Advances in times series forecasting models

Abstract

This paper evaluates the performance of Artificial Neural Networks (ANNs), Nonlinear Autoregressive (NAR) networks, wavelet transforms, and hybrid models in forecasting, focusing on global gold demand from 2010 to 2023. Each method's ability to handle complex predictive tasks is assessed: ANNs demonstrate strong potential due to their deep learning capabilities but encounter challenges such as overfitting and high computational demands. NAR networks, utilizing LSTM and GRU units, effectively capture temporal dependencies but are sensitive to data quality. Proper wavelet selection is essential for successful wavelet transforms, as they enable detailed analysis of nonstationary data at multiple resolutions. To overcome the limitations of each individual model, hybrid models were explored to leverage their combined strengths while mitigating their weaknesses. The performance evaluation used metrics such as Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE). Results indicate that the hybrid model integrating ANNs and wavelet transforms outperformed the standalone models, with a 54.74% reduction in MAPE and a 13.13% decrease in RMSE, highlighting improved forecasting accuracy and reliability, particularly in the context of predicting global gold demand. This study emphasizes the importance of methodological innovation in forecasting, providing valuable insights into optimizing model integration to enhance accuracy. These findings are anticipated to benefit sectors like finance and trade, where accurate gold demand forecasts are vital for strategic decision-making and sustainable economic policy development, especially in regions such as East Asia.