ML time series forecasting methods
summary
ES‑Hybrid pairs exponential smoothing with a dilated residual LSTM and wins M4; N‑Beats uses stacked blocks and ensembles 100+ models and claims better M4 scores; GluonTS provides probabilistic forecasting and Transformer support; Tsetlin machine learns per‑parameter stochastic distributions and updates after each prediction. Hands‑on uses an hourly air‑pollution set of 40,000 rows and 15 columns, forecasts NO2 with 6‑hour horizon using TFT and N‑Beats; TFT with quantile loss scores best and ensembling yields about 15% MAE improvement (authors report ~60% accuracy on 10‑year financial tests).
AI-written summary · about
chapters (22)
- 0:00ML time series intro
- 1:26From stats to ML methods
- 3:12M4 and M5 breakthroughs
- 4:55ES-Hybrid innovations
- 6:00N-BEATS advantages
- 7:51GluonTS overview
- 9:16Machine vs stochastic learning
- 12:45Advanced forecasting summary
- 14:31Hands-on session
- 15:14Data processing challenges
- 18:24Handling missing data
- 19:29Data transformations
- 20:31Baseline models
- 21:59Training setup and metrics
- 22:42Temporal Fusion Transformer
- 23:45N-BEATS performance
- 25:32Neural training techniques
- 25:53Model predictions
- 26:55Loss functions vs metrics
- 28:00Ensembling for accuracy
- 29:29Explainable AI
- 31:15Q&A
related talks
31mAdvanced machine learning time series....
39mForecasting With Classical and Machine Learning Methods
43mFeature Engineering for Time Series Forecasting
28mNixtla: Deep Learning for Forecasting...
21mFoundational Models for Time Series Forecasting: are we there yet?
54mTransformes for Time Series: Is the New State of the Art (SOA) Approaching?
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
31mFoundational Time Series Models in Practice: The Future of Forecasting, or Just Hype?