Predicting Money Laundering Using Machine Learning and Artificial Neural Networks Algorithms in Banks

Abstract This paper aims to build a machine learning and a neural network model to detect the probability of money laundering in banks. The paper's data came from a simulation of actual transactions flagged for money laundering in Middle Eastern banks. The main findings highlight that criminal networks mainly use the integration stage to integrate money into the financial system. Fraudsters prefer to launder funds in the early hours, morning followed by the business day's afternoon time intervals. Additionally, the Naïve Bayes and Random Forest classifiers were identified as the two best-performing models to predict bank money laundering transactions.


Introduction
As governments around the world have increased their scrutiny of the gaming and real estate sectors, there has been a shift in money-laundering activities in the financial industry.An International Consortium of Investigative Journalists investigation identified that more than US $2 trillion in transactions were flagged by financial institutions between 1999 and 2017 (Yang et al., 2020).In many of these instances, criminal networks were probing and detecting weaknesses in the control and anti-money-laundering (AML) frameworks to launder illicit funds and integrate them into ostensibly legitimate assets (Hutton, 2020).The end goal is the final integration of the funds into legitimate business activities or investments.To address these concerns, there have been calls to employ artificial intelligence (AI) to build learning algorithms that detect money-laundering transactions.Regulatory technology in the form of computational intelligence will be transformative in that AI techniques can help reduce false positives or type I errors.The present paper attempts to examine the following questions: 1. How does the probability of money laundering vary by personnel and type of activities?
2. Which features are the strongest predictors of money laundering?

Contribution to practice
In the past decade, banks have dramatically shifted the way they operate.
The traditional model of banking, which places reliance on human expertise to detect instances of money laundering, is unable to keep up with the rapid pace of change taking place in the world of finance.Banks must adopt cutting-edge technologies like machine learning (ML) and artificial neural networks (ANN) to stay competitive and detect financial crimes.ML is a form of AI that allows computers to learn from data, identify patterns, and make predictions.Banks are already using this technology to detect fraud, assess risk, personalize customer service, and process and interpret large amounts of data.ML technology is particularly well-suited for credit scoring and money laundering detection tasks.In the ever-evolving world of finance, financial institutions that use ML and ANN will be able to stay ahead of criminal networks and prevent launders from infiltrating their systems.
As finance becomes more digitized, so too do criminals' methods to launder money.In the past, banks could relatively quickly identify suspicious activity by looking for patterns in cheque deposits and withdrawals.However, because of the proliferation of internet banking and other forms of digital payment, it is now far more challenging to monitor unlawful financial activities.Banks are fighting back using ML and AI to detect and block suspicious transactions automatically.By analyzing large datasets, ML and AI can help banks identify trends that may indicate money laundering.These methods, when combined with business intelligence, have the potential to provide a potent instrument in the fight against money laundering.Computational technology may help with customer due diligence by identifying account holders and signatures, account numbers, the name of the bank, and the signature on the account.
The current article aims to predict money-laundering activities in the banking industry using supervised ML classification techniques and a feedforward neural network model.The rest of this paper is structured according to the following format: The next section provides a a critical overview of AML and the computational intelligence literature.Next, a discussion of the methodology and research design is conducted with emphasis on the data cleaning and processing process.The algorithms used in this paper are also discussed.Finally, a discussion of the findings carried out that highlights areas areas for future research.

