Tax-Invoice Relation Modeling for Financial Distress Prediction in Small and Medium Industrial Enterprises

Abstract

Small and medium industrial enterprises often lack complete public financial disclosures, making financial distress prediction difficult when relying only on balance-sheet indicators. Tax invoices, supplier-customer relations, sales concentration, purchase volatility, and payment behavior provide alternative signals for identifying early financial deterioration. This study proposes a tax-invoice relation modeling method for financial distress prediction in small and medium industrial enterprises. The method constructs an enterprise transaction graph using value-added tax invoices, supplier links, customer links, product-category flows, invoice cancellation records, tax-payment behavior, and regional industry labels. A graph neural network learns transaction-based risk representations, while knowledge rules infer distress signals from shrinking sales, abnormal invoice reversal, supplier loss, delayed tax payment, and customer concentration. The dataset contains 74,200 small and medium industrial enterprises, 3.12 million supplier-customer edges, 18.6 million invoice records, 960,000 invoice reversal events, and 13,700 confirmed financial distress cases over 40 months. The proposed model reduces median early-warning time from 83 days to 32 days compared with a tax-ratio scoring baseline. It identifies 5,860 enterprises with persistent sales contraction and 4,210 enterprises affected by major customer loss before formal distress confirmation. Knowledge-rule inference generates 9,740 interpretable risk chains, including supplier interruption, abnormal tax-payment delay, invoice reversal clustering, and downstream demand collapse. Full monthly assessment is completed in 18.3 minutes for all enterprises. The results indicate that tax-invoice relation modeling can strengthen industrial financial risk prediction when conventional financial statement data are incomplete or delayed