Contributed Talks II

Date & Time: Tuesday, September 29, 2026 | 17:00 - 18:40
Place: tba
Chairs: Thomas Sturges

Speaker: Lisa Leimenstoll (Karlsruhe Institute of Technology) 
Title: Boosted Progression
Abstract: Extrapolation beyond the observed range of the data is a fundamental yet difficult problem, especially in applications that require inference on extreme responses. It becomes even harder in multivariate settings, where several covariates may jointly influence the response through nonlinear effects and interactions, making reliable extrapolation substantially more demanding than in univariate problems. Building on the progression principle of Buriticá and Engelke (2024), we develop a method tailored to such multi-dimensional covariate settings. Our approach combines progression-based extrapolation with boosting, thereby increasing flexibility for multivariate problems while preserving the core idea of tail extrapolation. We focus on estimating the conditional median under an additive noise model and assess the performance of the proposed method in simulation studies, including scenarios with interaction effects.

Speaker: Lotta Rüter, Melanie Schienle (Karlsruhe Institute of Technology)
Title: Comparing forecast performance on panel data with unknown cluster structure
Abstract: We introduce a novel Diebold-Mariano type test for evaluating the equal predictive accuracy of forecast models in panel data settings. Our framework accommodates forecast errors that display substantial heterogeneity with unknown and clustered dependence structures in the cross-sectional dimension, as well as serial correlation over time. A key advantage of our approach is that it does not require prior knowledge of the number or composition of clusters and allows for overlapping or non-independent clusters, making it particularly well suited for complex data environments such as financial forecasting.

Speaker: Johannes Bracher, Theo Schäfer (Karlsruhe Institute of Technology) 
Title: Estimating voting behaviour: Does ecological inference actually work?
Abstract: In electoral research, there is often an interest in reconstructing how different groups of the electorate voted. As surveys are expensive and subject to many biases, a variety of statistical methods to recover these patterns from precinct-level data has been suggested. Put simply, these exploit geographical behaviour in the composition of the electorate and election outcomes to identify group-wise voting behaviour. However, this "ecological inference problem" is fundamentally underdetermined, meaning that strong assumptions are needed, which are hard to check in practice. As the true individual voting behaviour usually remains unknown, few empirical validations of ecological inference methods for election outcomes are available. We present work in progress on a validation study on multiple data sets where the true group-wise voting behaviour is known (using e.g., roll call votes and data from ranked choice voting system). Conceiving the ecological inference task as a prediction problem, we apply probabilistic evaluation techniques not previously applied to ecological inference, and explore the potential of model averaging.