[ORCID]
[Google Scholar]
[GitHub]
[Maastricht University]
[LinkedIn]
[CV]
filip@schlembach.be
I am a PhD candidate at Maastricht University’s Department of Advanced Computing Sciences (DACS), supervised by Mark Winands and Evgueni Smirnov. My research focuses on predictive uncertainty quantification for multivariate regression and time-series forecasts. Specifically, I work within the framework of conformal prediction applied to multivariate non-exchangeable data.
Before joining DACS as a PhD candidate, I graduated with an MSc in Data Science for Decision Making at the same department in 2022. My thesis on Conformal multistep-ahead multivariate time-series forecasting won the department’s Master Thesis Award. Prior to that, I obtained a BSc in Scientific Programming from the FH Aachen in 2020, focusing on data science and machine learning.
I started my PhD as an external candidate while working full time as a Data Science and AI consultant at adesso, building machine learning pipelines in close collaboration with clients’ domain experts. I pursued both roles in parallel for over a year and a half before I moved to my current full time research position at DACS.
Publications are grouped by the role I played in their creation. Within each group publications are listed in reverse chronological order.
In multivariate regression, one can build hyperrectangular conformal prediction regions from one conformal predictor
per output dimension.
The region’s global significance level can be interpreted as a coverage error budget, and the significance
level of every dimension’s predictor spends a share of it.
Boole’s inequality guarantees joint validity under exchangeability as long as the sum of each dimension’s
share does not exceed the budget.
Bonferroni correction satisfies this condition by simply splitting the budget equally among the output dimensions
(Vovk, 2013).
The resulting prediction regions are however unnecessarily conservative [1].
My thesis addresses two distinct causes of this conservative behaviour.
The first cause is how the coverage error budget is allocated among the dimensions.
An equal allocation, as prescribed by the Bonferroni correction, ignores the fact that the dimensions’ errors
may differ in scale, and shifting budget towards the dimensions with the larger errors yields a smaller region [2].
The second cause is that Boole’s inequality is not tight in general.
If dimensions miscover together rather than disjointly, the sum of the dimension-wise significance levels may exceed
the upper limit set by Boole’s inequality.
Max-Rank, introduced by Timans et al. (2025),
exploits this, and [3] corrects two issues in its implementation.
This work received the Alexey Chervonenkis Award for Best Student Paper at the Fifteenth Symposium
on Conformal and Probabilistic Prediction with Applications.
The respective solutions to the causes of the conservative behaviour presented in [2] and [3] assume
exchangeability, which may be violated in the time-series forecasting setting that motivated this work.
[4] extends them to non-exchangeable data using weighted conformal prediction
(Gibbs and Candès, 2021;
Barber et al., 2023).
Max-Rank may raise the sum of the dimension-wise significance levels, i.e. the total coverage error
budget, above the global significance level, but splits it equally among the output dimensions.
Therefore, it does not address the first cause of conservative prediction regions.
I am currently merging the two lines of work into a single method that optimises the allocation of the coverage
error budget and raises the total.
Bachelor’s and master’s thesis projects I co-supervised, published with the student as first author.