Time Series Forecasting: A New Approach

Time series forecasting presents significant challenges across various domains. A new study proposes a method to improve the accuracy of forecasts, based on the decomposition of the time series into trend and seasonal components.

Decomposition and Specific Models

The approach consists of applying distinct machine learning models to each component. Specifically, reversible instance normalization is effective only for the trend component. For the seasonal component, backbone models are used without normalization or scaling.

Results and Advantages

This strategy has reduced errors compared to existing state-of-the-art models. The results show an average reduction of 10% in Mean Squared Error (MSE) across several benchmark datasets. Furthermore, the approach was evaluated on a hydrological dataset extracted from the United States Geological Survey (USGS), achieving significant improvements while maintaining linear time complexity.

The introduced dual-MLP models prove to be computationally efficient solutions. The source code is available on GitHub.