Achieving Long-term Time Series Forecasting Models with Fewer Than 1k Parameters through Dynamic Sparse Training

Key Points

• Achieving long-term time series forecasting models with fewer than 1,000 parameters improves predictive performance.
• The analysis highlights the effectiveness of dynamic sparse training in enhancing model optimization and efficiency.
• A comprehensive review evaluates the impact of parameter reduction on forecast accuracy across various datasets.
• This work supports the need for efficient modeling in time series tasks, emphasizing model performance with reduced complexity.