contributions de l’Apprentissage Statistique à l’Actuariat et la Gestion des Risques Financiers. Gestion et management. Université de Lyon, 2019. Français. ⟨NNT : 2019LYSE1328⟩. ⟨tel-02496251⟩

Share

Metrics

Record views

186

Files downloads

161