Accounting Graph Transformer Boosts Small Business Financial Forecasts

Small businesses often operate with limited accounting history, typically 12-24 months, yet require coordinated forecasts across various financial statements for effective planning and risk management. Addressing this challenge, a new model called the Accounting Graph Transformer (AGT) has been developed. This model is specifically designed for joint 12-month forecasting of 13 key performance indicators (KPIs) spanning income statements, balance sheets, cash flows, and working capital, drawing from 71 monthly ledger series. The AGT represents each ledger series as a masked token and utilizes typed attention on a predefined accounting-relation graph to exchange information. It pools target-specific context and integrates it with a gated three-month recency path. In extensive evaluations across nearly 12,000 forecast origins from over 1,000 unseen companies, AGT significantly outperformed strong baselines like LightGBM, achieving a lower mean absolute error. Further ablations confirmed that relational attention, accounting topology, and the recency path each contribute to its improved accuracy. A single 5.3-million-parameter AGT model can produce 156 aligned forecasts without requiring company-specific fitting, offering a powerful tool for integrated financial analysis in small businesses.