Previous seminars 2026

Spring

Review Seminar 2026-01-14: Jakob Torgander

Speaker Jakob Torgander, Department of Statistics, Uppsala University

Opponent Mattias Villani, Department of Statistics, Stockholm University

Topic SimplexMCMC: An Auxiliary-Variable Framework for Mixed Discrete–Continuous Posterior Distributions

Abstract This paper introduces SimplexMCMC, a Markov chain Monte Carlo (MCMC) framework for sampling from mixed continuous-discrete posterior distributions. Discrete sampling is reformulated as sampling from the interior of the probability simplex, leading to an auxiliary-variable MCMC construction that admits the target discrete distribution as its stationary distribution. The framework is extended to Bayesian models with mixed discrete and continuous parameters and admits an HMC-based instantiation that makes use of the Gumbel–Softmax distribution. This instantiation, referred to as Relaxed Discrete Hamiltonian Monte Carlo (RD-HMC), provides a practical algorithm for approximate inference in mixed continuous-discrete models. Correctness of the SimplexMCMC framework is established through theoretical convergence results, and the practical behaviour of RD-HMC is assessed empirically on simulated and real data.

STATISTICS Review Seminar 2026-02-11: Väinö Yrjänäinen

Speaker Väinö Yrjänäinen, Department of Statistics Uppsala University

Opponent Pierre Nyquist, Department of Mathematical Sciences, Chalmers University of Technology and Gothenburg University

Abstract Data accuracy is crucial for reliable research, accurate decision-making, and high-performance machine learning. However, maintaining data accuracy, especially at scale, is complicated and diﬀicult to verify. By integrating concepts from software engineering, statistical quality control, and branching process theory, we formalize an iterative data curation framework that scales well to large data sets. We go on to prove that the proposed approach asymptotically eliminates all errors in the data with probability one. Additionally, we provide theoretical guarantees that data accuracy tests speed up error reduction. We corroborate these results through simulations on text and tabular data, and a real-world application to the Swedish Parliamentary Corpus, demonstrating the framework’s effectiveness in preserving high-accuracy historical records at scale.

STATISTICS Seminars Series 2026-03-11: Tilman Bretschneider

Speaker Tilman Bretschneider, Department of Economics, Lund University

Topic Factor-based imputation of missing values using cross-section averages

Abstract There is a fast-growing literature developing factor-based imputation methods of missing values in panel datasets. These methods typically estimate factors by applying principal component analysis to an observed subset of the data. In this paper, we propose a new method that uses cross section averages as factor estimates, obtains loadings by linear projections, and imputes missing values with their inner product. The main motivation behind our approach is its straightforward implementation and flexibility to accommodate general missing patterns. We derive the asymptotic properties of estimated factors, loadings, and imputed values while allowing the true number of factors to be overspecified. We showcase the good small sample properties and performance in a Monte Carlo study and in an imputation exercise with macroeconomic data. Finally, we illustrate how our method can also be used to construct monthly state-level GDP indexes from monthly macroeconomic data for a panel of U.S. states.

STATISTICS Industry Seminar 202603-31: Andreas Dahlström

Speaker Andreas Dahlström, Thalius AI

Topic Editable & Explainable Embeddings: Sliding Search, Taste-Vectors, and Early Semantic Arithmetic

Abstract Thalius is a deep-tech AI startup building methods for editable and explainable embeddings. In this session, we will demo a new interaction paradigm for semantic retrieval: sliding search, where users move gradually through an embedding space between arbitrary concepts instead of issuing discrete queries. We’ll show how this enables controllable product discovery and matching via learned “tastes” (preference vectors) that can be combined and tuned. Finally, I’ll share early results on semantic arithmetic: constructing, applying, and validating transformation directions to add, remove, or replace attributes in embeddings in a more predictable way. The goal is practical: make embedding systems that are not just accurate, but steerable, inspectable, and easier to debug.

STATISTICS Seminars Series 2026-04-15: Yingfu Xie

Speaker Yingfu Xie, Swedish Financial Supervisory Authority (Finansinspektionen)

Topic Time series forecast with statistical and Machine Learning models: a practitioner’s experience

Abstract In this talk, we go through the basic RegARIMA models for seasonal adjustment and forecasting, as well as a couple of ML models (LSTM and VAE) for time series. The talk will not be theoretical, but practical, and we will share some experiences participating in the European Statistics Awards Nowcast competition and possible issues for future research.

