LLM-Based Intelligent Risk Identification for SMEs: A Comparative Study of Prompt Engineering Strategies

Abstract

Small and medium enterprises (SMEs) are the backbone of economic growth and employment worldwide, yet they face significant challenges in risk management due to limited resources and expertise. Aligned with the United Nations Sustainable Development Goals (SDGs), particularly SDG 8 (Decent Work and Economic Growth) and SDG 9 (Industry, Innovation and Infrastructure), this paper proposes a novel approach leveraging Large Language Models (LLMs) for intelligent risk identification from unstructured textual data. We design a comprehensive risk classification framework covering financial, operational, market, and compliance risks tailored for SMEs. Through systematic comparison of three prompt engineering strategies—zero-shot, few-shot, and chain-of-thought (CoT)—using the DeepSeek-V3 model on a dataset of 100 samples from authoritative public sources (UCI German Credit, Kaggle Supply Chain, and Kaggle SME Financial datasets), we demonstrate that the CoT strategy achieves the best performance with an F1-score of 0.649, representing a 31.96% improvement over traditional keyword-based methods. Notably, CoT is the only strategy capable of identifying compliance risks (F1=0.649 vs. 0 for other strategies), highlighting the importance of step-by-step reasoning for complex regulatory risk detection. Extended validation with six LLMs (DeepSeek-V3, DeepSeek-R1, Qwen3-235B, GLM-4, Kimi-K2, and Qwen3-Max) on 2,000 samples confirms the robustness of our findings. Our research provides practical guidance for deploying LLMs in SME risk management while contributing to sustainable economic development.