Accounting Graph Transformer Improves Multi-KPI Forecasting for Small Businesses

A novel machine learning model known as the Accounting Graph Transformer (AGT) has been introduced to enhance financial forecasting for small enterprises with scant accounting records. The study, published on arXiv (2608.07037), tackles the issue of predicting 12-month results for 13 essential performance indicators (KPIs) from income statements, balance sheets, cash flow, and working capital, utilizing 71 monthly ledger series. Given that small businesses usually possess only 12-24 months of accounting data, making synchronized forecasts is challenging. By representing each ledger series as a masked token and employing typed attention on a fixed accounting-relation graph, AGT effectively exchanges information and integrates context with a gated three-month recency path. In evaluations involving 11,993 forecast origins from 1,060 unseen companies, AGT achieved a sample-weighted KPI-macro mean absolute error (MAE) of 0.6990 ± 0.0013, surpassing the top baseline, LightGBM, which recorded 0.7378 ± 0.0014. A paired company-clustered bootstrap at seed 42 revealed a significant difference favoring AGT, confirming its statistical advantage. This model's design capitalizes on the intrinsic connections among accounting items, potentially revolutionizing small business planning and risk management by delivering dependable financial forecasts despite limited historical data.