Using Large Language Models to Analysing SMEs Cash Flow

<p dir="ltr">This study examines whether Large Language Models (LLMs) can enhance cash flow forecasting for Small and Medium-sized Enterprises (SMEs) by improving the classification of financial transactions. Using transaction-level banking data, we evaluate the predictive performance of three open-source LLMs Flan-T5-base, Llama-2-7B, and Gemma-2B across four classification schemes: unclassified, binary (income vs. expense), a three-category classification aligned with the cash flow statement (CFO, CFI, CFF). And an IFRS-based eight-category structure. Our results show that the three-category structure offers the best balance between interpretability and predictive accuracy, outperforming both the unclassified and binary approaches. While the eight-category classification provides richer information, it often introduces noise and reduces model stability, particularly for smaller models. In contrast, high-capacity models like Gemma-2B demonstrate improved performance with increased classification granularity. We also find that prediction accuracy remains stable across SMEs of different sizes, suggesting the method’s broad applicability. These findings have practical implications for financial institutions, particularly in SME lending, where audited financial statements are often unavailable. By structuring transaction data around core cash flow categories, LLMs can generate timely and reliable insights into firm-level liquidity risk. This study demonstrates that even in the absence of formal financial statements, transaction data alone can support robust predictive modelling and enhance real-time financial visibility.</p>