Presenting at UAI 2026 in Amsterdam
I presented our paper, “Model-Agnostic Online Certificate-Driven Calibration for Time Series Forecasting Under Distribution Shift,” as an oral at UAI 2026 this week in Amsterdam. The conference was held at KIT (the Royal Tropical Institute). Our talk was on August 19 in Oral Session 4: Uncertainty and Calibration, in the Queen Máxima Hall.
After the acceptance in June, preparing the oral meant compressing the story: why i.i.d. PAC-Bayes is the wrong certificate for shifting time series, how a martingale bound gives a usable online certificate, and how a gated Bayesian head can calibrate a frozen backbone without throwing the source model away. The questions, both in the hall and later at the poster, kept coming back to that last point: when you should trust the certificate enough to adapt, and when you should fall back.
Giving the talk in that room was a different experience from ICDM last fall. An oral leaves little time, so every slide has to earn its place. I am grateful it landed in a session on uncertainty and calibration, where the audience already cared about making those statements mean something once you leave the i.i.d. setting.
The same afternoon I presented the poster in the Marble Hall (Marmeren Hal). Those conversations ran longer: people stayed with the bound, the predict-then-update protocol, and the backbone-agnostic design.
I am grateful to my advisor, Dr. George Michailidis, for his guidance throughout this work. The paper is in PMLR; code is on GitHub.