LSTM-MSNet: Leveraging Forecasts on Sets of Related Time Series With Multiple Seasonal Patterns.

Area of Science:

• Machine Learning
• Time Series Analysis
• Forecasting

Background:

• Accurate forecasting for time series with multiple seasonal cycles is crucial across various industries.
• Existing univariate methods independently model each time series, missing shared patterns.
• There is a need for advanced models that can capture complex multiseasonal dynamics.

Purpose of the Study:

• To propose LSTM-MSNet, a unified, decomposition-based prediction framework for multiseasonal time series forecasting.
• To develop a globally trained LSTM network that exploits cross-series knowledge.
• To integrate state-of-the-art multiseasonal decomposition techniques into the forecasting framework.

Main Methods:

• Developed a globally trained Long Short-Term Memory (LSTM) network, LSTM-MSNet.
• Integrated advanced multiseasonal decomposition techniques to enhance LSTM learning.
• Evaluated the framework on diverse datasets, including the M4 forecasting competition data.

Main Results:

• Demonstrated that a decomposition step is beneficial for disparate data sources.
• Showed that for homogeneous series, exogenous seasonal variables or no preprocessing are superior.
• Achieved competitive results, outperforming many existing state-of-the-art multiseasonal forecasting methods.

Conclusions:

• LSTM-MSNet offers a flexible and effective framework for multiseasonal time series forecasting.
• The choice of preprocessing (decomposition, exogenous variables, or none) depends on the data characteristics.
• The model successfully exploits cross-series information for improved forecasting accuracy.