STATISTICS Seminars Series 2026-05-13: Kristofer Månsson

Speaker Kristofer Månsson, Department of Economics and Statistics, Linnaeus University

Topic Nonlinear forecasting with many predictors using mixed data sampling kernel ridge regression models

Abstract Policy institutes such as central banks need accurate forecasts of key measures of economic activity to design stabilization policies that reduce the severity of economic fluctuations. Therefore, this paper develops a kernel ridge regression estimator in a mixed data sampling framework. Kernel ridge regression can handle many predictors with a nonlinear relationship to the target variable. Consequently, it has potential to improve the currently used principal component-based methods when the economic data follow a nonlinear factor structure. In a Monte Carlo study, we show that the kernel ridge regression approach is superior in terms of mean square error and is more robust than principal component-based methods to different nonlinear data generating processes. By using a dataset consisting of 24 economic indicators, we forecast Swedish gross domestic production. The results confirm the superiority of the kernel ridge regression approach. Therefore, we suggest that policy institutes consider the use of kernel-based approaches when forecasting key measures of economic activity.

Autumn

STATISTICS Seminars Series 2026-09-02: Yousef Kaddoura

Speaker: Yousef Kaddoura, Örebro University School of Business

Title: Estimating and Testing Kinks in Panel Data Models

Abstract: Many economic and financial relationships may change gradually rather than abruptly. We study panel data models in which the coefficient vector is continuous and piecewise linear in calendar time, with a finite number of unknown kink dates at which its slope changes. We propose a penalised least squares estimator that applies adaptive weighted group penalties to the second differences of the coefficient path, and develop asymptotic theory showing that it recovers both the number and the locations of the kinks with probability approaching one. To our knowledge, this is the first panel framework to estimate an unknown number of common kink dates in a time-varying coefficient path under fixed effects. We establish that endpoint slopes converge at the usual cubic regime-length rate and interior slopes at rates determined by their own and adjacent regime lengths. We also develop a coefficient-by-coefficient extension allowing individual regressors to kink at different dates. Monte Carlo evidence supports the good finite sample properties, and we illustrate the method through an application in macro-finance, specifically the relationship between debt and growth.

Here is the arXiv version of the paper https://arxiv.org/abs/2608.07162

Extra Seminar2026-09-02: Alexander Terenin

Speaker: Alexander Terenin, Independent Researcher, prev. Assistant Research Professor, Cornell University

Topic: A Mini-introduction on Bayesian Decision-making Algorithms

Abstract

Making decisions in the face of uncertainty is a ubiquitous part of life and intelligent behavior. To quantify uncertainty, one can introduce a stochastic model, and apply the paradigm of Bayesian learning: once this is done, the question becomes how to resolve the uncertainty in order to act. In this context, a consequence of uncertainty is the emergence of explore-exploit tradeoffs - that is, of tradeoffs between gathering new information, and using the information you’ve already gathered to achieve your goals. Model-based algorithms which carefully balance these tradeoffs are known as Bayesian decision-making algorithms, and provide a candidate paradigm for understanding optimal exploration in reinforcement learning, large language model problem-solving, and other areas. This talk follows my work-in-progress book (https://bayesianalgorithms.com/) on the topic, and will provide an introductory overview of the key notions and challenges in this research area.

Bio

Alexander Terenin, until extremely recently, was an Assistant Research Professor at Cornell University. Before this, he was with the University of Cambridge, and received his PhD from Imperial College London in 2022. His research focus is on decision-making under uncertainty. He is known for his work on probabilistic machine learning, where he made core contributions to Gaussian process scalability and co-developed the area of geometric Gaussian processes. These contributions include numerical methods which are now used in production in industry. He has picked up numerous honors for his work, including a best paper honorable mention at ICML and a best student paper award at AISTATS.

STATISTICS Seminars Series 2026-09-09: Jimmy Olsson

Speaker: Jimmy Olsson, Department of Mathematics, KTH

Topic: Online variational inference for dependent data

Abstract: Variational inference has become a cornerstone of modern machine learning, providing a scalable framework for parameter estimation and posterior approximation in complex latent-variable models. Most variational methods, however, are designed for batch settings, while in many applications observations arrive sequentially in a data stream.

In this talk, I will present a framework for amortised online variational inference in state-space models. The approach is based on maximising an asymptotic evidence lower bound (ELBO), obtained as the long-run average of finite-sample ELBOs. Recursive representations of the ELBO and its gradient enable online optimisation of this asymptotic objective using stochastic approximation. Combined with Monte Carlo sampling and flexible neural architectures, the resulting method supports simultaneous online learning of model parameters and the latent-state posterior distribution. Finally, I will discuss the theoretical foundations of the framework and demonstrate its effectiveness on both synthetic and real-world data.

