PRIOR-INFORMED MULTIVARIATE LSTM (PIM-LSTM) FOR ECONOMIC TIME SERIES

Date

2025-3-6

Author

Aydemir Aydın, Petek

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

48
views

13
downloads

Cite This

Deep Learning is a subset of machine learning that emphasizes algorithms influenced by the human brain, called artificial neural networks. Physics-Informed Neural Networks (PINNs) represent a distinct deep learning method that combines the strengths of neural networks with the physical principles that dictate particular systems. The main goal of this thesis is to enhance the PINNs model for multivariate time series by integrating causal relationships and cross-correlations to improve overall model performance. For this purpose, we developed a Prior-Informed Multivariate Long Short-Term Memory (PIM-LSTM) model. First, its application to the New Keynesian and Dividend-Augmented Goodwin-Keen (DAGKM) models is demonstrated. Then, the forecast performance of the PIM-LSTM model is compared to the LSTM and PINN models. Our findings indicate that the PIM-LSTM model demonstrates strong predictive performance on the New Keynesian Model for Turkiye and Mexico’s macroeconomic series, achieving lower MAE, RMSE, and MASE compared to LSTM and PINNs models. The PIM-LSTM model also performs well in the DAGKM model. Integrating the New Keynesian model for Turkiye and Mexico enhances the analysis by capturing country-specific monetary policies and economic dynamics. Similarly, incorporating the DAGKM model enhances the analysis by capturing cyclical growth and income distribution.

Subject Keywords

Time series analysis, Forecasting Methods, Deep Learning Algorithms

URI

https://hdl.handle.net/11511/114105

Collections

Graduate School of Natural and Applied Sciences, Thesis