PyData Amsterdam 2026

Forecasts are very brave little things: they happily tell you “3 minutes late” without telling you whether that estimate is solid or about to collapse into chaos. Standard forecasting models usually give point predictions, not prediction intervals with guarantees. That is where conformal prediction helps. This talk is about turning train delay forecasts into uncertainty aware predictions with coverage guarantees, using conformal prediction under realistic time-series drift.

However, time series are rude. Once drift, disruptions, and changing conditions appear, plain conformal prediction starts to struggle because yesterday’s calibration may no longer match today’s reality. The question easily now turns from “what is conformal prediction?” to which conformal method still works when the world shifts?"

Inspired by my daily nightmares, using open source Dutch train delay data, the core of the talk is a practical comparison of four conformal strategies for time series forecasting, presented as a deployment progression (Fixed split conformal prediction, Adaptive conformal inference (ACI), EnbPI). Those methods are approaching common problems in forecasting like drift, in different ways and that's why they were chosen. All of these approaches are implemented in well known libraries like MAPIE and are available to the user to support open source.

The outline:

• Commuter tragedy ( personal train commuting struggles)
• Why point forecasts are not enough
• How conformal prediction helps (add calibrated uncertainty on top of any forecaster)
• Wait a second, plot twist: time series ruin everything: plain conformal breaks in time series (drift, dependence)
• The method ladder:
  - Fixed Split CP: simple baseline (the naïve but lovable baseline)
  - ACI: adapt online to recent errors (the anxious adapter constantly updating)
  - EnbPI: handle sequential dependence with ensembles (the ensemble chaos manager)
• What to use when

Note that the above is an outline of the talk and not a per-slide equivalent. The main dissection, in terms of time, is that the first part will introduce the general problem (and my personal story) of point forecasts and the vanilla solution in 7 minutes. In the next section, the different approaches will be introduced in about 8 minutes. With the 10 minutes left, the direction will be on the practical answer of comparing strengths, weaknesses, and when to use which of the methods shown. The rest will be left for Q&A. The idea is to make this talk engaging with the audience.

This talk is unique among all prior PyData conformal prediction sessios. Conformal Prediction PyData talks have tended to emphasize on a broad overview of conformal prediction for time-series forecasting.

Unlike the time series focus at PyData Seattle 2023, the large‑scale forecasting angle at PyData London 2024, the energy‑grid case study at PyData Eindhoven 2023, the MAPIE library deep‑dive at PyData Global 2024, the gentle intro in Amsterdam 2024, the sktime/skpro probabilistic workshop in Amsterdam 2023, the regression only focus in London 2019 PyData and last year's PyData Amsterdam 2025 conformal prediction talk focused on either classification or general approach in time series, this talk shows the difference methods in Conformal Prediction and helps with "when to use which strategy in production". This is very crucial cause vanilla conformal prediction does break in production as mentioned above.

This talk is built are around only open source materials like libraries (Nixtla, Pandas), datasets (Rijdendetreinen). The code will be shared with the audience as well.