Macroeconomic Multivariate Time Series Forecasting Model Based on Transformer Architecture

Macroeconomic systems exhibit high complexity, multi-variable interactions, and dynamic evolution. Traditional methods struggle to capture dynamic dependencies between variables, leading to insufficient forecasting accuracy. We propose a Transformer-based multivariate time series forecasting model. It employs a self-attention mechanism to model long- and short-term dependencies uniformly and a hierarchical encoder structure to enhance cross-variable relationship capture. A multi-task learning strategy jointly optimizes predictions for multiple economic indicators to improve robustness. Experiments show: 1) Superior multivariate forecasting performance (MSE: 0.024) vs. LSTM (0.033), GRU (0.037), and Informer (0.029); 2) The hierarchical attention mechanism reduces prediction error by 44.8%; 3) With 30% missing data, multi-task learning lowers error volatility by 20.7%. The model effectively addresses complex macroeconomic data modeling, demonstrating advantages in capturing multi-variable interactions, long-range dependencies, and noise resistance. It provides a reliable tool for policymaking and economic monitoring.