Feature Engineering for Time Series Forecasting
summary
Claims that machine learning (not classical ARIMA) wins on large cross-series problems: LightGBM wins M5 and yet gains are modest (top 50 beat exponential smoothing by ~15–20) while 92.5% of teams fail to beat exponential smoothing. Shows exactly how to turn series into tabular data (lags, window stats, static metadata), pitfalls that leak future data, and multi-step approaches (direct vs recursive). Demos use concrete Python tools — darts, sktime, tsfresh, scikit‑learn — with code for recursive forecasts and covariate handling.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:45About this talk
- 1:24Why use machine learning for forecasting?
- 3:44Don't neglect simple baselines though!
- 4:12Forecasting with machine learning
- 4:24Time series to a table of features and a target
- 8:21Multi-step forecasting: Direct forecasting
- 9:40Multi-step forecasting: Recursive forecasting
- 11:32Cross-validation: Tabular vs Time series
- 12:55Machine learning workflow
- 14:47Feature engineering for time series forecasting
- 16:23An example
- 16:51Target variable
- 17:22Lag features: Past values of target & features
- 19:35Window features: Function over a past window
- 20:56Window features: Nested window features
- 22:23Static features: Target encoding
- 26:20Key takeaways
- 26:59Overview of some useful libraries
- 28:04Forecasting with tabular data using Darts
- 34:04Conclusions
- 34:41References
related talks
39mForecasting With Classical and Machine Learning Methods
31mAdvanced machine learning time series....
01h 27mIntroduction to Machine Learning with Time Series
01h 26mApplying Statistical Modeling & Machine Learning to Perform Time-Series Forecasting
44mWinning with Simple, even Linear, Models
34mTime series for scikit-learn people
29mTime Series Forecasting Techniques in Python
24mMachine learning and financial time series analysis