co-authored by Duc Thang Hoang
Anti-money laundering (AML) is a complex and regulated field involving composite data and intricate workflows. With tighter regulations and a prevailing reliance on manual processes, the heat is on for banks to get their risk management acts together. Machine learning can play a key role in transforming this sector.
Classical transaction monitoring tools perform based on static rules that were set up beforehand. For example, if a recorded transaction amount is higher compared to previous transaction amounts for a client, the monitoring tool will generate an alert. Up to 90% of these alerts are usually false positives. The AML specialists have to review all alerts without exception and mostly with limited structure and information on the alert.
Figure 1: Traditional alert handling within financial institutions
Using machine learning is possible to cluster the alerts generated by the transaction monitoring tool into different categories, for example, predicted true hits or false hits. This cluster allows the AML specialists to focus on hits that require faster review and decision-making to achieve regulators’ deadlines.
Figure 2: Possible alert handling using machine learning
In this new article, we will describe our approach to use machine earning within the AML transaction monitoring area.
How to leverage Machine Learning in AML Transaction Monitoring
Machine learning (ML) is one of the hottest topics in the last decade. One reason for this is that ML is used today to evaluate large amounts of data and because the costs for computing power and memory have fallen significantly. At the same time, computer performance has increased enormously. By storing and processing large amounts of data, companies have succeeded in applying ML and deep learning (DL) methods to data. Today, this takes place in almost every industry and should bring new insights. Examples from everyday life are the individually adapted search engine queries, suggestions to buy similar articles, targeted marketing, predictive maintenance, or clustering your AML alerts. According to our approach, we analyze and test each use case to understand if it is possible to solve the problem by applying machine learning. We will walk through each phase of the approach to explain the main steps to be performed.
Figure 3: Machine learning approach – model construction and training
Preprocessing: The first step in machine learning training is to understand the dataset. Analyzing the dataset means, to clear possible outliers or missing values, but also using feature engineering. Feature engineering is a process to extract or combine raw data into new values, to add further information. To clear the dataset there are many options, one of them is to use average values of a column for missing values. For a cleaner dataset, we normalize all input variables, which means that all values will convert between zero and one. Further possible tasks in this phase are to find correlations or transforming strings into the right data format. Pre-processing is very important and always takes the most time because better data beats fancier algorithms.
Modeling: The next step is to find the right and possible machine learning algorithms for the use case. First of all, we need a split on the dataset, one with the target variable to train (training set) and one without the target variable to test (test set). The most common split ratio is between 70-80% for the training set and 20-30% for the test set. To find out the best suiting model for the use case, it is necessary to use several algorithms and then compare the results.
After finding a suitable algorithm with good results, the next step is to further test and train the algorithm. The training set will be used to find the pattern and learn, the test set is for predicting the right outcome. In a binary classification problem, the measurements are shown in a confusion matrix.
Figure 4: Confusion matrix for a binary classification
The outcome of the analysis will be shown in a confusion matrix. The confusion matrix is a measurement for classification problems. With the outcome, we can calculate, for example, the accuracy (a percentage number) of all correctly classified outcome variables. The accuracy is given by all the wrong predicted outcomes (false negative and false positive), divided by all the right predicted outcomes, True positive and true negative.
Postprocessing: The most critical part to improve your algorithm is the right set of parameters. This method is called hyperparameter optimization. Two very common hyperparameter optimization options are grid search and random search. While ‘random search’ searches randomly with a set of parameters and selects random combinations to train the model, ‘grid search’ has a defined set of parameters and scores each combination on the testing data. In the end, both methods give you a result, which can be better than the default parameter setting. After finding the best possible combinations of parameters, a final training on the dataset should be conducted.
This is in a nutshell how to implement machine learning algorithms to cluster any kind of data.
Benefits of our approach
Machine learning algorithms can be easily deployed next to anti-money laundering transactions monitoring tools with a low impact on the IT infrastructure. There is no need to replace the current transaction monitoring system. In the AML area, it is possible to cluster all alerts coming from the transaction monitoring. The machine categorizes the alerts into different groups. This approach helps to focus and prioritize the efforts of the AML specialists and teams conducting the alert investigations in a structured way, allowing a firm’s highly specialized staff to focus on the tricky cases and accomplish within regulator deadlines. Last but not the least, ML augments the human intellect of an existing team. Applying machine learning algorithms augment the outcomes and reduce the effort and cost of an alert investigation.
Article authored by Gerardo Salonia and Duc Thang Hoang

Gerardo Salonia
Senior Principal
Gerardo Salonia is a senior principal within our financial services practice in Germany with a focus on compliance, AML and KYC areas. He has extensive consulting experience within the e-commerce and financial services domain. Gerardo has enabled several European companies and financial institutions to overcome the challenges posed by disruptive technologies and transform into digital-oriented organizations. Gerardo holds an MBA in business administration from the University of Mannheim. He is a certified AML officer and has a risk management certification from the Goethe Business School – Frankfurt University.

Duc Thang Hoang
Analyst & Data Scientist, Infosys Consulting
Duc Thang is an Analyst within the SAP Data Analytics practice in Germany with a focus in automotive and financial services areas. He gained technical backgrounds in data analytics and machine learning in his previous projects. His focus lies in business intelligence topics leveraging data-driven decisions using machine learning. Thang holds a master’s degree in automotive engineering at Technical University of Berlin.