The literature on AML and computational intelligence
Money laundering is the process of laundering proceeds earned through criminal activities into clean money that appears to come from a legitimate source.
In other words, it places illegally sourced funds into the standard financial cycle or money circulation process by disguising them as clean money (Ardizzi et al., 2014;Sobh, 2020).Money laundering involves processing funds from underground activities like terrorism, cybercrime, drug trafficking, corruption, tax evasion, and quasi-legal activities such as concealment of income from public authorities (Habib et al., 2018;Karim et al., 2020;Tiwari et al., 2020).Converting illegitimate gain into legitimate income disrupts the legal process of money supply and corrupts financial institutions, which in turn benefits criminal networks (Ardizzi et al., 2014).Moreover, according to Hendriyetty and Grewal (2017), money laundering leads to an increase in shadow economic and criminal activities while reducing the tax collections required for the growth of a country (as cited in Tiwari et al., 2020).Likewise, Drayton (2002) and Dowers and Palmreuther (2003) stated that money laundering could stymie a nation's economic growth, lead to financial distortion, socioeconomic and monetary instability, higher corruption, and increased vulnerability to financial institutions (as cited in Habib et al., 2018;Loayza et al., 2019).This claim can be exemplified by the research done on 91 Italian states in which the total identified laundered cash from 2005 to 2008 was equivalent to 7% of Italy's GDP, of which threefourths of the money was sourced from illegal trafficking activities.The remaining one-fourth was acquired through extortion (Ardizzi et al., 2014).Likewise, Loayza et al. (2019) highlighted in their paper that Colombia saw a phase in 2001 and 2002 when the total value of illicit income was equal to 12% of its GDP and the volume of laundered assets increased from 8% to 14% of Colombia's total GDP.As a result, illegal activities like as tax evasion, corruption, extortion, and drug trafficking result in income loss for the government, internal market instability, erosion of private-sector efforts, volatile currency and interest rates, and political upheaval (Ofoeda et al., 2020;Sobh, 2020).
Money laundering is primarily a three-step process that involves placement, layering, and integration stages (Sobh, 2020;Tiwari et al., 2020).The first step, known as placement, involves the introduction of illegal funds into the financial system; the second stage, known as layering, involves a series of fictitious transactions that mask the true source of the cash (Al-Suwaidi & Nobanee, 2020; Ardizzi et al., 2012;Ofoeda et al., 2020).In the third integration stage, illicit money is converted into a legitimate source of revenue by investing it in real estate, stocks, or businesses (Ardizzi et al., 2012).According to Loayza et al. (2019), an illicit process can be categorized into two types of activities: first, the production of illegal goods (such as drugs) that have value in the illicit market; and second, activities like kidnapping, extortion, robbery, and fraud that redistribute wealth among the various classes of people (from rich to poor) but do not contribute to the economy.The illegal money generated through these two types of illicit activities can be laundered through "front companies," gold dealers, currency exchange houses, insurance companies, shell companies, wire systems, offshore banking, automobile dealerships, casinos, lawyers, and accountants," among others (McDowell andNovis, 2001, as cited in Ofoeda et al., 2020, p. 4).

Determinants of money laundering
To manage the illegal activities of an area, it is necessary to comprehend the elements that contribute to or determine the likelihood of money laundering.According to Karim et al. (2020), the four variables of the fraud diamond theory-rationalization, pressure, capacity, and opportunitydrive the illegal behaviors engaged in the money-laundering process.Following the fraud diamond theory, high living standards, greed for power and money, bad habits or financial need, loopholes in the current system, insecure e-money facilities, loose control of access to information, poor supervision, wealth distribution, and urgency, as well as a propensity to commit fraud and use specific skills to carry out laundering activities, represent the pressure, opportunities, rationalization, and capability legs of the diamond theory, respectively (Lokanan, 2019).
To identify and prevent money-laundering operations, it is vital to understand the other characteristics that should be considered when developing AML legislation and risk-mitigating AI algorithms.In this context, Reganati and Oliva (2018) have shown that the factors determining illegal behavior might differ by geography.In their paper, Reganati and Oliva (2018) demonstrated, for instance, that a region's education and corruption level influenced the mafia crime rate and money-laundering activities in northern and central Italy, whereas gambling and gaming habits heavily contributed to the presence of money-laundering activities in the southern region of Italy.Similarly, Amara and Khlif (2018) found that the rate of financial crime is strongly correlated with tax evasion and corruption as a nation's primary driver of financial crimes.
In addition, Ferwerda (2009) revolutionized the research on money laundering by demonstrating that "a) the probability of being caught for money laundering, b) the sentence for money laundering, c) the probability of being convicted for the predicate crime, and d) the transaction costs of money laundering are negatively related to the amount of crime" (p. 1) and that constructing laws and policies based on these factors will aid in reducing crime.In addition to the four factors of the diamond theory, the issues of corruption, education, organizational culture, working environment, money lust, the strictness of laws, the strength of the adopted audit standards, and the gender, age, source of funds, and number of bank accounts held by account holders are the major determinants of money laundering that determine the likelihood of the presence of illegal activities.

