Dynamic quantification anti-fraud machine learning model for real-time transaction fraud detection in banking

As the scenario of telecom network fraud intensifies, the development of anti-fraud models has become a focal point in financial technology, including banking sectors. Traditional anti-telecom fraud models, which primarily rely on expert rules and machine learning methodologies, are beset with limitations such as poor real-time performance and high false favourable rates, leading to biased predictive outcomes. This paper introduces a novel anti-fraud model: The Dynamic Quantification Anti-Fraud Model grounded in Real-time Transaction Flows to address these issues. This model combines multiple F-Beta features linearly, employs a DML (Double Machine Learning) framework for feature weight quantification, and conducts instantaneous risk assessments of account transactions. By dynamically adjusting feature weights based on historical single-feature false favourable rates and the branch’s anti-fraud tolerance capacity, it significantly improves model performance and the efficiency of fraud prevention. Using transaction detail data of implicated cards from a case bank spanning January to June 2023 as an example for modelling, the validation sample recall rate reached 0.506. Post-deployment, the model facilitated a 30-percentage point reduction in the proportion of implicated cards within six months at the bank, alongside a substantial drop in its ranking among peers concerning the number of fraud cases. This comprehensive improvement solidly validates the precision and efficacy of the proposed model.