Emergence of Artificial Intelligence and Machine Learning 

The rise of ChatGPT has brought generative artificial intelligence (GenAI) into the mainstream, accelerating adoption across industries ranging from healthcare to banking. The pace at which (Gen)AI is being used is outpacing prior technological advances, putting pressure on individuals and companies to adapt their ways of working to this new technology. While GenAI uses data to create new content, traditional AI is typically designed to perform specific tasks such as making predictions or classifications. Both approaches are built on complex machine learning (ML) models which, when applied correctly, can be highly effective even on a stand-alone basis.  

Figure 1: Model development steps

Though ML techniques are known for their accuracy, a major challenge lies in their complexity and limited interpretability. Unlocking the full potential of ML requires not only technical expertise, but also deep domain knowledge. Asset and Liability Management (ALM) departments can benefit from ML, for example in the area of behavioral modelling. In this article, we explore the application of ML in prepayment modelling (data processing, segmentation, estimation) based on research conducted at a Dutch bank. The findings demonstrate how ML can improve the accuracy of prepayment models, leading to better cashflow forecasts and, consequently, a more accurate hedge. By building ML capabilities in this context, ALM teams can play a key role in shaping the future of behavioral modelling throughout the whole model development process.

Prepayment Modelling and Machine Learning

Prepayment risk is a critical concern for financial institutions, particularly in the mortgage sector, where borrowers have the option to repay (a part of) their loans earlier than contractually agreed. While prepayments can be beneficial for borrowers (allowing them to refinance at lower interest rates or reduce their debt obligations) they present several challenges for financial institutions. Uncertainty in prepayment behaviour makes it harder to predict the duration mismatch and the corresponding interest rate hedge.

Effective prepayment modelling by accurately forecasting borrower behaviour is crucial for financial institutions seeking to manage interest rate risks. Improved forecasting enables institutions to better anticipate cash flow fluctuations and implement more robust hedging strategies. To facilitate the modelling, data is segmented based on similar prepayment characteristics. This segmentation is often based on expert judgment and extensive data analysis, accounting for factors like loan age, interest rate, type, and borrower characteristics. Each segment is then analyzed through tailored prepayment models, such as a logistic regression or survival models.¹

ML techniques offer significant potential to enhance segmentation and estimation in prepayment modelling. In rapidly changing interest rate environments, traditional models often struggle to accurately capture borrower behaviour that deviates from conventional financial logic. In contrast, ML models can detect complex, non-linear patterns and adapt to changing behaviour, improving predictive accuracy by uncovering hidden relationships. Investigating such relationships becomes particularly relevant when borrower actions undermine traditional assumptions, as was the case in early 2021, when interest rates began to rise but prepayment rates did not decline immediately.

Real-world application

In collaboration with a Dutch bank, we conducted research on the application of ML in prepayment modelling within the Dutch mortgage market. The applications include data processing, segmentation, and estimation followed by an interpretation of the results with the use of ML specific interpretability metrics. Despite being constrained by limited computational power, the ML-based approaches outperformed the traditional methods, demonstrating superior predictive accuracy and stronger ability to capture complex patterns in the data. The specific applications are highlighted below.

Data processing

One of the first steps in model development is ensuring that the data is fit for use. An ML technique that can be commonly applied for outlier detection is the DBSCAN algorithm. This clustering method relies on the concept of distance to identify groups of observations, flagging those that do not fit well into any cluster as potential outliers. Since DBSCAN requires the user to define specific parameters, it offers flexibility and robustness in detecting outliers across a wide range of datasets.

Another example is an isolation forest algorithm. It detects outliers by randomly splitting the data and measuring how quickly a point becomes isolated. Outliers tend to be separated faster, since they share fewer similarities with the rest of the data. The model assigns an anomaly score based on how few splits were needed to isolate each point, where fewer splits suggest a higher likelihood of being an outlier. The isolation forest method is computationally efficient, performs well with large datasets, and does not require labelled data.

