Johan Hallberg Szabadváry
Industrial PhD Student in Computational Mathematics
Stockholm University and Jönköping University
Industrial PhD Student in Computational Mathematics
Stockholm University and Jönköping University
I am a mathematician working at the intersection of non-parametric statistics and reliable machine learning, concurrently serving as a Data Scientist at Qamcom.
My overarching research perspective is pragmatic: it makes little sense to predict anything unless that prediction informs a decision. Therefore, my work focuses on generating predictions that are mathematically guaranteed to be reliable, even in dynamic environments, and integrating them into formal decision-theoretic frameworks.
My current research is divided into two main tracks, which I ultimately aim to unify:
1. Theoretical Conformal Prediction in Non-Exchangeable Settings I investigate the fundamental limits of conformal prediction when exchangeability is violated. A major focus of my published work involves characterizing $A$-cryptic deviations—distribution shifts that are fundamentally invisible to a fixed conformity measure $A$, leaving the resulting p-values deceptively uniform. For non-cryptic shifts, my ongoing theoretical work draws heavily on the defensive forecasting paradigm from game-theoretic probability. Inspired by Vovk et al.'s protected probabilistic regression and classification, I am developing frameworks to dynamically detect deviations from p-value uniformity online and recalibrate predictions, establishing rigorous theoretical guarantees for inference when standard exchangeability assumptions fail.
2. Predictive Decision-Making Systems (PDMS) Drawing on imprecise probability and classical decision theory, I develop systems that account for Knightian uncertainty. By utilizing Venn-Abers predictors, we obtain multiple predictive distributions rather than a single point estimate. Applying standard von Neumann-Morgenstern utility theory (or regret minimization) to these sets yields multiple expected utilities (or regrets). Applying subjective decision criteria to these bounds creates an uncertainty-aware PDMS.
The Ultimate Goal I aim to combine these tracks to create a Protected PDMS—an uncertainty-aware system that is robust to dynamic, non-exchangeable environments. Such a system mathematically "knows what it does not know," guaranteeing cautious behavior—such as abstention or defensive actions—when uncertainty is critically high.
Background in the Performing Arts Prior to my career in mathematics, I spent over a decade as a professional actor and musician. I view mathematical proof and musical performance as similar disciplines—both requiring rigorous preparation and structural clarity. I remain active in the music scene on the Swedish West Coast.
Recent Highlights
- Honor: Selected to attend the 13th Heidelberg Laureate Forum (HLF) as a Young Researcher (2026).
- Paper Accepted: "A Fast, Closed-Form Bandwidth Selector for the Beta Kernel Density Estimator" to appear in Journal of Computational and Graphical Statistics.
- Software Ecosystem: The Beta KDE "HS" bandwidth selector is now available across Python (
beta-kde), R (kdensityCRAN package), and Julia (BetaKDE). - Book Chapter: Co-authored a chapter in the Alexander Gammerman Festschrift, The Importance of Being Learnable (Springer, 2026).