Forecasting With Classical and Machine Learning Methods
summary
Claim: machine learning can beat classical forecasting but only with careful preprocessing and feature design. The demo uses SKTime to run auto ARIMA, differencing, window features and a make_reduction to wrap tree models like LightGBM. On M5 (42,000 series, ~5,000 teams) only ~7.5% beat an ES-BASED benchmark and the top entry nets ~22.4% improvement; out-of-the-box ML often fails. The unified toolbox includes ~70 estimators, optional backends and conformal prediction, and the example pipelines materially improve accuracy.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:06Time series forecasting
- 3:10Classical vs ML methods
- 8:44Classical models outperform ML
- 11:52ML improves forecasting
- 15:27M5 competition insights
- 18:56SK time toolbox
- 20:19Simple SK time forecasting
- 23:07ML reduction for forecasting
- 26:35Preprocessing in SK time
- 28:22Global forecasting with SK time
- 31:55PyTorch in SK time
- 34:50Choosing forecasting model
- 35:35Q&A
related talks
43mFeature Engineering for Time Series Forecasting
29mMachine Learning: Better than Traditional Techniques?
31mAdvanced machine learning time series....
37mForecasting using machine learning methods
31mML time series forecasting methods
25mTimeSeries Forecasting with ML Algorithms and there comparisons
22mPredict the future (or don't) with machine learning in applications
26mPredicting the future (of equipment) using ML.NET