Segmentation

Following the data processing step, where outliers are identified, evaluated, and treated appropriately, the next phase in model development involves analyzing the dataset to define economically meaningful segments. ML-based clustering techniques are well-suited for deriving segments from data. It is important to note that mortgage data is generally high-dimensional and contains a large number of observations. As a result, clustering techniques must be carefully selected to ensure they can handle high-volume data efficiently within reasonable timelines. Two effective techniques for this purpose are K-means clustering and decision trees.  

K-means clustering is an ML algorithm used to partition data into distinct segments based on similarity. Data points that are close to each other in a multi-dimensional space are grouped together, as illustrated in Figure 2. In the context of mortgage portfolio segmentation, K-means enables the grouping of loans with similar characteristics, making the segmentation process data-driven rather than based on predefined rules.  

Figure 2: K-Means concept in a 2-dimensional space. Before, the dataset is seen as a whole unstructured dataset while K-means reveals three different segments in the data

Another ML technique useful for segmentation is the decision tree. This method involves splitting the dataset based on certain variables in a way that optimizes a predefined objective. A key advantage of tree-based methods is their interpretability: it is easy to see which variables drive the splits and to assess whether those splits make economic sense. Variable importance measures, like Information Gain, help interpret the decision tree by showing how much each split reduces the entropy (uncertainty). Lower entropy means the data is more organized, allowing for clearer and more meaningful segments to be created.    

Estimation

Once the segments are defined, the final step involves applying prediction models to each segment. While the segments resulting from ML models can be used in traditional estimation models such as a logistic regression or a survival model, ML-based estimation models can also be used. An example of such an ML estimation technique is XGBoost. The method combines multiple small decision trees, learning from previous errors, and continuously improving its predictions. It was observed that applying this estimation method in combination with ML-based segments outperformed traditional methods on the used dataset. 

Interpretability

Though the techniques show added value, a significant drawback of using ML models for both segmentation and estimation is their tendency to be perceived as black-boxes. A lack of transparency can be problematic for financial institutions, where interpretability is crucial to ensure compliance with regulatory and internal requirements. The SHapley Additive exPlanations (SHAP) method provides an insightful way for explaining predictions made by ML models. The method provides an understanding of a prediction model by showing the average contribution of features across many predictions. SHAP values highlight which features are most important for the model and how they affect the model outcome. This makes it a powerful tool for explaining complex models, by enhancing interpretability and enabling practical use in regulated industries and decision-making processes.

Figure 3 presents an illustrative SHAP plot, showing how different features (variables) influence the ML model’s prepayment rate predictions on a per-observation basis. These features are given on the y-axis, in this case 8 features. Each dot represents a prepayment observation, with its position on the x-axis indicating the SHAP value. This value indicates the impact of that feature on the predicted prepayment rate. Positive values on the x-axis indicate that the feature increased the prediction, while negative values show a decreasing effect.  For example, the feature on the bottom indicates that high feature values have a positive effect on the prepayment prediction. This type of plot helps identify the key drivers of prepayment estimates within the model as it can be seen that the bottom two features have the smallest impact on the model output. It also supports stakeholder communication of model results, offering an additional layer of evaluation beyond the conventional in-sample and out-of-sample performance metrics.

Figure 3: Illustrative SHAP plot

Conclusion

ML techniques can improve prepayment modelling throughout various stages of the model development process, specifically in data processing, segmentation and estimation. By enabling the full potential of ML in these facets, future cashflows can be estimated more precisely, resulting in a more accurate hedge. However, the trade-off between interpretability and accuracy remains an important consideration. Traditional methods offer high transparency and ease of implementation, which is particularly valuable in a heavily regulated financial sector whereas ML models can be considered a black-box. The introduction of explainability techniques such as SHAP help bridge this gap, providing financial institutions with insights into ML model decisions and ensuring compliance with internal and regulatory expectations for model transparency.

