Deep Learning for Time Series Forecasting: Key Design Choices Matter More Than Architecture
Analysis
This paper critiques the current state of deep learning for time series forecasting, highlighting the importance of fundamental design principles (locality, globality) and implementation details over complex architectures. It argues that current benchmarking practices are flawed and proposes a model card to better characterize forecasting architectures based on key design choices. The core argument is that simpler, well-designed models can often outperform more complex ones when these principles are correctly applied.
Key Takeaways
- •Fundamental design principles (locality, globality) are crucial for time series forecasting.
- •Implementation details significantly impact model performance.
- •Simple, well-designed architectures can outperform complex ones.
- •Current benchmarking practices need improvement.
- •A model card is proposed to characterize forecasting architectures based on key design choices.
“Accounting for concepts such as locality and globality can be more relevant for achieving accurate results than adopting specific sequence modeling layers and that simple, well-designed forecasting architectures can often match the state of the art.”