GEVReg-MIDAS-SSVS for Mixed-Frequency Time Series Interval Forecasting

This study addresses the challenge of frequency mismatches between covariates and responses, known as mixed-frequency data, by using low-frequency variable information to predict high-frequency GEV-distributed variables with the GEVReg-MIDAS model. Furthermore, using an unlimited number of covariates increases computational load and model complexity. To address this issue, a covariate selection process based on Bayesian principles is integrated into the GEVReg-MIDAS model to form the GEVReg-MIDAS-SSVS. This proposed integrated model is tested on two classification datasets of clothing import prices, designed to produce price ranges (interval forecasts) as an early warning system for tax avoidance. The results show that the GEVReg-MIDAS-SSVS model outperforms the GEVReg-MIDAS model without feature selection, as evaluated by the AE and QRMSE metrics. GEVReg-MIDAS-SSVS also demonstrates the flexibility of the information contained in the model, where each modeling mismatch is more informative, representing the influence of its covariates. The import price range forecasting results indicate a price range that aligns with the training and testing data intervals. The results of this study can serve as a basis for policymaking to evaluate the prices of imported goods.