Implementing Adaptive Machine Learning Frameworks for Financial Fraud Prevention

Abstract

Adaptive machine-learning frameworks for fraud prevention combine several design choices — the learning algorithm, the treatment of class imbalance, the training window, the alert threshold and the policy for updating the model — yet these are usually evaluated together, so implementers cannot tell which choices matter. This study implements the temporal-drift framework of Zuberi et al. [1] as the monitoring and retraining core of an adaptive fraud system and evaluates each design choice in a one-factor-at-a-time ablation. A simulated 40-week mobile-money stream of 320,000 transactions contained a mule-network fraud pattern emerging in week 15, a gradual shift in legitimate transfer behaviour and a two-week label delay. Fourteen configurations were compared over five replications using prequential evaluation. Adaptation was the decisive factor: a static model reached an average precision of 0.629, against 0.846 for the drift-triggered reference (gradient boosting, undersampling, four-week window, F2-tuned threshold). Among the remaining choices, the treatment of class imbalance and the length of the training window mattered most: SMOTE raised average precision to 0.876 and an eight-week window to 0.867, both improving on the reference in all five replications, mainly by stabilising models trained on few fraud cases. Combining them gave 0.878 and the highest F2 (0.817) with 3.6 retraining events, compared with 0.859 for weekly retraining with 35 events and ten times the training time. Learner choice had smaller effects, incremental stochastic-gradient learning was cheap but weaker (0.777), and re-tuning the threshold left average precision unchanged while modestly stabilising alert volumes. The results yield an implementation checklist: monitor and retrain on evidence of drift, give the learner enough fraud examples through resampling and window length, and gate every retrained model before deployment.