Conformal Prediction for Time Series: Uncertainty Quantification for Trustworthy Systems
Speaker
Vincenzo Ventriglia
A results-driven data professional, focused on hype-free solutions tailored to business needs.
I currently create value at the National Institute of Geophysics and Volcanology, where I develop machine learning models in the Space Weather domain. My work is complemented by finding the hidden stories in data and make them accessible to stakeholders. I studied Physics in Italy (Napoli) and Germany (Frankfurt am Main), previously worked in Analytics within the strategic division of the world's largest professional services network, as well as in the Data Science department of Italy’s leading publishing group.
I am also an organiser of PyData Roma Capitale, actively involved in building the local Python and data science community. Outside of work, I enjoy theatre, discussing finance, and learning new languages.
Abstract
How can we quantify uncertainty in time series forecasts, without unrealistic assumptions, and with rock-solid guarantees?
This talk introduces Conformal Prediction (CP), a framework to generate prediction intervals with guaranteed coverage. Whether you're forecasting energy demand, markets volatility, or space weather disturbances, CP helps you move from point forecasts to reliable intervals — even in non-stationary settings.
Description
Uncertainty quantification is a key component of any forecasting system deployed in the real world. Yet most models provide only point estimates, which can be dangerously misleading, especially in non-stationary or high-risk domains.
Conformal Prediction (CP) offers a statistically sound way to build prediction intervals with valid coverage guarantees, without assuming a specific data distribution or retraining the model. However, applying CP to time series requires care: sequential dependencies violate the exchangeability assumption, and naïve implementations can lead to invalid prediction intervals.
In this talk, we will:
- Start from the core idea behind CP
- Explain how it can be adapted to non-exchangeable data
- Compare approaches like Adaptive Conformal Inference (ACI), Ensemble Batch Prediction Intervals (EnbPI), and even Conformalized Quantile Regression (CQR)
- Show implementations for real-word time series problems
This session is aimed at data scientists and ML engineers working with temporal data and looking for better ways to express model confidence. A working knowledge of time series forecasting (with statistical or ML models) is beneficial, but not necessary.
Talk outline
- 0-5: Why point forecasts aren't enough — introduction to uncertainty quantification
- 5-10: A primer on CP
- 10-18: Beyond exchangeability — ACI, EnbPI, CQR: when and how to use
- 18-25: Applications from energy, finance, and space industries
- 25-30: Q&A