Random Forest Model with Prediction Error Decomposition Function

This webinar introduced the IAA AI Task Force and provide an overview of the “Case Studies & Tools” workstream and its objectives.

The session includes a walkthrough of the Task Force's GitHub account and repositories covering how they are structured, how they function, and how you can explore and contribute to them. A curated collection of GitHub case studies is then presented, with a detailed walkthrough of one fully developed case study illustrating how AI can be applied to a real actuarial problem.

The webinar also touches on the GitHub tools repository showcasing the AI tools through short presentation videos. a more detailed walkthrough of developed case study to illustrate how AI can be applied to a practical actuarial problem.

IAA's CAT Modelling Task Force facilitated the webinar with Verisk and Moody's examine the evolving landscape of climate risk and catastrophe (CAT) modeling, exploring how advancements in modeling methodologies are shaping risk assessment across the insurance and reinsurance industry.

In the first part of this webinar, we will analyze the effect of using weighted maximum likelihood estimation for survival models where the assigned weights depend on the sum assured – a common practice in life insurance to focus on the financial consequences of survivorship. In this context, we consider how these estimates may change when more granular mortality models are used that are based on the socioeconomic characteristics of policyholders. During the presentation, we will highlight several practical considerations for developing weighted models for liability estimation and discuss the impact of such approaches on annuity liabilities. To illustrate this, we will present results using a real-world insurance dataset providing survival information on individuals who purchased annuities between 2015 and 2022 in the Netherlands. The analysis will include a financial test of fit to evaluate the performance of models. In the second part of the webinar, we will discuss research findings on the potential influence of socioeconomic factors on annuitants' longevity, based on publicly available, area-level indicators. The analysis will present detailed results on variables that could be useful for modelling policyholder mortality in the Netherlands, and will compare these with the traditional approach of using age, gender and pension information as covariates.


Speaker: Andrey Ugarte Montero
Andrey Ugarte Montero is a postdoctoral researcher at the Research Center for Longevity Risk (RCLR) of the University of Amsterdam. He holds a master’s degree in Actuarial Science from the University of Lisbon in Portugal and a PhD in Actuarial Science from HEC Lausanne in Switzerland. His interests include actuarial modelling in life and health insurance and the financial consequences of longevity within an insurance context. He is also interested in the applications of machine learning and data science in the insurance sector. From December 2013 to August 2017, Andrey worked as an actuarial consultant for the firm EY (Ernst & Young) in Latin America. He’s also pursuing certification as an Associate/Fellow of the UK’s Institute and Faculty of Actuaries


Moderator: Ernst Visser
Ernst is a Dutch actuary working for KPMG Financial Risk Management in The Netherlands. He has a master degree in physics and a master degree in actuarial science. In 2025 Ernst obtained an MBA degree with a thesis investigating the potential for disruption in the insurance industry. Ernst is a fully qualified member of the Dutch actuarial society. He has Over 30 years experience in consulting in the insurance industry

AFIR-ERM Webinar: Optimal Preventive Strategies in Insurance Risk Models. We develop an integrated framework for optimal risk management in insurance that considers preventive strategies and/or risk sharing in risk models. We odel an insurer facing risks whose frequency and severity can be influenced by preventive efforts, as well as by exogenous risk-sharing instruments. Within a dynamic risk model, we formalize the insurer’s problem as either minimizing ruin probabilities or maximizing a mean–variance criterion for the terminal surplus process, while accounting for the implementation costs of preventive measures and the cost of risk transfer. We derive analytical conditions characterizing the optimal prevention efforts and optimal risk-sharing structure, and we study how these choices depend on risk characteristics. Numerical illustrations based on classical risk models show that jointly optimizing prevention and risk sharing can substantially reduce ruin probabilities compared to treating these instruments in isolation.


Speaker: Li Jingchao
Session moderator: Guangyao Liu

This webinar explores how Behavioral Economics can bridge the gap between complex financial products and actual consumer decisions within the pension and insurance industries. By shifting the focus from purely technical designs to the cognitive reality of "System 1" and "System 2" thinking, we will analyze why even the most robust actuarial models often face inertia and low uptake. The presentation moves beyond theory to examine proven strategies from leading global markets who have successfully maximized participation and savings rates. We will discuss how to apply Choice Architecture to overcome deep-seated biases that hinder effective risk protection across both retirement and insurance sector. Ultimately, this talk redefines the actuary's role as a "Decision Architect," capable of building scalable infrastructure that is both mathematically sound and humanly effective in a modern regulatory landscape.


Speaker: Diego Valero Carreras
Session moderator: Abraham Hernández Pacheco

With the advancement of technology, insurance companies are increasingly adopting usage-based insurance (UBI) supported by vehicle telematics. Vehicle telematics refers to data collected from in-vehicle sensors or smartphone applications during driving, such as speed, acceleration, braking, and steering. It provides a rich, high-frequency record of how a vehicle is driven, offering insights into driving habits, behaviour, safety, and potential risk. However, many current approaches rely on aggregated metrics and do not fully capture the detailed time-series patterns in telematics data. This presentation introduces a flexible framework based on a continuous-time hidden Markov model (CTHMM) to analyze trip-level telematics data directly. Our approach avoids predefined thresholds for harsh events or assumptions about accident probabilities, and uses only telematics data, requiring no traditional demographic covariates. Using an unsupervised anomaly detection technique, we identify deviations from normal driving patterns linked to higher accident risk. The framework is tested on both controlled and real-world datasets, and the results reveal clear behavioural differences between drivers with and without claims, offering practical insights for insurance, accident analysis, and prevention.