Reliable Machine Learning Methods for Payment Risk Prediction in Supply Chain Financing

Abstract

Supply Chain Financing (SCF) has become an essential financial solution for enhancing liquidity and stabilising cash flows within increasingly complex global supply chains, especially for small and medium-sized enterprises (SMEs). However, a series of SCF failures and financial scandals have revealed fundamental weaknesses in current risk prediction methods. These approaches largely depend on firm-level financial indicators, are constrained by limited data availability, lack detailed operational insight, and provide little interpretability. Among the different types of SCF risks, payment delays constitute a particularly severe operational hazard, directly affecting suppliers’ liquidity and potentially setting off cascading disruptions throughout supply chain networks. Overcoming these challenges requires innovative analytical methods that are robust, privacy-preserving, uncertainty-aware, and explainable. This thesis investigates the application of advanced machine learning techniques can be used to enhance the prediction of payment delay risk in SCF, with a particular emphasis on order-level modelling and industrial feasibility. Three complementary methodological frameworks are developed to address distinct but interrelated research gaps. First, a Federated Learning (FL) framework is proposed to enable privacy-preserving, collaborative risk prediction among supply chain partners without sharing raw operational or financial data. This approach mitigates data insufficiency and imbalanced data, especially for SMEs, while improving predictive performance through shared model training. Second, a Hierarchical Bayesian Model (HBM) is introduced to deliver uncertainty-aware payment risk predictions and behaviour similarity sharing via partial pooling. By generating posterior distributions rather than point estimates, the HBM provides more robustness and interpretability in data-sparse supply chain environments. Third, a Causal Artificial Intelligence framework is introduced to move beyond purely correlation-based prediction by uncovering the causal factors of payment delays and converting these findings into actionable operational actions. Using causal effect estimation and interpretable policy learning, this framework supports targeted interventions, such as urgent deliveries or pricing strategy adjustments, customised to individual suppliers. The proposed methods are empirically evaluated using a real-world aerospace supply chain case study involving order-level operational data. Results demonstrate that the proposed frameworks significantly outperform conventional models, improve explainability in risk assessment, and support more effective decision-making for SMEs and financial intermediaries. The findings highlight the importance of considering privacy preservation, uncertainty modelling, and causal reasoning into SCF payment risk predictions. Overall, this thesis advances the state of the art in SCF by providing an applicable and generalisable foundation for explainable and resilient payment risk prediction for industrial application.