Factor-Based Framework for Multivariate and Multi-step-ahead Forecasting of Large Scale Time Series.

Area of Science:

• Econometrics
• Machine Learning
• Data Science

Background:

• Current multivariate forecasting methods struggle with high-dimensional data, non-linear relationships, and long horizons characteristic of Big Data.
• Deep learning models excel but demand substantial data and computational resources, lacking interpretability.
• Existing Dynamic Factor Model (DFM) approaches offer a foundation but require enhancements for modern forecasting challenges.

Purpose of the Study:

• To extend the Dynamic Factor Model for Large-scale (DFML) framework to improve multivariate forecasting capabilities.
• To integrate and assess both linear and non-linear factor estimation techniques.
• To evaluate model-driven and data-driven factor forecasting strategies within the enhanced DFML.

Main Methods:

• Proposed an extension to the DFML framework, incorporating advanced factor estimation and forecasting techniques.
• Implemented and assessed linear and non-linear factor estimation methods.
• Evaluated model-driven and data-driven factor forecasting approaches.

Main Results:

• The enhanced DFML framework demonstrated competitive forecasting accuracy on large-scale real-world datasets (>100 variables, >1000 samples).
• The proposed technique achieved significant computational efficiency compared to existing methods.
• Integration of various methods within DFML yielded robust performance across multiple forecasting tasks.

Conclusions:

• The extended DFML framework offers a powerful and efficient solution for complex, large-scale multivariate forecasting.
• This hybrid approach balances the strengths of DFM with advanced estimation and forecasting techniques.
• The method provides a viable alternative to deep learning for big data forecasting, offering improved interpretability and resource efficiency.