A unified ML LLM decision support framework for SME financial analytics forecasting risk and advisory recommendations

In the financial decision-making process of Small and Medium Enterprises (SMEs) and individuals, integration of forecasting, risk quantification, and advisory generation is a tripartite requirement that is not addressed by existing systems. We present MoneyMind AI, a modular, system integration pipeline-based Financial Decision Support System (FDSS) with tightly integrated components: (i) Data ingestion & preprocessing, (ii) Structured financial analytics engine computing per financial metric like income, expenses, Net Cash Flow (NCF), and Profit Margin (PM) from the transaction-level CSV data, (iii) Machine Learning (ML) forecasting module using Ordinary Least Squares (OLS) Linear Regression (LR) and Random Forest (RF) ensemble methods, which offers 1, 2, and 3-month ahead financial forecasts, and (iv) a composite Financial Health Scoring (FHS) algorithm that integrates the four primary metrics the PM, Cash Flow Stability (CFS), Revenue Growth (RG) rate, and Expense Volatility (EV) into a tri-class risk stratification schema, (v) A Large Language Model (LLM) driven recommendation engine that supports the generation of action-oriented, natural language-based financial advisory based on a prompt-engineered API call, converting the structured financial summary results. The RF forecasting model benefits the ARIMA(1,1,1) baseline model, with a mean MAPE of 11.2% ± 1.5% (95% CI; MAE ≈ ₹50,400; RMSE ≈ ₹67,100) achieved over a 24-month observations hybrid transactional dataset of 480 transaction records across six expense categories using four-fold cross-validation (walk-forward). The Diebold–Mariano test with HLN small-sample correction indicates that RF is significantly more accurate than the ARIMA(1,1,1) baseline at the 1% level (HLN-corrected DM = 3.94, p = 0.002, n = 12) and then LR at the 5% level (HLN-corrected DM = 2.24, p = 0.046). The FHS risk classifier achieves an overall accuracy of 84.0% and a macro-average F1-Score of 83.9% and a Cohen’s κ of 0.76, showing substantial inter-class agreement. The recommendations of LLM (50 outputs per category from 5 domain experts) result in a mean quality score of 4.19 ± 0.31/5.0 (Krippendorff’s α = 0.79). On a full four-module functional coverage, which it is alone among the four benchmark methods, MoneyMind AI cuts the time required to complete financial analysis tasks by 97.9–99.0% relative to manual financial analysis methods across a small battery of standard tasks. The results presented show the potential of MoneyMind AI as an accessible, replicable, and actionable solution to empower SMEs and individual financial decision-makers with financial intelligence.