1D Convolutional Neural Networks for Time Series Modeling
summary
Claims that learned 1D convolutional filters plus a small context network beat the existing TBATS baseline by about MASE 0.85–0.86 vs TBATS 0.90 on revenue‑per‑click hourly data, trained on eight weeks and tested on the final week. Prototype in MATLAB is slow (≈20 minutes) but porting to PyTorch + fastai on GPU cuts training time dramatically; a richer filter‑net ensemble gives another accuracy bump. Model shares information across categories, resists short training windows, but degrades quickly with many missing points.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:15Motivation for 1D CNNs
- 3:23Signal processing and 1D conv
- 4:551D CNN architecture
- 7:43Hyperparameter tuning
- 8:48Data and evaluation metrics
- 11:34Prototype results
- 13:23GPU framework transition
- 14:27PyTorch implementation
- 15:52Comparison with TBATS
- 17:14WaveNet structure
- 18:17Feature engineering
- 19:22Model architecture
- 20:44Performance comparisons
- 22:06Model generalization
- 23:53Testing jagged data
- 25:16Cross-category context
- 26:41Training data impact
- 27:23Conclusion
- 28:04Q&A
related talks
25mLooking beyond LSTMs: Alternatives to Time Series Modelling using Neural Nets
31mAdvanced machine learning time series....
31mML time series forecasting methods
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
21mFoundational Models for Time Series Forecasting: are we there yet?
31mFoundational Time Series Models in Practice: The Future of Forecasting, or Just Hype?
32mA word is worth a thousand pictures: Convolutional methods for text
54mTransformes for Time Series: Is the New State of the Art (SOA) Approaching?