SparseTSF: Lightweight and Robust Time Series Forecasting via Sparse Modeling.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Long-term time series forecasting (LTSF) presents challenges in modeling complex temporal dependencies with limited computational resources.
• Existing methods often require substantial parameters and computational power, hindering their application in resource-constrained environments.

Purpose of the Study:

• To introduce SparseTSF, an extremely lightweight and novel method for LTSF.
• To address the need for efficient and robust time series forecasting models with minimal computational overhead.
• To demonstrate competitive performance against state-of-the-art methods using significantly fewer parameters.

Main Methods:

• Developed SparseTSF, a novel forecasting method.
• Implemented the Cross-Period Sparse Forecasting technique, involving downsampling sequences for trend prediction.
• Focused on reducing model complexity and parameter count while enhancing robustness through implicit regularization.

Main Results:

• SparseTSF utilizes fewer than 1,000 parameters, achieving competitive performance in LTSF.
• Demonstrated significant advantages with longer look-back windows (e.g., 720), effectively exploiting periodicity and trend information.
• Showcased remarkable generalization capabilities, performing well with limited data, small samples, or low-quality data.

Conclusions:

• SparseTSF offers an optimal balance between performance and computational efficiency for LTSF.
• The method is highly suitable for scenarios with resource limitations, small datasets, or noisy data.
• Publicly available code facilitates adoption and further research in lightweight time series forecasting.