Since 2018, following significant changes in the Latvian financial sector aimed at preventing Latvia's inclusion on the FATF (Financial Action Task Force) "grey list," the cross-border payment flow associated with Latvia has experienced a substantial decline. However, the analysis of cross-border payment data from credit institutions operating in Latvia is very resource-intensive, given the enormous volume of data involved.
At Latvijas Banka, we have developed an unsupervised machine learning tool to detect the level of anomalies in data from cross-border payments made by customers of Latvian credit institutions. This tool helps improve the efficiency of analysing data from cross-border payments made by customers of credit institutions by assessing potential threats of money laundering. In developing this tool, we also drew inspiration from the cross-border money flow analysis conducted by the International Monetary Fund [1].
Methodology
Payment data represent a relatively large dataset that does not include indications of historical anomalies. Therefore, an adequate and effective method is an unsupervised machine learning method, given that anomaly parameters have not been pre-identified. This technique not only facilitates efficient data processing but is also neutral to potential changes. The selected Isolation Forest algorithm functions by identifying observations that can be isolated from the rest of them with a small number of choices.
Payment data are anonymous and aggregated into batches, which means the observations are not perfectly isolated. Nonetheless, several observational parameters enable the creation of an appropriate set of variables for the algorithm to operate. Several parameters were developed with a focus on the critical characteristics of money laundering (ML).
- First, the volume of payments and the average transaction amount are used.
- Second, the economic indicators of the destination countries (i.e. the countries to which payments from Latvia have been dispatched or from which they have originated) are used – the proportion of these payments in relation to each country's gross domestic product, trade figures (whether imports or exports depending on the transaction direction), and direct investment indicators.
- Third, risk indicators in the ML field pertaining to destination countries are employed. The AML index [2], the corruption perceptions index , and the financial secrecy index [3] are used as indirect indicators. All indices are scaled from 0 to 1, reflecting a range from the lowest to the highest risk, and are then multiplied by the payment volume, thereby indicating the significance of the risk. The deviation of the specific payment group from seasonal patterns is also assessed, examining all payments with the respective country across relevant calendar quarters.
These parameters collectively offer nine dimensions, enabling a simultaneous analysis of all data across these dimensions. Thus, the analysis considers the degree to which each case within a payment group diverges from other cases, both in terms of how much its volume is attributable to Latvia's economic ties with the respective country and the impact of that country's risk factors relative to the payment size. Furthermore, the analysis incorporates the intrinsic characteristics of the payment itself, including its size, average transaction amount, and seasonal variations.
The model also incorporates a calibration factor known as the contamination index, which provides an estimate of the prevalence of anomalies within the dataset.
Results
The graph illustrates the trend in anomaly counts since the inception of the dataset, utilising a calibration factor of 0.05%. Anomalies related to payments with Russia see a marked decline in both incoming and outgoing payments, reflecting the impact of the AML regulatory framework reform implemented in 2018 and the geopolitical shifts triggered by Russia's invasion of Ukraine in 2022. On the other hand, the number of anomalies in payments with nearby neighbouring countries might suggest that the algorithm does not consider geographic proximity as a factor in explaining substantial payment volumes. While close economic and financial ties are evident in trade and investment data, the payment volumes and associated anomalies in transactions with neighbouring countries markedly differ from other cross-border flows. This disparity remains pronounced even when compared to more distant countries with similar trade links. In such cases, it is important to focus on the dynamics, such as observing fluctuations in the level of payment anomalies, whether they are increasing or decreasing.
Observations and application
It is important to note that the identified payment anomalies provide additional information for understanding threats related to cross-border payment flows; however, they do not necessarily indicate suspicious transactions. The applicable outcome provides an additional tool for enhancing risk awareness related to cross-border flows by automatically identifying notably atypical payments, with a specific emphasis on the field of ML risks. A major advantage of this solution is its minimal need for user input in defining anomaly typologies in a rapidly changing world. This tool also does not guarantee that anomalies necessarily indicate high-risk payments; therefore, expert interpretation is crucial for accurately assessing the results. Consequently, the distribution of anomalies by country can illuminate evolution of money flows and pinpoint the directions that require closer scrutiny. It is also essential to ascertain whether experts understand the underlying causes of changes. Changes in the overall level of anomalies over time can offer insights into the evolving ML risk across the entire sector.
Finally, Latvia has made considerable strides in mitigating and managing the ML risk; however, continuous and vigilant supervision is essential.
This piece was originally published at the Latvian central bank's economics website www.macroeconomics.lv, where you can also find more analysis of topical financial issues.
