The research group in Mathematical Statistics and Data Science studies advanced methods and models for analysing and representing data. We employ probability theory and stochastic processes to rigorously model uncertainty and randomness, and abstract and linear algebra to understand the structure of statistical models and the relationships between their parameters.
 |
Pauliina Ilmonen Professor Multivariate extreme values, functional data analysis, cancer epidemiology |
 |
Kaie Kubjas Associate Professor Algebraic statistics |
 |
Lasse Leskelä Associate Professor Mathematical statistics, network analysis, probability theory |
 |
Vanni Noferini Associate Professor Network analysis, random matrix theory |
 |
Jukka Kohonen University Lecturer Statistics, combinatorics |
 |
Pekka Pere University Lecturer Statistics |
 |
Jonas Tölle University Lecturer Stochastic processes, probability theory |
Predicting food aid demand requires balancing theoretical soundness with computational practicality while handling the messiness of zero-inflated humanitarian data. This thesis develops a natural gradient boosting approach for probabilistic demand forecasting by implementing the Extended Quasi-Likelihood (EQL) function within NGBoost, enabling native support for Tweedie-parametrized Compound Poisson-Gamma distributions without the computational barriers of exact likelihood estimation. The method generates full predictive distributions, which allows supply planners to quantify uncertainty and make risk-aware decisions in prepositioning and predictive procurement. The approach is evaluated on handover data from the World Food Programme (WFP) from April 2024 to March 2025, along with backtesting across six years. Results show conservative calibration errors of for central quantiles and relatively strong performance at extreme values where humanitarian logistics decisions are most consequential. While point prediction accuracy doesn't outperform the benchmark LGBM approach, the superior distributional calibration directly addresses WFP's stated need for quantified uncertainty in supply chain planning and makes the trade-off justified. The work bridges quasi-likelihood statistical theory with practical machine learning, allowing feasibility of learning on standard hardware, as opposed to the full likelihood approximation methods of these kinds of distributions, which are too heavy for repeated computations. However, the theoretical regularity (in the Riemannian sense) of the quasi-likelihood statistical manifolds, which is proposed in this work, remains unproven, leaving important theoretical questions about why the approach works empirically.
 |
 |
 |
Page content by: webmaster-math [at] list [dot] aalto [dot] fi