In the coming years, (Gen)AI and ML are expected to continue expanding their presence across the financial industry. This creates a growing need to explore opportunities for enhancing model performance, interpretability, and decision-making using these technologies. Beyond prepayment modelling, (Gen)AI and ML techniques are increasingly being applied in areas such as credit risk modelling, fraud detection, stress testing, and treasury analytics.

Zanders has extensive experience in applying advanced analytics across a wide range of financial domains, e.g.:

• Development of a standardized GenAI validation policy for foundational models (i.e., large, general-purpose AI models), ensuring responsible, explainable, and compliant use of GenAI technologies across the organization.
• Application of ML to distinguish between stable and non-stable portions of deposit balances, supporting improved behavioural assumptions for liquidity and interest rate risk management.
• Use of ML in credit risk to monitor the performance and stability of the production Probability of Default (PD) model, enabling early detection of model drift or degradation.
• Deployment of ML to enhance the efficiency and effectiveness of Financial Crime Prevention, including anomaly detection, transaction monitoring, and prioritization of investigative efforts.  

Please contact Erik Vijlbrief or Siska van Hees for more information.

We have developed a machine learning model for a leading Dutch bank with over EUR 300 billion in assets to detect potential money laundering activities within its high-net-worth client segment.

Challenge​

Anti-Money Laundering (AML) models typically struggle with a common issue: a limited number of true positive cases for effective model training. To address this, most AML models incorporate some form of anomaly detection to identify unusual patterns in client behavior.​

We focused on the bank’s wealthiest clients, identified by a minimum asset threshold. This presents a unique challenge because, by nature, these clients are already statistical outliers. As a result, we needed to identify anomalies within this group of outliers, significantly increasing the model's complexity.​

The bank has an existing model in place, and our role is to enhance its performance, with a focus on:

• Redeveloping peer groups
• Reducing false positives in AML detection

Model Development

Like most machine learning projects, the development lifecycle is divided into three key phases: feature engineering (which takes up most of the time), modeling, and testing/implementation.​

We designed model features to ensure that normal client behavior corresponds to lower values, while anomalous behavior triggers higher values. This approach enhances the effectiveness of anomaly detection models.​

Collaborating closely with operations analysts, we refined these features to minimize obvious false positives among top-scoring cases. As a result, clients with legitimate activities are less likely to receive high anomaly scores in the final model.​

Peer Groups​

Detecting anomalous behavior among high-net-worth clients—who are all outliers by nature and exhibit highly diverse transaction patterns—requires a nuanced approach. To address this, we grouped clients based on their transaction behaviors to form peer groups.​

Key features were then evaluated by measuring how much a client’s behavior deviated from that of their peers. This method identifies anomalies by comparing clients to peers with similar transaction patterns.​

Our role involved revisiting and refining these peer groups to enhance the effectiveness of peer-based features, ultimately improving the model’s overall performance.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.

VAT Carousel Fraud (VCF) is a significant issue in the EU, costing an estimated €25-50 billion annually. We helped one of the largest Dutch banks become the first to develop a machine learning model specifically tailored to detect this type of fraud.

Challenge​

A team within the bank manually identified several cases of VCF each month. The objective was to develop a machine learning model to replace these investigations while leveraging the team’s expertise. The challenge was to create an effective model despite having a limited number of true positive cases for training.

Model Development​

Most Anti-Money Laundering (AML) models focus on detecting a broad range of money laundering activities. However, anomaly detection models are particularly effective at identifying outliers across diverse behaviors.

In the case of VAT Carousel Fraud (VCF), the behavioral patterns of the missing trader role are more distinct and consistent. To address this, we implemented a hybrid approach that combines supervised and unsupervised machine learning models.

The approach is summarized in the following steps:​

1 - Feature Engineering​

• Converted risk indicators into features, focusing on aspects like network structures and rapid movement of funds.​

2 - Supervised Model​

• Employed XGBoost to identify missing traders within the carousel fraud.​
• Utilized all available true positives to train the model on recognizable patterns.​

3 - Unsupervised Model​

• Implemented Isolation Forest to detect other roles in the carousel fraud.​
• Focused on outlier detection to identify anomalous behavior.