Smart analytics for money laundering detection
This section will address the role of technology in combating money laundering challenges.Even though several academics have presented numerous anomaly-detection and money laundering risk (MLR) mitigation models, it stands to reason that the intended outcomes would be achieved if technological or software solutions were integrated with those models (Lokanan, 2019).Similarly, various researchers and practitioners have favored technology to enhance the efficacy of anomaly detection and risk-mitigation models (Kansal, 2021;Singh & Best, 2019).These models need intelligent analytics technologies to identify suspicious activity via pattern recognition.Analytical methods such as link analysis and interactive data visualization have proven critical in identifying anomalous patterns and transforming them into visual representations for further human examination (Dilla & Raschke, 2015;Singh & Best, 2019).
More specifically, innovative analytics tools can be used to help detect and prevent money laundering.By analyzing transaction data, these tools can flag patterns that may indicate criminal activity (Ferwerda, 2009;Ferwerda et al., 2013).These transactions include but are not limited to money laundering using cash transactions, electronic transfers and payments, bank accounts, investment-related transactions, offshore activities, secure and unsecured lending, and laundering involving intermediaries.The information detected from these transactions can then be used by authorized personnel to form the basis of an investigation.In some cases, smart analytics tools may even provide real-time alerts to authorized persons when suspicious transactions are detected and allow for quick and effective action to be taken, potentially preventing large sums of money from being laundered (Ardizzi et al., 2018;Singh & Best, 2019).In conclusion, smart analytics tools can be valuable in the fight against money laundering.However, analytics is not particularly useful if data, such as financial transactions, are created rapidly and in a large volume since analysts would find it difficult to make quick and accurate choices when dealing with such a dynamic quantity of data (Ferwerda, 2009;Singh & Best, 2019).Moreover, money launderers regularly use system vulnerabilities and current laws to launder dirty money; hence, the linear and pattern-based analysis will be unsuccessful unless the system learns from its previous patterns and creates a new algorithm each time something new is observed (Sobh, 2020).
Another issue to consider is the cost of implementation.The overall cost of implementing a money laundering detection algorithm in real-time applications depends on several factors, including the type of algorithm used, the implementation's complexity, and the deployment's scale (Ferwerda, 2009;2013).For example, a simple rule-based algorithm could be implemented for a few hundred dollars, while a more complex machine learning-based algorithm could cost several thousand dollars.The cost also scales with the deployment size, so a large bank with millions of customers would incur much higher costs than a small financial institution (Dre_ zewski et al., 2015).In general, however, the cost of implementing anti-money laundering measures is typically a small fraction of the overall budget for compliance and risk management (Dilla & Raschke, 2015;Ferwerda et al., 2020).

Machine learning and artificial intelligence for money laundering detection
Despite these challenges, ML algorithms are popular and in demand because they can change as they read new data or patterns (Kansal, 2021;Lokanan & Sharma, 2022;Zhang & Trubey, 2019).For instance, Semmelbeck and Besaw (2019) used a random forest (RF) classification algorithm to identify the factors that could be important to detect whether a terrorist group is engaged in criminal activities or not and found that the temporal change in the organizational structure of the terrorist group is a red flag for money-laundering activities.Generally, ML algorithms used for detecting money-laundering activities can be of two types: supervised and unsupervised ML algorithms (Chen et al., 2018;Lopez-Rojas & Axelsson, 2012).Badal-Valero et al. (2018) proposed integrating Benford's rule with ML techniques such as logistic regression (LR), decision trees (DT), neural networks (NN), and random forests (RF) and employing any applicable approach depending on whether the data was balanced or imbalanced.
They found that Benford's law in association with NN works best on unbalanced data, whereas RF and LR perform best with Benford's law when balancing methods are applied or when Synthetic Minority Oversampling Technique (SMOTE) is applied to the data.Zhang and Trubey (2019) have compared the five supervised ML algorithms -DT, RF, support vector machine (SVM), artificial neural network (ANN), and Bayes Logistic Regression (BLR) against the standard Maximum Likelihood Logistic Regression (MLLR) and found that ANN performs best as a rare event classification algorithm.In contrast, SVM and RF can also generate comparable results when amalgamated with sampling methods.However, the challenge with supervised techniques is that the data must be devoid of biases and mistakes, the events in the historical data must be precisely characterized, and each input variable must be precisely recognized; otherwise, suboptimal results will be generated by the models (Zhang & Trubey, 2019).
Further, these suboptimal results can be avoided by using the XGBoost ML algorithm.Jullum et al. (2020) demonstrate through their paper that XGBoost is useful in fighting suboptimal results as it considers nonreported alerts, normal alerts, and flagged alerts equally to develop a detection algorithm that predicts the probability of money laundering based on the senders' or receivers' background information, as well as their previous actions and transaction history.Other than the use of boosted algorithms, it is recommended to use unsupervised ML algorithms to overcome this drawback of supervised algorithms.According to Salehi et al. (2017), unsupervised data techniques are more helpful in identifying money-laundering patterns and can be instrumental in improving the learning capacity of classification methods.For example, Chao et al. (2019) used data-mining methods to monitor abnormal behaviors in trade-based money-laundering activities.Improvements were observed in management efficiency, which will be beneficial to restraining cross-border capital flow and arbitrage for emerging markets and developing economies.Another way to overcome the drawbacks of supervised learning is to combine visuals with the deep learning ML algorithm called graph learning or clustering algorithm (Dre_ zewski et al., 2012;Weber et al., 2018).Indeed, Li et al. (2020) proposed using FlowScope-a multipartite graph and scalable algorithm-to plot the complete flow of monetary transaction money from source to destination.According to Li et al. (2020), FlowScope can outperform the state-of-the-art baselines in identifying the fraudulent accounts used in the synthetic and realworld datasets.