STATISTICS Seminars Series 2026-09-16: Sachi Dilami Ilangasekara

Speaker: Sachi Dilami Ilangasekara, Department of Economics and Statistics, Linnaeus University

Topic: Feasibility failure in VaR-implied risk-aversion estimation: a simulation comparison of classical, Stein-type and non-Stein regularized plug-in estimators

Abstract: This paper studies the finite-sample feasibility and stability of Value-at-Risk (VaR)-implied risk-aversion coefficient estimators under quadratic and exponential utility functions. Existing closed-form expressions link VaR-based portfolio choice to implied risk aversion, but their empirical use requires estimation of the mean vector and inverse covariance matrix. In large-portfolio and high-concentration settings, these plug-in quantities may become unstable, and the estimated feasibility condition may fail. In such cases, the implied risk aversion coefficient is undefined. We treat this event as Method Failure (MF) and report its frequency as a primary performance measure, together with estimation accuracy conditional on feasibility. The study compares the classical sample plug-in estimator with Stein-type shrinkage and regularized estimators, as well as Serdobolskii-type non-Stein precision regularization. Monte Carlo simulations are conducted under multivariate normal returns across different portfolio dimensions, concentration ratios, covariance structures, and VaR confidence levels. The simulation evidence is complemented by an empirical application to S&P 500 constituent stocks using weekly and monthly log returns. The results show that feasibility failure is not only a numerical inconvenience but an important feature of estimator performance. Regularization, especially of the precision matrix, substantially improves stability in large-portfolio settings.

STATISTICS Seminars Series 2026-09-23: Marcus Piil Petersen

Speaker: Marcus Piil Petersen, Department of Mathematical Sciences, University of Copenhagen

Topic: The Linear Path-Dependent Volatility Model

Abstract: I introduce a new path-dependent volatility (PDV) model, which shows promising performance in three respects: (1) Tractability: Volatility and log-price are linear mappings of a high-dimensional polynomial process, yielding closed-form expressions for all cross moments and term structures. (2) Realism: The model is intuitive and replicates stylized facts, including roughness, the leverage effect, and volatility clustering. (3) Predictive power: The model’s forecasts of implied and realized volatility across equity indices are on par with or exceed those of state-of-the-art models from the literature.

STATISTICS Seminars Series 2026-09-30: Mikkel Sølvsten - Cancelled!

Speaker: Mikkel Sølvsten, Department of Economics and Business Economics, Aarhus University

Topic: Approximately Parallel Trends: Event Study Inference under Trend Violations

Abstract: Difference-in-difference designs based on parallel trends have become a favorite tool for applied researchers estimating causal effects. Often however, they are applied in settings where exact parallel trends can be hard to justify and/or where pretend tests show small but statistically significant violations of parallel trends. To rationalize this behavior, we recast difference-in differences designs as relying solely on an approximate version of parallel trends, which allows parallel trends to fail in some realizations of the data. Our reformulation delivers new inference procedures that account for uncertainty about possible deviations from parallel trends.

STATISTICS Seminars Series 2026-10-07: Stepan Mazur

Speaker: Stepan Mazur, Handelshögskolan vid Örebro University

Topic: Likelihood ratio test for covariance matrix under multivariate 𝑡 distribution with uncorrelated observations

Abstract: In this paper, estimators for the unknown parameters under two types of matrix-variate 𝑡 distributions are determined, and their basic statistical properties, including bias and sufficiency, are investigated. These estimators are then applied to test hypotheses concerning the covariance structure of a multivariate 𝑡 distribution associated with a collection of uncorrelated, though not necessarily independent, observation vectors, using two types of matrix-variate 𝑡 distributions. A likelihood ratio test is proposed, and its distributional properties under the null hypothesis are examined, assuming either a fully specified covariance matrix or one specified up to a constant. Furthermore, it is demonstrated that the asymptotic distribution for the type I matrix-variate 𝑡 distribution under both hypotheses coincides with that under the normality assumption. Finally, for testing a fully specified covariance matrix, the asymptotic distribution of the likelihood ratio test statistic is determined.

