Reproduce the experimental results using the provided scripts:
- Long-Term Forecasting:
bash scripts/reproduce_ltsf_results.sh
Configuration files: config/ltsf/.
- Short-Term Forecasting:
bash scripts/reproduce_stsf_results.sh
Configuration files: config/stsf/.
Insights
- Current supervised long-term point forecasting models (e.g., DLinear, PatchTST, iTransformer) struggle with intricate data distributions.
- Current supervised short-term probabilistic forecasting models (e.g., GRU NVP, TimeGrad, CSDI) face challenges in extended forecasting horizons.
Takeaways
- It is important to consider both long-term and short-term evaluation scenarios.
- Leverage both point and distributional metrics for more comprehensive insights.
Insights
- Current Supervised Non-Autoregressive (NAR) Models (e.g., PatchTST, iTransformer, CSDI)
- Primarily developed for long-term forecasting scenarios.
- Suboptimal for short-term forecasting, and some models are memory-intensive.
- Current Supervised Autoregressive (AR) Models (e.g., GRU, GRU NVP, TimeGrad)
- Primarily developed for short-term forecasting scenarios
- Perform well with strong seasonality but struggle with long-term, strong trends
Takeaways
- It is crucial to select the right methodological design based on the specific data characteristics.
- There are tremendous re-design opportunities, given the comprehensive forecasting needs.
Insights
- Reversible Instance Normalization (RevIN): Essential for Long-term Forecasting Scenarios
- Our observation: AR models in the literature are scarce for long-term forecasting
- Our finding: RevIN + AR => A simple yet highly effective baseline that has been overlooked
- Normalization Choices under Short-term Forecasting Scenarios
- No dominating normalization strategies
Takeaways
- The co-design of normalization techniques and model architectures warrants further research attention.
- The challenges and opportunities in time-series normalization persist in balancing short-term and long-term forecasting needs.
Long-Term Forecasting Benchmarking
Table 1. Results (
Short-Term Forecasting Benchmarking
Table 2.Results (