Time series

What you'll learn:

• How to craft time-structured features from heterogeneous sources from with skrub: temporal resampling, temporal merges, calendar features, lag features and aggregate features over rolling windows.
• Methods to adapt state-of-the-art tabular data regressors and classifiers such as Gradient Boosted Decision Trees or any scikit-learn pipeline into time series forecasters, either manually or via third-party time-series oriented libraries.
• Some pitfalls of converting t+1 regressors to forecasters via recursion and what to do instead.
• How to train forecasters that yield sound predictions intervals rather than just point-estimates.
• Methodologically sound model selection and evaluation via time-structured cross-validation to account for time dependencies, both for point-wise average value predictions and probabilistic predictions. This includes the choice of informative evaluation metrics such as strictly proper scoring rules for point estimates, prediction interval boundaries and fine-grained distributional predictions and the coverage of prediction intervals.
• How to visually diagnose systematic over and under-estimation of extreme values via reliability diagrams and pros and cons of various strategies to mitigate those problems.

This hands-on workshop will use both real world production and environmental data to illustrate how to deal with common data-quality problems. We will also use some specially crafted synthetic data to emphasize catastrophic pitfalls related to edge cases to gain deeper insights about our modeling choices and their limitations. 

What to expect:

• Direct access to experts: Gain insights from Olivier and Guillaume, core developers of scikit-learn, and learn the latest features of the library.
• Hands-on learning: A small group format allows for deeper engagement and one-on-one interaction and real time feedback.
• Network with industry leaders: Wrap up with a cocktail hour overlooking the Eiffel Tower, connecting with key figures in tech.
• Personalized learning: This workshop is limited to 20 participants only, ensuring personalized, in-depth learning and high quality networking.