Deep learning for time series forecasting and classification in practice
summary
Walks through hands‑on examples using Flow Forecast, PyTorch Geometric Temporal, SHAP and ELI5 to forecast retail, healthcare, climate and sales series and to do anomaly detection. Shows Flow Forecast config-driven runs, a case where Informer fails vs DAR RNN out‑of‑the‑box, a probabilistic model producing wide confidence bands, and a 116-run hyperparameter sweep that changes outcomes dramatically. Reports small datasets (≈300 rows), common pitfalls (missing values, scaling, unstable seeds) and which builds actually work.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:50Tutorial overview and challenges
- 3:16History of DL time series models
- 4:41Forecasting and classification challenges
- 6:07Problem formulation and data
- 8:56Model landscape overview
- 10:22Pitfalls with SOTA models
- 12:47Training deep learning issues
- 14:55Metrics don't guarantee performance
- 17:27Hands-on with Flow Forecast
- 21:18Reproducibility in Flow Forecast
- 23:45Hyperparameter tuning and steps
- 26:34Training and logging with W&B
- 29:22Evaluating predictions and features
- 33:56Probabilistic forecasting
- 36:23Multi-region forecasting with Transformers
- 38:30Uncertainty with confidence intervals
- 41:20Classification and data length
- 44:28Hyperparameter sweep example
- 51:14Future directions in DL
- 55:48Q&A
related talks
40mDeep Learning and Time Series Forecasting for Smarter Energy
01h 26mApplying Statistical Modeling & Machine Learning to Perform Time-Series Forecasting
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
44mMethods and Tools for Time Series Data Science Problems
01h 29mFull-stack Machine Learning and Generative AI for Data Scientists
31mAdvanced machine learning time series....
43mFeature Engineering for Time Series Forecasting
34mTime series for scikit-learn people