ISBA Young Researchers Day / YRD

06/02/2026 - 09:00 - ISBA C115 - 

More information coming soon

Mathias Dah Fienon 
Title : "Dynamic risk-parity portfolio with independent components"

Abstract:
Portfolio diversification involves investing in many different securities and types of assets in order to mitigate risk and maximize returns. One approach to achieve such diversification is the factor-risk-parity model. It assumes that the portfolio weights of uncorrelated factors are equal, up to sign. One criticism of factor-risk-parity is that any orthogonal rotation of these factors will also yield uncorrelated factors, the choice of which is arbitrary. However, if the factors are non-Gaussian, then there is only one rotation that gives independent components, or that minimizes a mutual information criterion. We propose a new methodology that searches for maximally independent factors in a dynamic approach by updating the information from the past using a score-driven model. This fits into a general framework for modelling time varying parameters in time series analysis.
 

Mengxue Li
Title: "Learning population and individual structure in dynamic networks with degree heterogeneity"

Abstract:
Dynamic networks provide a powerful framework for characterizing time-varying functional connectivity in neuroimaging studies. In practice, such networks are typically collected from multiple subjects across time and exhibit both temporal dynamics and subject-specific heterogeneity. Brain functional connectivity networks also contain hub nodes, defined as highly connected regions that play critical roles in understanding brain functional connectivity. In this talk, we propose a mixed-effect dynamic stochastic block model with degree heterogeneity, which simultaneously disentangles the population connectivity structure from individual variability and recovers the trajectories of hub regions through time-varying degree parameters. We develop an efficient local approximate estimation procedure and evaluate its performance through extensive simulations and a case study of dynamic functional connectivity from the Human Connectome Project.
 

José Miguel Flores Contro 
Title : "Poverty Trapping: A Ruin Theory Perspective"

Abstract
Trapping refers to the event when a household falls into the area of poverty. Households that live or fall into the area of poverty are said to be in a poverty trap, where a poverty trap is a state of poverty from which it is difficult to escape without external help. Ruin theory, on the other hand, studies stochastic processes and their fluctuations, with its classical application being the modelling of an insurance company’s surplus over time. Since the seminal works of Lundberg
(1903) and Cramér (1930), ruin theory has remained a fundamental area of research in actuarial science. The canonical model in this setting is the Cramér–Lundberg risk process, which has been extensively studied and generalised in the literature. This talk introduces the fundamental principles of ruin theory and examines their application to poverty trapping. In particular, we consider a risk process with deterministic growth and multiplicative jumps, as introduced in Kovacevic and Pflug (2011), to model household capital, incorporating both exponential growth and losses proportional to the current level of capital. Within this framework, we derive closedform expressions for trapping (ruin) probabilities and for the Gerber–Shiu expected discounted penalty function in specific cases, and we discuss how these results relate to the role of policy interventions in the stochastic evolution of household capital. These results illustrate how the mathematical tools of ruin theory provide a stochastic framework for the analysis of poverty dynamics. 

Gabriel Bailly
Title : "Satterthwaite Approximation and Gaussian Time Series"

Abstract
Satterthwaite (1941, 1946) proposed a very simple approximation to the distribution of linear combinations of Chi-squared random variables. It can be used in univariate time series analysis to approximate the distribution of the sample variance and the periodogram of Gaussian time series; we provide Wasserstein bounds and rates of convergence of the approximation towards the true distribution. Similarly, Tan & Gupta (1983) proposed an approximation to the distribution of linear combinations of Wishart random matrices. This, however, has not yet been applied to the framework of multivariate time series: we take advantage of a special case of the matrix normal distribution to propose a feasible approximation to the distribution of the sample covariance matrix of Gaussian time series.

Robert Paulus
Title : "Adaptive regionalization for extreme precipitation: A neural network-weighted independence likelihood approach" 

Abstract
Recent European floods underscore the high cost of underestimating extreme rainfall. Designing resilient infrastructure depends on estimating precipitation return levels: rainfall amounts associated with very rare events. The challenge is that observational records contain only a small number of extremes, so fitting an extreme-value model separately at each site often yields unstable estimates and very wide confidence intervals. Pooling data across space can reduce uncertainty, but it may also introduce substantial bias when nearby locations do not share the same extreme-behavior patterns.
We propose an adaptive pooling strategy that learns how much information to borrow from surrounding sites. For each target location, we fit an extreme-value model using a neural network–weighted independence likelihood, where the network assigns weights to neighboring observations based on distributional similarity. Extensive simulation studies show a clear bias–variance tradeoff: as the network effectively shrinks the sample size from broad pooling toward purely local fitting, return-level errors first decrease and then increase again, indicating that performance is best at an intermediate level of pooling.

LIDAM seminar series / Statistics Seminar by Clement Berenfeld

27/03/2025 - 14:30 - ISBA C115 - 
 

Clement Berenfeld 

(Institut national de recherche en sciences et technologies du numérique) 

Will give a presentation on : 

A causal framework for reliable membership inference attack evaluation

Abstract :  
Quantifying memorization is central for assessing privacy risks in machine learning. The standard tool for this purpose, membership inference attacks (MIAs), traditionally re-quires multi-run evaluations (repeated retraining) that are computationally prohibitive for modern large-scale models. This has led to the adoption of one-run methods (training once with a randomized subset of points) and zero-run methods (evaluating models “as-is”), though their statistical soundness remains unclear. We address this by reframing MIA evaluation as a causal inference problem, defining memorization as the causal effect of a data point’s inclusion in the training set. Our work reveals systematic issues: one-run regimes introduce interferences between jointly inserted points, and zero-run regimes introduce confounding from non-random membership assignments. We formalize these challenges by introducing a new interference model for treatment effect estimation, derive causal counterparts to standard MIA evaluation metrics, and propose estimators that are provably consistent by leveraging learning-theory properties.

LIDAM seminar series / Statistics Seminar by Swati Chandna

28/05/2025 - 14:30 - ISBA C115 - 
 

Swati Chandna 

(Birkbeck, University of London) 

Will give a presentation on : 

Understanding structure in networks observed with covariates via profile least squares

Abstract: 
Network data across various domains (e.g. social, political, biological) are commonly observed with additional node- and/or edge- specific attributes that may explain the intensity of interactions between nodes within the network. For example, in the study of international relations, attributes such as conflict intensity and volume of trade are observed for each country pair, and it is important to understand the extent to which these covariates explain the formation of military alliances between countries (nodes). We study the model where probabilities of edge formation between any two nodes in the network are given by the sum of a linear edge-covariate term and a residual term to model the remaining heterogeneity from unobserved factors. We approach estimation of the model via profile least squares and show how it leads to a simple algorithm to estimate the linear covariate term and the residual structure that is truly latent in the presence of observed covariates. Our framework lends itself naturally to a bootstrap procedure which is used to draw inference on model parameters, such as to determine significance of the homophily parameter or covariates in explaining the underlying network structure. We illustrate the usefulness of our methodology through simulations and application to four real network datasets.