Application of empirical mode decomposition to improve deep learning for US GDP data forecasting

Area of Science:

• Econometrics
• Data Science
• Machine Learning

Background:

• Deep learning, particularly Long Short-Term Memory (LSTM) networks, shows promise for econometric forecasting.
• However, standard LSTM models exhibit instability, complexity, and limitations in handling econometric data, leading to prediction errors.
• Existing methods struggle with the nuances of predicting economic trends like US GDP growth.

Purpose of the Study:

• To propose an improved deep learning framework for US GDP trend and data prediction.
• To address the instability and complexity issues associated with standard LSTM models in econometric applications.
• To enhance the accuracy of econometric forecasting using a novel data decomposition technique.

Main Methods:

• Application of deep neural networks, specifically Long Short-Term Memory (LSTM) models.
• Introduction of Empirical Mode Decomposition (EMD) to preprocess time-series data.
• Decomposition of US GDP growth rate into Intrinsic Mode Functions (IMFs) using EMD before LSTM prediction.

Main Results:

• Standard LSTM prediction of US GDP growth rate yielded a Root Mean Squared Error (RMSE) of 2.7274.
• EMD-based LSTM prediction achieved a significantly lower RMSE of 0.93557.
• The EMD method effectively improved the predictive accuracy for US GDP growth rate.

Conclusions:

• Empirical Mode Decomposition (EMD) combined with Long Short-Term Memory (LSTM) offers a more stable and accurate approach to econometric forecasting.
• This hybrid method overcomes limitations of traditional LSTM models for complex economic data like US GDP.
• The findings suggest a promising direction for enhancing deep learning applications in economic trend prediction.