Electiveness of AI and other methods in money laundering detection
Models based on ML and AI are increasingly being used in a variety of financial crime applications, including money laundering prediction.While rule-based methods have long been the standard for detecting and preventing money laundering, AI models offer a more sophisticated approach considering various factors.Studies have shown that ML and AI models can outperform rule-based methods in several ways, including accuracy, speed, and scalability (see Arnone & Borlini, 2010;Ba & Huynh, 2018;Jullum et al., 2020;Singh & Best, 2019).Money laundering models based on ML and AI are said to have higher predictive accuracy than the traditional rule-based approach to detection (Chen et al., 2018;Jullum et al., 2020).Other studies showed that ML and AI models could handle more volumes of data than traditional rule-based systems and could do so in a fraction of the time (Lokanan, 2019;Salehi et al., 2017;Sarker, 2022).Finally, AI models are more effective than rule-based methods at detecting previously unknown money laundering schemes (Singh & Best, 2019;Zhang & Trubey, 2019).
Even though AI is still in its infancy regarding AML compliance, several financial institutions are already adopting it for transaction monitoring.However, AI is not the only method currently used to detect suspicious transactions (Singh & Best, 2019;Sobh, 2020).Banks have long relied on rules-based systems to flag suspicious transactions, and these systems are continuously being refined and updated (Mathuva et al., 2020).In addition, banks are also increasingly using behavior-based prediction models that focus on identifying anomalous patterns of behavior.Money laundering detection models based on ML and AI are trained on historical data to look for red flags signs of money laundering, such as sudden changes in account activity or large transfers to high-risk jurisdictions.By combining different prediction methods, banks can create a more comprehensive approach to detecting and preventing money laundering.
Although a lot has been done using ML algorithms in the field of fraud detection and credit default, there is a dearth of scholarship on the application of ML and AI algorithms to detect laundering activities (Jullum et al., 2020;Lokanan & Sharma, 2022).The lack of scholarship can be attributed to the complexity of money-laundering events and the unavailability of quantitative data (Chen et al., 2018;Tiwari et al., 2020;Zhang & Trubey, 2019).As noted by Canhoto (2021): [D]ue to the unavailability of high-quality, large training datasets regarding money laundering methods, there is limited scope for using supervised machine learning.Conversely, it is possible to use reinforced machine learning and, to an extent, unsupervised learning, although only to model unusual financial behaviour, not actual money laundering.(p. 1) Lopez-Rojas and Axelsson (2012) believe that synthetic data, in the absence of real data, can be used to stimulate the required dataset for ML algorithms; however, the downside of this approach is that a biased dataset can be generated, depending on how it has been simulated.That said, synthetic dataset does provide an avenue to build and train algorithms to detect money-laundering activities.They advise using synthetic data for experimentation and Multi-Agent Based Simulation (MABS) until alternative mechanisms for developing more realistic user datasets become accessible.The present paper attempts to fill this gap by using a simulated dataset of banking data from Middle Eastern banks.
H 1 : Ceteris paribus, the likelihood of money-laundering events is an increasing function that depends on gatekeepers and the timing of the transaction.