Performance​

Given the large scale of VAT fraud within the EU and the well-defined transactional typologies, we expect the model to deliver strong performance.

The first VAT Carousel Fraud (VCF)-specific models are now in production. A set of alerts was generated using real transaction data and reviewed by experienced analysts, achieving a 20% precision rate in identifying suspicious activities.

​VCF explained ​

VAT carousel fraud, also known as missing trader fraud, exploits the VAT system by allowing companies to import goods VAT-free within the EU, sell them domestically while collecting VAT, and then fail to remit the VAT to tax authorities.

The goods are sold through a chain of companies and eventually exported again, enabling the final exporter to reclaim the VAT. This cycle can be repeated multiple times, leading to substantial tax losses for governments.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.

We have helped a Dutch bank with over 500 billion in assets understand and realize their data ambitions regarding customer due diligence, sanctions, transaction monitoring, and fraud.

Challenge​

Entering the final stages of remediation, the bank wishes to develop a best-in-class data strategy in the financial crime domain. Having spent the last few years focusing on ensuring short-term compliance, the management team requested Zanders’ help to transition into an institution capable of tackling financial crime proactively.

This project aligns with a wider trend in the financial industry, where institutions, after addressing regulatory findings, invest in augmenting and automating their financial crime systems as a stepping stone toward an integration phase with a holistic view of client risk.

Solution

Zanders proposed a three-step approach:

• Step 1: Determine the current maturity state​
• Step 2: Work along with stream leads to determine ambition
• Step 3: Create and execute a roadmap to guide the bank through until 2027​

In Step 1, we determined the current state of data management regarding customer due diligence (CDD), sanctions, transaction monitoring (TM), and fraud. Since data is a multifaceted and all-encompassing element of an institution’s fight against financial crime, we divided our investigation into six key themes. This structure allowed for better alignment with stream leads within the bank while also enabling comparisons with best practices across the financial industry.

During Step 2, we assisted stream leads in identifying pain points and future objectives, thereby developing ambitions for each of the six themes. These ambitions balanced the bank’s desire to foster a cutting-edge data policy while still being actionable given the available resources—technical and otherwise.

Finally, in Step 3, not only did we bring all ambitions together into a coherent roadmap defining the data strategy through 2027, but we also began executing this roadmap immediately, minimizing the time between vision and realization to maximize value creation.

Moreover, a key deliverable was a comprehensive overview of the primary dataflows between departments. Our experience shows that, in trying to ensure short-term compliance, financial institutions often inherit a legacy of tangled dataflows, where data origins are obscured and key features are redundantly recalculated.

The first step toward resolving this issue was to carefully analyze how data flows between different pillars (Transaction Monitoring, KYC, Fraud Detection, and Sanctions). Once identified, inefficiencies and vulnerabilities were addressed through improved architecture and governance.

Making the Data Transformation Visual

The broad scope of this project, combined with the large amount of data it involves, poses a risk that stakeholders may struggle to stay informed about developments and decisions.

That’s why, from the very beginning, we committed to visualizing the ongoing transformations by creating a Data Initiative Dashboard to track progress on key data initiatives. This tool enables leadership to monitor and adjust priorities throughout the execution phase and establishes a gold standard for reporting future initiatives in an informative and intuitive manner.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.

Our client faced a challenge with a large volume of KYC/CDD reviews that needed to be conducted periodically, but many did not contain material CDD risks. Zanders supported the development and management of the PR automation model, which automates cases requiring limited research and allows analysts to focus on more complex, high-risk cases.

Challenge​

The traditional approach to CDD case handling requires significant manual effort by analysts. However, many cases that require minimal investigation are still reviewed manually. As a result, the current approach is neither risk-based nor cost-effective.

