Leveraging Partner Country Factors in Deep Learning for Thailand’s Forecasted Inflation Accuracy Enhancement

Abstract

This paper focuses on improving the accuracy of headline inflation forecasts in Thailand. By evaluating the performance of deep learning models, time series forecasting models, and hybrid models in 1-, 3-, 6-, and 12-month advance forecast periods are investigated. In addition, the efficacy of including partner countries' inflation variables in the model is evaluated. There is a comparative analysis of various models, including ANN, RNN, LSTM, VAR, the hybrid model (VAR-ANN), and the BOTMM benchmark model of the Bank of Thailand. This study aimed to identify the most efficient model and demonstrate the impact of including partner countries' inflation on forecast accuracy. The results reveal that the hybrid model (VAR-ANN) consistently outperforms other models over several forecast periods, showing its superiority in capturing inflation trends. Specifically, the hybrid model (VAR-ANN) shows an average RMSE improvement of 50.36% over the BOTMM benchmark model from 2020 to 2022, with performance improvements of 52.94% in 2020, 56.56% in 2021, and 47.25% in 2022. In addition, the inclusion of partner countries' inflation significantly increases the accuracy of the predictions. These results are helpful for policymakers and practitioners working on inflation forecasts and emphasize the practical advantages of the hybrid model for enhancing prediction accuracy for Thailand's economic indicators.