Data collection
Data for this project came from a simulation of money-laundering activities in Middle Eastern banks based on a real dataset.The data were simulated using similar features and data points that mirror actual transfers of the original dataset.The features used for the simulation were similar to the processes used in actual transactions.Both the production of money laundering and non-money laundering were considered.Attempts were made to simulate all aspects of money laundering and non-money laundering transactions and to provide a relatively complete simulator.The simulation is based on financial institutions' three processes of money-laundering techniques: placement, layering, and integration.In simulating each process, a rule was created to represent cash-in transfers and one for transferred-out funds (See Ferwerda et al., 2020).An important feature of the stimulated dataset is that it is flexible and produces a dataset with different parameters.

Type of transaction
The type of transaction is classified as either cash-in or transfer-out.The type of transaction was coded as a categorical variable and then transformed into dummy variables.

Level of crime
The level of crime refers to whether the money-laundering activities were committed by the head of the financial institution or by a colleague working in the same institution.The level of crime was coded as a categorical variable and then transformed into dummy variables.

Amount of money
The amount of money is a random continuous variable and represents the actual amount of funds processed through the transaction.

Date
Date is simply the day, month, and year of the transaction.The date was further reformatted using the date and time format in Python as categorical variables representing the days of the week and months of the year.

Time
The time of the transaction is a continuous measure rounded to the nearest hour.

Type of money laundering
The type of money laundering was coded to represent the three stages of the laundering process: placement, layering, or integration.Each of these variables was later transformed into a categorical variable.

Money laundering
The target variable was whether money laundering occurred or not.The target variable was coded to represent 1 when the transaction was classified as money laundering and 0 when it was classified as non-money laundering.The formula to represent the target variable is shown in Equation 1: y ¼ f1, money laundering 0, noÀ money launderingg

Statistical tool and performance metrics
Coding and analysis were conducted using the Python programming language in a Jupyter notebook.Scikit-learn was the library of choice.Scikitlearn is a popular ML library used to build and analyze ML algorithms.The Keras open-source software library was used to perform the ANN analysis in the Jupyter notebook.The ML algorithm and ANN model were evaluated using the accuracy scores.As this is a classification model, the confusion matrix was employed to identify false positives or type 1 error.The algorithm with the highest predictive accuracy was selected as the best classifier.A confusion matrix was used to describe the performance of the best classification algorithm on the test set (or unseen data).A classification report with the following performance metrics was also employed to evaluate whether the model using precision, recall, and the F-1 score was also used to assess whether the model was capturing the money-laundering category and not only the non-money-laundering classification.The Receiver Operating Characteristic (ROC) evaluation metric was used to plot the tradeoff of the false positive (x-axis) against the true positive (y-axis).The Area Under the Curve (AUC) performance metric distinguishes between the money-laundering and non-money-laundering classifications.A completed random model will produce an AUC of 0.5, and a perfect model will have an AUC of 1.In this regard, the higher the AUC, the better the model distinguishes between the positive class and the negative class.

Data preprocessing
In the data preprocessing stage, the dataset was checked for missing values.The level of crime feature is missing in 38% of the observations.As this is a categorical variable, the mode was used to impute the missing observations.All duplicate values were deleted from the dataset.The specific and unique values were identified for all the features.In cases where Not a Number (nan) values were identified, they were replaced with zero.Some features, such as source and destination ID, were dropped from the dataset.There was no way to determine the location (i.e., latitude and longitude) from the data.

Feature engineering
In the data preprocessing stage, the columns for some of the variables were changed to make them more readable.For example, "isfraud" was changed to "moneylaundering" to reflect whether a money laundering transaction occurred or not.The same was done for another type of money laundering.
As this was a large dataset comprising 2,340 observations, the data was divided into a 60/40 train/test split.A standard scalar technique was used to normalize the numeric features.The model will not be able to analyze dates and times as raw data.As such, the date and time data were first converted into categorical variables for Exploratory Data Analysis (EDA) and then converted back to numeric data for the ML and ANN algorithms to analyze.I further transformed the time into three categorical variables reflecting the morning, afternoon, and evening.There were only eight feature variables.Even though Spearman's rank correlation coefficient independent check for attributes listed the top five features, all 14 features were employed to build the final model.