Typically, a Client Risk Rating model classifies clients as low, medium, or high risk. These clients are then reviewed at set intervals, such as every five, three, or one year(s), respectively. During these periodic reviews, analysts spend considerable time reviewing cases with no significant changes since the last manual review. This process is inefficient, and improvements can be made to make it more risk-based.

Solution

The PR automation model identifies cases that require minimal research and processes them automatically. The process begins with analyzing the current group of clients. From this large dataset, a subset of low-risk clients is identified based on expert knowledge combined with data analysis.

Next, the model determines which additional automated checks are necessary to ensure that the case has not undergone material changes since the last manual review. With these additional checks in place, the case can be processed automatically.

Automating PRs is only possible with a strong data foundation. Zanders assisted not only in developing the PR model but also in ensuring that data quality meets the necessary standards.

Performance

The PR automation model delivers significant cost savings while improving the efficiency and effectiveness of CDD case handling. Additionally, Zanders supports clients in demonstrating to regulators that this model helps transition to a more risk-based CDD approach.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.

We conducted a comprehensive review of the model landscape and governance across all KYC & AML models in a leading Dutch bank with over 900 billion in assets.

Challenge​

Although the shift toward a risk-based approach for models and processes in a bank’s first line of defense (1LoD) is widely adopted across the industry, this approach is often lacking in second-line (2LoD) activities. As a result, limited compliance and validation resources are allocated to low-complexity, low-impact models, leaving less time for rigorous evaluation of high-impact, complex models.

We conducted our review by focusing on four key areas: first, clarifying the definition and classification of models; then building a risk-based approach to model risk management given this classification. Finally, we covered model ownership and the overarching landscape.

Model Definition and Classification

What qualifies as a model? This question is as straightforward as it is fundamental. Without a clear definition, institutions cannot build a coherent approach to model risk management. Under the existing framework, several "expert opinion"-based systems were subject to model validation requirements. By leveraging Zanders’ extensive market experience, we refined the definition to align with industry standards, effectively removing these expert systems from the model inventory.

The bank’s existing model risk classification, based on the likelihood and magnitude of reputational risk, did not require an overhaul but benefited from key refinements. Zanders recommended enhancing the framework by factoring in additional elements, such as automation and model complexity. These refinements resulted in a classification system suitable for leading industry models.

Risk-Based Model Risk Management

With a robust model classification established, a risk-based approach to model risk management was implemented. Each model was assessed based on its potential reputational risk and intrinsic complexity, with oversight measures adjusted accordingly.

For example, models with lower risk and complexity were validated using a more qualitative approach, eliminating unnecessary benchmarking and confidence level requirements.

Model Ownership

Ownership is a critical component of the financial crime model landscape. Without clear ownership structures, model maintenance and necessary improvements, such as those prompted by validation findings, become challenging.

A well-defined ownership structure ensures sufficient independence between model development, ownership, and validation, reinforcing accountability and governance.

Model Landscape

Within the Compliance and KYC domains, a wide range of models exist, broadly categorized under:

• Customer Due Diligence (CDD)
• Transaction Monitoring (TM)
• Screening
• Market Abuse
• Investment Risk

Following our review, we provided a detailed visual representation of this often highly complex model landscape. This offered the management team a unique, high-level perspective, helping identify key areas for improvement based on industry best practices and regulatory expectations.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.

We restructured the Financial Crime Prevention (FCP) model landscape of a bank with over 500 billion in assets to prioritize efficient model development and reduce analyst workloads.

Challenge​

Despite having a mature and well-resourced FCP department, this Dutch bank’s existing model landscape was overly complex. It lacked transparency in the alert generation process and produced too many unnecessary alerts. Additionally, the model development process was extremely time-consuming, causing a lengthy delay between development and implementation. This resulted in frustration within the department and exposed the bank to operational risks.

Our mandate was to develop a model landscape that:

• Improves the effectiveness of models within the landscape
• Enhances the efficiency of the model development process

Model Development

Any mature model development process follows key stages: data preparation, experimentation/modeling, implementation (e.g., within a pre-existing codebase), and model validation.

