Detecting money laundering transactions with machine learning

Abstract

Purpose The purpose of this paper is to develop, describe and validate a machine learning model for prioritising which financial transactions should be manually investigated for potential money laundering. The model is applied to a large data set from Norway’s largest bank, DNB. Design/methodology/approach A supervised machine learning model is trained by using three types of historic data: “normal” legal transactions; those flagged as suspicious by the bank’s internal alert system; and potential money laundering cases reported to the authorities. The model is trained to predict the probability that a new transaction should be reported, using information such as background information about the sender/receiver, their earlier behaviour and their transaction history. Findings The paper demonstrates that the common approach of not using non-reported alerts (i.e. transactions that are investigated but not reported) in the training of the model can lead to sub-optimal results. The same applies to the use of normal (un-investigated) transactions. Our developed method outperforms the bank’s current approach in terms of a fair measure of performance. Originality/value This research study is one of very few published anti-money laundering (AML) models for suspicious transactions that have been applied to a realistically sized data set. The paper also presents a new performance measure specifically tailored to compare the proposed method to the bank’s existing AML system.

Application of Gradient Boosting Algorithms for Anti-money Laundering in Cryptocurrencies

“…Research published [19,22,39,49,57] in subsequent years reported that tree-based classifiers can effectively detect money laundering activities, e.g., Savage et al [39] reported that RF in conjunction with network analysis and community detection can effectively classify whether transactions were associated with money laundering. Jullum et al [22] employed XGBoost when detecting money laundering at a transaction level, due to its: (i) efficiency, (ii) scalability, and (iii) ability to reduce training time by utilising the GPU, and they noted that it outperformed a rule-based system with manual inspection (a common approach used in banks). Both RF and XGBoost are Decision Tree (DT) ensembles; however, the latter is a boosting algorithm, while RF is a bagging approach.…”

“…These models were implemented using xgboost, 1 lightgbm, 2 and catboost. 3 Hyperparameter optimisation was applied, as gradient boosting algorithms have a significant amount of parameters, and so, tweaking these parameters could improve results [15,22,37,52]. We opted to make use of Bayesian hyperparameter optimisation, more specifically, Tree Structured Parzen Estimator (TPE) [5], since it showed its effectiveness in highdimensional search spaces [37,52], when compared to other approaches [52].…”

“…In this offline setting, the following splits were used: (i) Elliptic, timestep ≤ 34 was used for training, with the following timesteps for testing; (ii) Ethereum Illicit Accounts, first 3276 instances for training with the remaining for testing (stratified sampling on class); (iii) NOAA, timestep ≤ 424 was used for training, with the following timesteps for testing. Data with temporal information were split based on timestep, so to not break the sequence of time [22,49]. -Experiment 2 (offline setting with data-sampling): We hypothesise that DT based gradient boosting, together with data-sampling, can further improve on the classification of licit/illicit at a transaction level.…”

“…A rule-based system typically consists of a set of conditions that check whether certain events occur or thresholds are exceeded, to signal suspicious activity. However, this technique has its drawbacks, a major one being that it leads to a high false-positive rate and require domain experts to create rules [22]. Machine learning overcomes some of these drawbacks by inferring patterns from previous data, which can decrease the falsepositive rate while keeping the false-negative rate relatively low.…”

Review Article on Machine Learning and its applications

“…Another intelligence, the infrared image based analysis, can also be implemented and used to predict the harvest time. Same as in machine learning in the money laundering [8].…”