The applied high-order spatial weight matrix gstarimax with principal component analysis for google trends data on gold price

This study aims to forecast gold prices with missing values in several G20 countries based on the best GSTARIMA model. Missing values have been handled by imputing last observation carried forward (LOCF) and moving average smoothing (MA). The results of the imputation then combined with the GSTARIMA model, principal component analysis (PCA), and dummy variables derived from Google Trends data to model and forecast gold prices in each country so 16 different models are considered in thesis study. In addition, Google Trends data also has been used to build a high-order spatial weight matrix on the GSTARIMA model. The eighth order is the highest order which is formed from the distance between countries. Based on the smallest MAPE value, the GSTARIMA model (1,0,0) with PCA and LOCF imputation is recommended for forecasting the next 1 month, the GSTARIMA model (1,0,0) with MA imputation for forecasting the next 1 year, and the GSTARIMA model (1,0,0) with PCA and MA imputation to forecast the next 2 years. The three best models explain that the gold prices in the future is relatively increasing.