Skip to content
Publications

Talk · IEEE International Conference on Smart Information Systems and Technologies · May 13, 2026

Calibration-Aware Gradient Boosting With Temporal Behavioral Features for Anti-Money Laundering Detection

T.K. Azbergen, Arailym Tleubayeva, Askar Aituov

Abstract

Anti-money-laundering (AML) systems increasingly rely on machine-learning models to prioritise suspicious transactions under severe class imbalance and limited investigative resources. Although gradient boosting methods provide strong ranking performance, their probabilistic outputs are often poorly calibrated, limiting their value for risk-based decision-making. This paper presents an empirical evaluation of calibration-aware gradient boosting models augmented with temporal behavioural features for transaction-level AML detection. CatBoost and XGBoost are combined with post-hoc probability calibration techniques, including Platt scaling and isotonic regression, while behavioural features are computed over rolling 1-, 7-, and 30-day windows to capture short- and long-term transaction dynamics. Experiments are conducted on the publicly available IBM Anti-Money Laundering dataset comprising approximately seven million synthetic transactions. Models are assessed using ranking metrics (AUPRC, AUROC), calibration metrics (Brier score and reliability curves), and operational metrics (Precision@Top k and Lift@Top k). Results show that calibrated CatBoost achieves the best overall performance (AUPRC =0.367, Lift @1%=40.9) while substantially improving probability reliability without degrading ranking quality. The study provides a reproducible benchmark demonstrating the importance of combining temporal behavioural modelling and probability calibration for operational AML risk scoring.

Imbalanced Data Classification TechniquesAnomaly Detection Techniques and ApplicationsFinancial Distress and Bankruptcy Predictiongradient boostingprobability calibrationCatBoost

More by Askar Aituov