Dienstag, 21.10.2025, 12-13 Uhr in W9-109
Dr. David Winkelmann
Universität Bielefeld
Data-Driven Approaches for Reducing Food Waste in Grocery Retailing
Grocery retailers face the trade-off between overstocking, potentially generating food waste, and risking customer dissatisfaction through stock outs. To ensure high service levels, many retailers accept excess inventory, contributing to the approximately 11 million tons of food discarded annually in Germany, over 10% of which originates in the retail sector. In collaboration with the German government, major retailers have therefore committed to halving food waste by 2030. Achieving this goal requires more accurate demand forecasts that capture operational complexities. While demand forecasting has been widely studied by the scientific literature, specific challenges remain. For example, retailers frequently apply price reductions to units nearing expiration: such discounts stimulate short-term customer demand and mitigate the risk for food waste, but if these inflated sales figures are not modelled appropriately, forecasts systematically overestimate future demand and perpetuate overstocking. Using data from a major European grocery retailer, this study aims to contribute to the reduction of food waste in the retail sector by improving forecasts for these specific challenges.
Dienstag, 04.11.2025, 12-13 Uhr in W9-109
Prof. Dr. Dietmar Bauer
Universität Bielefeld
The more, the merrier: the role of exogeneity in global state space models
Often multivariate time series data has matrix time series structure such that at each point in time the same variables are observed for a number of individual entities. Joint unrestricted modelling using vector autoregressions (VAR) contains a prohibitively large number of parameters. This can be countered in a number of fashions (for example, MaTS, aDFM, Bayesian VARs). These approaches, however, do not include structural information on the relations between the various individuals. The global VAR approach of Pesaran and co-workers provides a different idea: It partitions the model into regional models, each of which is estimated and specified separately, allowing for the imposition of structural relations.
Partitioning uses heavily different exogeneity concepts. In this talk we examine the implications of these exogeneity concepts. We also propose how to proceed, if exogeneity is violated.
Dienstag, 18.11.2025, 12-13 Uhr in W9-109
Sophie Potts
Georg-August-Universität Göttingen
Breaking up is hard (to model): Joint models for longitudinal and time-to-event data on divorce data and rare events
In time-to-event analyses in social sciences, endogenous time-varying variables often occur, where the event status is correlated with the covariate’s trajectory. Ignoring this causes biased estimates. In biostatistics, this is addressed with joint models for longitudinal and time-to-event data, which handle endogeneity correctly. Despite their usefulness, joint models remain rare in social sciences. We give an introduction to this method, highlight its advantages for social sciences, and demonstrate it with an example on marital satisfaction and dissolution, comparing results with classical time-to-event models. In a second part, we address rare events in joint models. Standard estimation struggles with few events and imbalanced designs, leading to monotone likelihood. Firth’s correction, which adjusts the score function, provides stable estimates even with very small event counts. We extend this correction to joint models, enabling their use in rare event studies, particularly in observational data where event numbers cannot be controlled.
Dienstag, 02.12.2025, 12-13 Uhr in W9-109
Aktuelle Forschungsbereiche des ZeSt
Dienstag, 16.12.2025, 12-13 Uhr in W9-109
Jan-Ole Koslik
Universität Bielefeld
Fast approximate likelihood inference for indirectly observed hidden Markov models
Hidden Markov models are powerful tools for analysing time series data that depend on discrete underlying but unobserved states. Owing to their flexible hierarchical framework, which separates the observation process from the unobserved state process, they have gained prominence across numerous empirical disciplines. In many applications, however, the data-generating mechanism involves an additional layer: the state-dependent process is itself latent and only observed through a separate measurement model. Examples include movement trajectories observed with GPS error or time series contaminated by an unknown smooth trend. Incorporating this third level substantially complicates inference, as the HMM must be fitted to a process that is not directly observed. We consider modelling such triply stochastic processes, comprising (i) a discrete hidden state process, (ii) a latent continuous process driven by these states, and (iii) noisy observations conditional on the latent process. Parameter inference requires integrating over all possible sequences of hidden states and all possible latent process values consistent with the observed data. This can, in principle, be achieved by combining the forward algorithm for the state process with a Laplace approximation for the latent process. However, practical implementations of the Laplace approximation rely on sparsity in the Hessian of the log-likelihood with respect to the latent process, a property not possessed by the state-dependent process of an HMM. We address this challenge by introducing a banded approximation to the forward algorithm that imposes the required sparsity pattern by setting conditional dependence exactly to zero beyond a certain lag. This enables efficient inference using standard computational tools. The approach is demonstrated through a simulation study of movement data subject to measurement error and a real-data application to the detection of stellar flares in photometric data.
Dienstag, 13.01.2026, 12-13 Uhr in W9-109
Anna Hager
Universität Bielefeld
Titel folgt
Dienstag, 27.01.2026, 12-13 Uhr in W9-109
Pauline Baur
Technische Universität Dortmund
Titel folgt