Transformer-Based Time Series Forecasting for Credit Limit Utilization and Risk Profiling in Retail Lending

This paper proposes a hybrid framework for credit risk assessment and forecasting of credit-limit utilization in a retail lending portfolio. In contrast to traditional binary credit-scoring models that estimate only the probability of default, the proposed approach formulates credit portfolio dynamics as a multivariate, credit policy-controlled time-dependent system, enabling trajectory-based forecasting of borrower behavior and risk. The framework employs a Transformer architecture to capture complex temporal dependencies. The input representation consists of time series describing the evolution of assigned credit limits, their utilization dynamics, and client-specific risk profiles, where logistic-regression-based risk estimates are embedded as structured components of the system state. Unlike conventional ensemble approaches, the proposed method integrates statistical risk estimates directly into a sequential modeling framework and performs multi-step autoregressive forecasting of both utilization and risk trajectories. Empirical evaluation on a real-world credit portfolio dataset demonstrates that the proposed approach outperforms classical and neural baselines in forecasting accuracy while maintaining stable out-of-sample performance. Overall, the results highlight the effectiveness of Transformer-based architectures for modeling dynamic credit systems and their potential as a decision-support tool for proactive credit portfolio management, as demonstrated under stylized economic assumptions.