By developing a landscape where preparation, implementation, and validation are streamlined, model developers can dedicate more time to in-depth analysis and leverage cutting-edge modeling techniques. The result is a landscape populated by high-performing, advanced models.

To achieve this, we provided several key recommendations:

• Model Validation: By enacting a toll-gated approach to model validation, potential issues can be flagged earlier in the development process.
• Data: By emphasizing reusable and well-documented data elements (e.g., through feature stores or derived layers), features and tables can be shared, drastically reducing data preparation time.
• Implementation: By standardizing the model design process and adopting a robust MLOps framework, model implementation can become seamless and consistent.

Model Landscape

Within such a complex domain, having high-quality models is only the first step in tackling the risks associated with financial crime. To be truly effective, models must have minimal overlap while collectively maximizing coverage of perceived risks.

Once this balance is achieved, models can be embedded into a landscape designed for targeted signals (i.e., ongoing due diligence) rather than the former periodic review regime.

Below is a schematic view of a best-practice FCP model landscape. The trigger-based approach ensures that analysts assess only those clients or transactions worth investigating. This shift results in higher-quality Suspicious Activity Reports (SARs) and more engaged analysts.

Implementing our recommendations would reduce the time-to-market for a model in development from 24 months to 9 months, significantly decreasing analyst workloads while improving the efficiency and effectiveness of financial crime detection.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.

We assisted a private Dutch bank in managing the overwhelming volume of queries from bankers regarding Know Your Customer (KYC) and Customer Due Diligence (CDD) by developing an AI-powered chatbot.

Challenge​

Bankers frequently needed specific information to assess customer profiles, verify identities, and evaluate potential risks to comply with regulatory requirements and ensure due diligence. However, this information was scattered across approximately 30 documents in various formats, including Word documents and PDFs. As a result, CDD specialists were overwhelmed by a constant stream of queries, making the process inefficient and increasing the risk of inconsistent or inaccurate responses.

To address these inefficiencies, the client sought a solution to streamline information retrieval and provide accurate, consistent answers.

Solution

Zanders proposed a step-by-step approach to developing the CDD chatbot:

1 - Investigating requirements from CDD specialists and bankers.

2 - Developing the chatbot using Retrieval-Augmented Generation (RAG).

3 - Optimizing performance through rigorous validation.

4 - Deploying the chatbot for live use by the client.

In the initial phase, we held discussions with both CDD specialists and bankers to identify their key requirements. It became evident that the chatbot needed to deliver accurate, reliable information while ensuring response consistency. Given these requirements, the unstructured nature of the data, and the chatbot’s internal use case, we determined that a Generative AI (GenAI) chatbot would be the most effective solution.

For development, we leveraged Retrieval-Augmented Generation (RAG), a method that combines fast and relevant information retrieval with the power of advanced AI to generate accurate, context-aware responses. This ensured a reliable and informative user experience. The processes in this approach are shown in the figure below.

To validate the chatbot’s performance, we created a dataset with expected responses, fine-tuned hyperparameters, and conducted extensive accuracy testing.

Finally, to ensure a seamless deployment, we established a structured system for development, deployment, and continuous improvement. By leveraging pre-trained large language models (LLMs), we were able to rapidly deploy the chatbot and refine it based on real-world user feedback.

Performance

As a result, the client successfully implemented the CDD chatbot, allowing users to query the document corpus directly and receive responses in plain English, along with a list of reference sources used by the LLM. Thanks to the RAG approach and thorough validation, the chatbot consistently produced accurate and reliable answers.

The chatbot has significantly improved efficiency, enabling CDD specialists to manage inquiries more effectively while helping bankers conduct due diligence with greater speed and accuracy. This has led to a more streamlined and reliable CDD process.

For more information, visit our Financial Crime Prevention page, or reach out to Johannes Lont, Senior Manager.