International Journal of Science, Engineering and TechnologyAuthors: Dr. Pankaj Malik, Vaidika Kaul, Gautam Jagthap, Harshit Jamley, Hitansh Chopra
Abstract: Traditional credit scoring models depend on static, historical data and fail to adapt to the rapidly changing financial behavior of borrowers. This paper introduces a Dynamic Credit Scoring Framework that leverages Deep Reinforcement Learning (DRL) to update borrower creditworthiness in real time using transactional and behavioral data streams. The credit assessment process is formulated as a Markov Decision Process (MDP), where a DRL agent continuously learns optimal credit decisions—such as loan approval, limit adjustment, or monitoring actions—based on evolving borrower states. The model employs a temporal feature encoder for real-time transaction analysis, coupled with an actor–critic architecture for decision optimization and a reward function that balances profitability, default risk, and fairness constraints. Experimental evaluation was conducted using three years of anonymized banking transaction data from 12,000 customers. Results show that the proposed DRL-based system improves default prediction accuracy by 18.7%, enhances long-term portfolio profitability by 23.4%, and reduces false approval rates by 21.6% compared to traditional gradient-boosted and logistic regression models. Furthermore, the model demonstrates strong adaptability under concept drift, maintaining performance stability with only minor retraining. These findings indicate that integrating DRL with real-time behavioral analytics can significantly enhance credit risk assessment, enabling financial institutions to make faster, fairer, and more dynamic lending decisions.
