A Multi-Task GRU-Attention Model for Predicting Enterprise Investment and Financing Behavior from Multi-Source Economic Data

Abstract

Accurately predicting corporate investment and financing behavior is crucial for improving financial intelligence and capital allocation efficiency. This article proposes an economic data-driven multi-task deep prediction model that integrates Gated Recurrent Unit (GRU) networks with a multi-head attention mechanism to process multi-source heterogeneous economic variables, including macroeconomic indicators, corporate financial data, and market sentiment factors, under a unified structure. The model constructs multivariate time-series samples through sliding windows and employs a dual-output architecture to perform regression prediction of financing intensity and classification recognition of behavioral states into three classes (expansion, wait-and-see, contraction). To enhance responsiveness to behavioral transition patterns, a feature cross-attention mechanism and a joint loss function optimization strategy are introduced, improving nonlinear behavior learning capability and generalization robustness. Based on empirical data from 232 A-share listed companies, covering 12,840 training samples over the past decade, the experimental results showed that the model achieved a coefficient of determination (R²) of 0.862 in the financing prediction subtask, an accuracy of 88.3% in the classification task, and a Macro-F1 value of 0.841. Compared with baseline machine learning methods including Support Vector Regression (SVR), Random Forest (RF), and Multi-Layer Perceptron (MLP), the model demonstrated superior error control and trend fitting ability. Overall, the model exhibits high prediction accuracy, stability, and industry adaptability, providing a feasible technical path and empirical basis for building a data-driven intelligent investment and financing analysis system for enterprises.