Coupling LSTM Neural Networks and State-Space Models through Analytically Tractable Inference

This repo contains the matlab codes to reproduce the results for the paper:

Vuong, Nguyen & Goulet (2024), Coupling LSTM Neural Networks and State-Space Models through Analytically Tractable Inference, International Journal of Forecasting.

(1) To load the saved predictions and calculate the test metrics: run scripts in the /metrics folder, e.g. metrics_electricity.m

(2) To run the code and obtain the predictions for each dataset: run scripts in the /config folder , e.g. electricity_2014_03_31.m

(3) To run examples using TAGI-LSTM and the TAGI-LSTM/SSM hybrid model: runs scripts in the /examples folder

• The synthetic_LSTM_smoothing.m file is to perform smoothing in TAGI-LSTM. In this example, smoothing is used to infer the past observations before the training time.
• The synthetic_coupling_normal.m file is to decompose a time series with linear trend using the TAGI-LSTM/SSM hybrid model.
• The synthetic_coupling_exponential_smoothing.m file is to decompose time series with a complex non-linear trend using the TAGI-LSTM/SSM hybrid model.

The Python implementation of the TAGI-LSTM method can be found in the pyTAGI library at https://www.tagiml.com/