Naive Bayes algorithm
The Naive Bayes (NB) theorem is one of the classifiers employed to predict money-laundering activities in financial institutions.Recall from Lokanan and Liu (2021) that the NB theorem is based on the probability that the output in class C given that X ¼ x can be estimated by P(y j x) from P(y), P(x), and P(x j y) and is represented by the following equation: where P(y j x) is the posterior probability of the target variable (y) given predictor variables (x, features).P(y) is the prior probability of class x P(x j y) is the likelihood which is the probability of predictor given class x P(x) is the prior probability of predictor The NB classifier works well for big datasets that include a large number of features and operates on the assumption that the input features are independent of one another for multinomial distributed data.To build the algorithm for this model, GaussianNB was use along with n_sample and n_features as the parameters.

Logistic regression algorithm
Logistic regression is one of the simplest and most established algorithms in ML classification models.Despite its name, logistic regression is a linear model use for classification when the target variable is binary or consists of multinomial indicators.Logistic regression uses the sigmoid function.The sigmoid function maps the real value to a value bounded between 0 and 1 (hence the logistic regression is used for classification models).The parameters employed for the logistic regression algorithm is solver ¼ liblinear.The basic assumption with logistic regression is that of the linear function b0þb1X, which is transferred using the sigmoid function S(t); then, no matter what values b0, b1, b2 … and bk X take, y (target variable) will always have values of 0 and 1 (e.g., fraud and nonfraud; money laundering or no money laundering, spam or no-spam).Logistic regression models use this equation to estimate the probability that y ¼ 1 given its size X as follows:

Random forest algorithm
Random forest (RF) is an ensemble classification method that is useful because it adds additional randomness to the data.Unlike the NB algorithm, RF trains many strong decision trees and combines their predictions through a bagging process.The RF model was trained on the following parameters: criterion ¼ "entropy," n_estimators ¼ 100, and random_state ¼ 123.A diagraphic illustration of the RF model is shown in Figure 1 below.As you can see from Figure 1, there are two sets: a training set labeled in blue and a test set (unseen data) labeled in green.After training the RF model on the training set (blue circles), the model is then evaluated on the test set (green circles).The scores from the trees of the test set (in this case, two) are then averaged to form the RF score for the classification model.
When using the Gini index to determine the branching off of nodes in the decision trees, the mathematical formula for the RF algorithm for classification data is represented by the following equation: CatBoast is an open-source gradient boosting library developed by Yandex (a Russian-based search engine) and is easy to use.CatBoast is very useful for datasets where a large number of the features are categorical variables.The CatBoast algorithm is based on gradient boosting and ML; it works great when data comes from different sources.Hence, it is useful for this dataset since the data was a concatenation of two sources of data: one that involves the transition amount, type of laundering, and date and time, and the other that contains the people (i.e., head or colleague) who were involved in the laundered activities.The parameters of the CatBoast algorithm are iterations ¼ 50, depth ¼ 3, learning_rate ¼ 0.1.

GridSearchCV
Based on the performance metric used to evaluate the model, the single algorithm outlined above will project the performance of the parameters that come with those algorithms.For these parameters, the only choice is to try all the possible values and then choose the best one.To further finetune the model and enhance the performance metrics, GridSearchCV ("grid search") is used.The grid search approach generates the best candidate from a specified list of parameters.There are two types of grid search: exhaustive and randomized.The exhaustive grid search approach optimizes the parameters to be included in the model, whereas the randomized grid search approach automatically selects the best parameters for the model.In this project, an exhaustive grid search was employed to select the best parameters for the model.The exhaustive grid search approach was chosen because all the possible parameters are evaluated, and the best possible parameters are retained.

Artificial neural networks
The ANN method is a generalized model that processes many layers of data to make a decision.As can be seen in Figure 2 below, ANN is a multilayered layer perceptron (MLP) approach, where the input features are given values or weights.The MLP method is a deep learning neural network approach composed of several perceptrons.As can be seen in Figure 2, the MLP method consists of three layers: the input layer, which receives the signals from the modes; the output layer, which calculates the weighted average of the single features; and the output layer, which receives the weighted sum from the output layer to make a decision.In classification problems, the decision will be based on the percentage or accuracy of the model to predict the outcome.The input layers are typically the feature variables of the model.The input layers pick up the signals (coefficients) and pass them on to the hidden layer, where the weighted average for each feature is calculated and passed on to the output layer, which delivers the results.ANN is like a black box.It is not supposed to be interpretable in terms of feature importance; rather, it is a useful algorithm for predictive models and can be analyzed using the same performance metrics as classical ML models.