STATISTICS Seminars Series 2026-10-14: Leoni Grossmann

Speaker: Leoni Grossmann, Rück Research Group, Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institute

Topic: Suicide risk after involuntary psychiatric care, descriptive and predictive findings

Abstract: I am planning to talk about three studies from my ongoing PhD project on suicide prediction. The papers are i) a descriptive study on suicide risk among individuals discharged from involuntary psychiatric care, ii) risk indicators for suicide, and iii) risk prediction for adverse outcomes (suicide, self-harm and involuntary readmission) after discharge from involuntary care.

Review Seminar 2026-10-21: Chamika Porage

Speaker: Chamika Porage, Department of Statistics, Uppsala University

Opponent: Camila Olarte Parra, Institute of Environmental Medicine, Karolinska Institutet

Topic: FPGS: An R Package for Causal Inference using Full Prognostic Scores

Abstract: Full prognostic score (FPGS) methods provide an approach to confounder adjustment to estimate average treatment effects by summarizing pretreatment covariates through their associations with potential outcomes under both treatment conditions. This paper introduces the FPGS package for R, which provides tools for estimating the FPGS and a selection of treatment-effect estimators. The package implements regression imputation, matching, stratification, inverse probability weighting, normalized inverse probability weighting, and augmented inverse probability weighting estimators. Continuous and binary outcomes are supported. The FPGS can be estimated parametrically using linear or logistic regression or using random forests with cross-fitting. In addition, externally estimated FPGS components can be supplied directly, allowing treatment-specific prognostic models fitted using other methods to be used with the treatment-effect estimators implemented in the package. Since the estimators condition on an estimated rather than a known score, we state the properties of the estimated score that the implemented methods rely on. The package provides analytic variance estimation for parametrically estimated FPGS models and bootstrap inference for supported estimators. Its functionality is illustrated using continuous- and binary-outcome applications and a simulation-based example in which the FPGS components are estimated externally using XGBoost. The FPGS package provides a unified implementation of FPGS-based methods for causal effect estimation in R.

STATISTICS Seminars Series 2026-11-04: Gilbert Mutungi

Speaker: Gilbert Mutungi, Makerere University, Uganda

Topic: Time Series Features and Forecasting

Abstract: Selecting an appropriate forecasting approach remains a central challenge, as no single method consistently performs best across different time series. Conventional model selection and forecast combination approaches face challenges related to accuracy, computational demands, and generalizability. Using the M4 forecasting competition dataset, we explore how statistical features relate to forecasting model performance and how these relationships can inform model selection and combination. We further investigate whether knowledge learned from M4 can be transferred to improve the forecasting of Uganda’s macroeconomic indicators. Results demonstrate the potential of archive-based forecasting knowledge transfer, with the proposed probabilistically weighted combination achieving the best overall performance among the selection and combination strategies considered.

STATISTICS Seminars Series 2026-11-11: Simon Steiger

Speaker Simon Steiger, Department of Medicine, Karolinska Institutet

Topic Generative modelling in clinical epidemiology

Abstract In my work at the clinical epidemiology division at Karolinska Institutet, I study rheumatic diseases with the help of Swedish health registries. Linked via personal numbers, these health registries provide many unique research opportunities. However, the statistical training of clinicians and epidemiologists predisposes the field to certain types of analyses, which—as I will argue—leave many impactful questions unanswered. I will discuss probabilistic generative modelling as a toolset that allows domain experts and statisticians to tackle these challenging research projects together.

STATISTICS Seminars Series 2026-11-18: Ye Tian

Speaker: Ye Tian, Handelshögskolan, Umeå University

Topic: Measuring episodic memory longitudinally with a reduced test battery: a test equating approach

Abstract Longitudinal studies of episodic memory often rely on composite scores derived from multiple tasks. When some tasks are no longer administered at later time points, scores from the reduced test battery are not directly comparable with earlier scores. Test equating can link scores from different instruments, but it typically requires equivalent groups, common items, or background covariates. We propose incorporating previously observed test scores from longitudinal data into kernel equating under a nonequivalent groups with covariates design, in which propensity scores balance the groups that completed the full and the reduced battery. The method was evaluated with data from the Betula study, using a validation wave at which both batteries were available, and in a simulation study with different attrition patterns and instruments. In the empirical study, the proposed method yielded residual errors close to those of single-group equating and z-score standardization, and smaller than those of a Betula-specific transformation, while providing equated values over the whole score range. In the simulation study, it was more accurate than transformation and linear mixed-effects predictions when the two instruments differed in difficulty. A propensity score model based only on historical scores performed comparably to models with demographic covariates.