Welcome at ISBA !
The Institute of Statistics, Biostatistics and Actuarial Sciences (ISBA) of the Université catholique de Louvain (UCLouvain) is a research centre of high international reputation.
ISBA collaborates with the teaching unit Louvain School of Statistics, Biostatistics and Actuarial Sciences (LSBA) and the technological platform Statistical Methodology and Computing Support (SMCS).
Upcoming Events at ISBA
EOS workshop - UCLouvain/KULeuven
27/03/2025 - 08:30 - D.251 -
EOS workshop
Program
8h30 : Yevhen Havrylenko, Lausanne University(45 min)
Synthetic tabular data for ratemaking: deep generative models vs imputation-based methods
Actuarial ratemaking depends on high-quality data, yet access to such data is often limited by the cost of obtaining new data, privacy concerns, and competitive sensitivity. In this talk, we explore synthetic-data generation as a potential solution to these issues. In addition to generative methods previously studied in the actuarial literature, we explore and benchmark fundamentally different class of approaches, which is based on Multiple Imputation by Chained Equations (MICE).
In a comparative study using an open-source dataset, we evaluate MICE-based models against other generative models like Variational Autoencoders and Conditional Tabular Generative Adversarial Networks. We assess each model on three dimensions. First, we assess data fidelity, namely how well the synthetic data preserves marginal distributions and multivariate covariate relationships. Second, we evaluate data utility, measuring the consistency of Generalized Linear Models (GLMs) trained on synthetic versus real data. Third, we assess the ease of use of each method by analyzing its implementation complexity, and the need for customization and finetuning. Furthermore, we investigate the impact of generically augmenting the original data with synthetic data on the performance of GLMs for predicting claim counts. Our results highlight the potential of MICE-based methods in creating high-fidelity tabular data while offering lower implementation complexity compared to deep generative models.
This research talk is based on a joint work with Meelis Käärik and Artur Tuttar, both affiliated with the University of Tartu. The pre-print version of our paper is available at https://arxiv.org/abs/2509.02171
9h15 : Paul Wilsens, KULeuven (30 min)
Claim dynamics of weather-related insurance claims
Weather events account for a substantial share of property-related insurance claims, making careful modelling of such claims crucial for prudent portfolio management. Highly granular open-source weather-related data are increasingly available at short notice, enabling new opportunities for improved claim development and reserving. In this paper, we provide a methodology to handle the claim dynamics of weather-related property insurance claims, i.e. the occurrence and reporting processes, and apply it to a real insurance portfolio. Specifically, we apply an EM-XGBoost framework, incorporating publicly available weather-related covariates such as precipitation and wind speed, enriched with date- and policyholder-specific information. Weather covariates are used directly to model the occurrence of claims, as well as indirectly to construct different reporting regimes for claims under varying weather conditions. Using data from multiple public sources, we assess the use of observational versus reanalysis weather data for both claim occurrence and reporting. Additionally, we examine the impact of temporal and spatial aggregation on model accuracy for IBNR reserving. This allows us to identify aggregation levels that still yield reliable estimates of occurrence and reporting dates while providing insights at both the individual and portfolio levels.
9h45 : José Miguel Flores-Contro, UCLouvain (30 min)
Linear Risk-Sharing of Losses at Occurrence Time in the Compound Poisson Surplus Model
Denuit and Robert (2023) study a conditional-mean risk-sharing scheme within the classical Cramér–Lundberg risk process, where losses are allocated among pool participants at the time of occurrence. They show that this scheme reduces the infinite-time ruin probability of each participant, highlighting the benefits of risk sharing. However, the conditional-mean approach is not easily interpretable and may be difficult to implement in practice, particularly in contexts where transparency is essential, such as low-income communities using community-based insurance. Motivated by these considerations, we propose a linear risk-sharing rule in which each participant covers a fixed fraction of losses incurred by the pool. We show that when individual losses belong to the same scale-family or share a common mean, pooling under this rule also reduces the infinite-time ruin probability for each participant. This provides a more intuitive and practical risk-mitigation mechanism. Numerical examples illustrate and support our findings.
coffee break (30 min)
10h45 : Mathias Lindholm, Stockholm University (45 min)
Discrimination, fairness and how this can be measured
We will start by going through different notions of fairness in relation to proxy-discrimination. This includes both classical definitions from the algorithmic fairness literature, but also more recent contributions more directly targeting the insurance pricing setting. A natural follow up question relating to defining fairness and discrimination is how this should be measured. Apart from presenting recent contributions on this we will also discuss ongoing work in this direction.
11h30 : Freek Holvoet, KULeuven (30 min)
Deep learning methods for loss simulation in a multi-peril insurance portfolio
In a portfolio where each policyholder is exposed to losses under multiple perils, calculating a value-at-risk requires an understanding of the dependency structure among perils. In this work, we investigate deep learning architectures to model the joint distribution of multi-peril losses. We explore two frameworks: a direct estimation approach using conditional variational autoencoders, and a flexible two-step approach that decouples the marginal distributions from the dependency structure. For the marginals, we utilize neural networks based on the Tweedie compound Poisson distribution, paying special attention to the mixed discrete-continuous nature of insurance losses and the probability mass at zero. For the dependency structure, we investigate a normalizing flow model called RealNVP, which allows for exact density estimation via invertible transformations of pseudo-residuals. We validate these methodologies using synthetic data to benchmark their ability to recover the underlying joint density and aggregate risk measures against known ground truth.
12h00 : Charlotte Jamotton, UCLouvain (30 min)
A Multi-Criteria Fair Gaussian Regressor for Insurance Premium
This article studies how multiple notions of fairness can be incorporated into a single Bayesian non-parametric regression framework for insurance pricing, with a focus on claim frequency modeling under a log-link. We consider a Generalized Gaussian Process Regression (GGPR) model for count data with risk exposure and introduce fairness interventions in its architecture. Specifically, we address notions of individual fairness by altering the kernel structure to control the similarity between policies (e.g., to mitigate omitted variable bias). We also address group-level fairness by enforcing demographic parity through linear constraints affecting the posterior. This modified GGPR architecture allows us to jointly enforce multiple fairness definitions, spanning both group and individual-level criteria, within a single probabilistic model. We empirically explore trade-offs with actuarial fairness, and how different fairness criteria interact when combined. The results highlight the importance of adopting a multi-criteria, context-aware approach to fairness in insurance pricing.