Univariate analysis
Table 1 presents the descriptive statistics of the numerical features.It is important to note that the maximum amount of funds laundered in a single transaction was $7.95 million.There is also no significant difference between the average amount of funds laundered ($2.51 million) and the standard deviation ($2.34 million).
Table 2 shows the descriptive statistics of the categorical features.As shown in Table 2, there are five categorical features.This type of laundering represents the stages of money laundering: placement (type 1), layering (type 2), and integration (type 3).A closer look at Table 2 indicates that integration (type 3) was the top method used to launder funds.Note also that there are more transfer-out than transfer-in transactions.Interestingly, more colleagues (or employees) are involved in money laundering transactions than managers.
Figure 3 shows the most important features for predicting money-laundering activities in financial institutions.As seen in Figure 2, the hour of the day has the highest positive impact in predicting money-laundering transactions, followed by cash-in and transfer-out.These findings make sense because the amount of funds entering and leaving the system is more likely to be flagged by compliance officers if they are over the allotted amount or if there is any anomaly with the transaction.Surprisingly enough, the amount of money does not seem to be an important feature in predicting money laundering in financial institutions.Continuing from Figure 3, the timing of the laundering activities is significant.As can be seen in Figure 4, launderers are more likely to launder funds through on-site transactions and during business hours.Most of the  laundered funds occurred in the morning hours, followed by the afternoon time intervals.Not surprisingly, laundering activities do not occur during the nights and evenings because these are outside business hours.According to these results, there is a clear need for a heightened focus on monitoring transactions in the early morning hours of work days.

Bivariate analysis
Figure 5 presents the correlation matrix of the numerical features with money-laundering activities.There is a moderately positive correlation (0.57) between the head of the bank and money-laundering transactions.Conversely, there is an inverse or moderately weak relationship (À0.57) between colleagues and money-laundering transitions.These findings indicate that the more funds laundered by colleagues, the less likely the money-laundering transactions will be successful.Note also from Figure 1 that the placement stage of the laundering process is positively correlated (0.67) with money laundering.Funds that are transferred out have a negative (À1) correlation with money laundering, which indicates that once the transactions (cash-in) are placed in the financial system, they are integrated and become part of the criminal network's portfolio of assets.

Evaluating model performance
Evaluating classification models can be complex because of the different possible performance metrics to consider.The present study is based on a binary classification model to predict money-laundering transactions and is denoted as 1 (for money laundering) and 0 (for no money laundering).The classification of a particular observation can fall within one of several different outcomes, as shown below: Accuracy is simply the number of correctly classified observations (TP þ TN) divided by the total number of observations.Precision is how precise or accurate the model predicts the true class (in this case, money laundering).Recall, or sensitivity is the positive rate of the true class that has been correctively classified.The F1-measure is simply the harmonious mean of recall and precision and might be a better measure if there is a need to balance the scores between precision and recall in an imbalanced dataset (See Lokanan & Sharma, 2022;Kansal, 2021).
Table 3 presents the accuracy score of the algorithms.As noted in this study, there was not much difference in the respective scores.Quite notably, the NB and RF classifiers were the two best-performing models, both with 77.46% accuracy.Significantly enough, grid search, which involved hyperparameter tuning, did not improve the accuracy score of the model.Given that the NB and RF had the highest accuracy rates, it is logical to look further into their classification scores.As seen in Table 4 below, when compared to the RF model (.87), the NB (1.00) classification did an excellent job of not labeling an observation as money laundering that was not money laundering.On the other hand, RF did a better job of capturing more money laundering observations (0.72 versus 0.63).
It is essential to have a look at the confusion matrix of the NB and RF classifiers in order to get further knowledge.Figure 6, below, shows a sideby-side comparison of the NB and RF confusion matrix.There are two possible outcomes from the predicted class: money laundering and no money laundering.A closer look at Figure 6 shows that the NB classifier shows that the model correctly predicted money-laundering transactions 39.4% of the time and no-money-laundering transactions 38% of the time.Together, those numbers represent 77.4% classification accuracy.On the other hand, the RF model correctly predicted money-laundering transactions 32.9% of the time and no-money-laundering transactions 43.8% of the time.Taken together, those numbers represent 76.7% classification accuracy.The presence of money laundering was present 38% of the time and absented 62% of the time.Conversely, the RF classifier predicted money-laundering transactions 50.3% of the time and no-money-laundering transactions 49.7% of the time.Interestingly enough, the false positive or Type 1 error, where the models predicted no money laundering, but money laundering occurred, was 0% for the NB model compared to 6.5% for the RF model.When one considers the large percentage of true positives and negatives for both models, they are fairly good classifiers to predict money-laundering transactions.

