South African inflation modelling using bootstrapped long short-term memory methods

Area of Science:

• Economics
• Econometrics
• Data Science

Background:

• Economic stability relies on effective inflation targeting.
• The COVID-19 pandemic has created unprecedented economic conditions, necessitating updated policy guidance.
• Previous South African inflation research primarily utilized statistical models such as ARFIMA, GARCH, and GJR-GARCH.

Purpose of the Study:

• To explore the application of deep learning techniques for South African inflation forecasting.
• To compare the predictive performance of deep learning models against established statistical methods.
• To identify the most accurate forecasting model for informing economic policy.

Main Methods:

• Implementation of deep learning models, including clustered bootstrap Long Short-Term Memory (LSTM).
• Evaluation of model performance using metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Root Mean Squared Percentage Error (RSMPE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).
• Application of the Diebold-Mariano test to statistically compare forecast accuracy between models.

Main Results:

• Clustered bootstrap LSTM models demonstrated superior forecasting performance.
• Deep learning models significantly outperformed traditional ARFIMA-GARCH and ARFIMA-GJR-GARCH models in predicting South African inflation.
• The Diebold-Mariano test confirmed the statistical significance of LSTM model superiority.

Conclusions:

• Deep learning, particularly clustered bootstrap LSTM, represents a significant advancement in inflation forecasting accuracy for South Africa.
• These findings offer valuable insights for policymakers seeking to navigate current economic challenges.
• The study highlights the potential of advanced machine learning techniques in economic modeling and policy formulation.