Multi-Task Learning for Macroeconomic Forecasting Based on Cross-Domain Data Fusion

Abstract

This paper addresses the challenges of multi-source heterogeneous data fusion and multi-indicator joint modeling in macroeconomic forecasting. It proposes a model framework based on multi-domain sample representation and joint prediction mechanisms. The goal is to improve both the prediction accuracy and structural modeling capability for key economic indicators such as CPI and GDP. The proposed method introduces a Domain-aware Representation Compression module. It encodes structured economic data and unstructured text data in a unified way. This enables efficient compression and alignment of multi-domain features. In parallel, a Joint Indicator Alignment mechanism is designed. Under a multi-task learning framework, it performs trend alignment and feature decoupling on the prediction outputs. This enhances the dynamic relationship modeling between different economic indicators. To validate the effectiveness of the proposed approach, a joint sample set is constructed. It integrates multi-domain information, including structural economic variables, news texts, and policy indicators. Multiple comparative experiments and ablation studies are conducted. The experimental results show that the proposed method outperforms mainstream models across various macroeconomic forecasting tasks. It demonstrates clear advantages in accuracy, robustness, and generalization. In particular, it maintains stable performance in cross-country transfer and multi-step forecasting scenarios. These findings confirm the model's adaptability and effectiveness in complex economic systems.