Author(s)

Qingquan Huang

Corresponding Author

Qingquan Huang

Abstract

To address core bottlenecks in traditional government budget performance evaluation—static lag, data silos, and disconnection between results and application—this paper proposes a dynamic evaluation model based on DeepSeek multimodal analysis. The model integrates structured and unstructured financial data (e.g., government accounting, policy texts, and project images) through a finance-specific multimodal fusion engine. Using time series analysis (LSTM) and reinforcement learning, it enables real-time performance tracking and deviation alerts throughout the budget cycle. Empirical analysis of three typical financial projects in Dongguan, China (livelihood, infrastructure, informatization), along with five comparative and ablation experiments, demonstrates that the model reduces the evaluation response cycle from an average of 186 days to within 72 hours, achieves 94.7% accuracy in expenditure deviation warnings (>15%), and improves the F1-Score to 93.4%. The study offers an engineering-ready intelligent solution to chronic issues like “funds waiting for projects” and “last-minute spending,” supporting the shift from compliance control to efficiency governance in public finance.

Keywords

Budget Performance, Multimodal Analysis, Dynamic Evaluation, Time Series Prediction, Intelligent Government Affairs