Research on Dynamic Assessment of Accounts Receivable Default Risk Using Long Short-Term Memory Network (LSTM) Combined with Attention Mechanism

Abstract

Dynamic assessment of accounts-receivable default risk requires time-series modeling of nonlinear financial and transaction behaviors. This study develops a Long Short-Term Memory network combined with an attention mechanism for early default-risk prediction. Financial and transaction data from 2844 manufacturing supply-chain enterprises over 72 consecutive months are preprocessed using seasonal decomposition, normalization, leakage-controlled feature construction, and imbalance correction. A three-layer LSTM extracts temporal dependence from 42 variables within a 12-month observation window, while the attention mechanism adaptively weights key time steps for interpretable risk classification. Experimental comparison with logistic regression and traditional LSTM shows that the LSTM–attention model achieves an accuracy of 0.892, recall of 0.874, F1 score of 0.878, and AUC of 0.921. Feature-weight analysis identifies aging structure, accounts-receivable turnover, leverage, and cash-flow ratio as key risk indicators. The model provides an engineering-oriented framework for financial time-series signal modeling, intelligent risk warning, and decision support.