Artificial neural network
The ANN model performed slightly better than the classical ML models.The accuracy of the training and test sets is 78% and 80%, respectively.These results indicate that the ANN model did an excellent job predicting money-laundering transactions.More importantly, the model did not suffer from underfitting or overfitting the data.The ANN model is, therefore, very good at generalizing from the test set.The precision score is 87%, and the recall score is 72%; these findings indicate that the model is very good at predicting money-laundering transactions and correctly identifying individuals who are laundering money through the financial system.Figure 7 presents the results for the ROC curve for both the training and test sets.
The AUC for the test set is 78% (rounded), which indicates that the model performance was decent when predicting whether there were money-laundering transactions or not.

Conclusion
Money laundering is all about converting dirty money into clean funds.The involvement of financial institutions in money laundering cannot be underestimated.The ML and ANN algorithms employed in this paper perform reasonably well in identifying and labeling money-laundering transactions (see also Jullum et al., 2020;Tiwari et al., 2020;Zhang & Trubey, 2019).At the very least, compliance officers should use the findings presented here to scrutinize the features related to the laundering of funds.Features such as the time of the day and the amounts of money coming in and transferred out should be comprehensively monitored and scrutinized with regulatory technology.Other factors that should be monitored closely are international payments, sudden changes in the source of income, considerable anomalies in the amounts of money transfers, and any other suspicious activity that should be immediately scrutinized as part of the due-diligence process (Ba & Huynh, 2018).Financial institutions should use the findings from this paper to maintain lower money-laundering risks and conduct due-diligence background checks on the source of the cash-in funds and the destination of the transfer-out funds (Ba & Huynh, 2018;Tran & Nguyen, 2017).The machinelearning and ANN algorithms can be used to inform and continuously update money-laundering risks for each customer and incorporate new features such as salary, occupation, and source of income (Tran & Nguyen, 2017).Any abrupt changes in a customer's profile will act as red flags that would be eligible for scrutiny.Indeed, there should be proper training of bank employees and frontline workers to ensure that they are capable of identifying hot spots identified by the algorithms (Isa et al., 2015;Usman Kemal, 2014), while not ignoring their qualitative capabilities and phenomenologically lived human experiences and expertise to identify and report unusual activities (Usman Kemal, 2014).The findings presented in this paper support the claim that there is scope to develop ML and AI models to detect illicit activities in financial institutions.
Financial institutions should use the findings from this paper to maintain lower money-laundering risks and conduct due-diligence background checks on the source of the cash-in funds and the destination of the transfer-out funds (Ba & Huynh, 2018;Tran & Nguyen, 2017).The ML and ANN algorithms can be used to inform and continuously update moneylaundering risks for each customer and incorporate new features such as salary, occupation, and source of income into the models (Tran & Nguyen, 2017).Any abrupt changes in a customer's profile will act as red flags that would be eligible for scrutiny.Indeed, there should be proper training of bank employees and frontline workers to ensure that they are capable of identifying hot spots identified by the algorithms (Isa et al., 2015;Usman Kemal, 2014), while not ignoring their qualitative capabilities and phenomenologically lived human experiences and expertise to identify and report unusual activities (Usman Kemal, 2014).The findings presented in this paper support the claim that there is scope to develop ML and AI models to detect illicit activities in financial institutions.
the relative frequency of the binary class, and n represent the number of classes CatBoast algorithm

Figure 6 .
Figure 6.Confusion matrix of Naïve Bayes and random forest classifiers.

Table 1 .
Descriptive statistics of numerical features.

Table 2 .
Descriptive statistics of numerical features.

Table 3 .
Accuracy score of algorithms.