Application of empirical mode decomposition, DeepAR, and N-Beats for gold price analysis and forecasting

Abstract

Gold has long served to secure funds in times of financial instability, adding value to currencies and protecting investments from inflation and economic downturns. Its price reflects broader economic conditions, making it extremely important to investors and necessitating the development of an effective method to forecast the price of gold. Financial time series typically exhibit non-linear relationships and are also characterized by significant volatility, requiring more sophisticated models to understand these dynamics and inform decision-making and strategic financial planning in an uncertain economic environment. The thesis evaluates the performance of innovative time series forecasting methods such as DeepAR and N-Beats. Comparisons are made with Linear Regression, ARIMA, VAR, and LSTM models to evaluate their effectiveness. The thesis also examines the effectiveness of using Empirical Mode Decomposition (EMD) in combination with various models to improve forecasting accuracy. Gold price data from 01.07.2012 to 31.03.2024 is used to compare the models. In addition to the gold price, the dataset includes 11 other explanatory variables to improve forecasting accuracy. The main results show that before applying EMD, the most efficient model is LSTM. After applying EMD, the most efficient model is ARIMA. The DeepAR and N-Beats models are comparable to traditional models in terms of forecasting accuracy due to their ability to handle complex non-linear time series such as historical gold prices. The application of Empirical Mode Decomposition significantly improved the accuracy of ARIMA and Linear Regression models, demonstrating their potential for accurate forecasting and the effectiveness of the EMD method for analyzing complex time series.