Compliance Hub

Fraud Detection Using Machine Learning in Banking

Site Logo
Tookitaki
10 min
read

The financial landscape is evolving rapidly. With this evolution comes an increase in financial crimes, particularly fraud.

Financial institutions are constantly seeking ways to enhance their fraud detection and prevention mechanisms. Traditional methods, while effective to some extent, often fall short in the face of sophisticated fraudulent schemes.

Enter machine learning. This technology has emerged as a game-changer in the banking sector, particularly in fraud detection.

Machine learning algorithms can sift through vast volumes of transaction data, identifying patterns and anomalies indicative of fraudulent activities. This ability to learn from historical data and predict future frauds is revolutionising the way financial institutions approach fraud detection.

An illustration of machine learning algorithms analyzing transaction data

However, the implementation of machine learning in fraud detection is not without its challenges. Distinguishing between legitimate transactions and suspicious activity, ensuring data privacy, and maintaining regulatory compliance are just a few of the hurdles to overcome.

This article aims to provide a comprehensive overview of fraud detection using machine learning in banking. It will delve into the evolution of fraud detection, the role of machine learning, its implementation, and the challenges faced.

By the end, financial crime investigators and other professionals in the banking sector will gain valuable insights into this cutting-edge technology and its potential in enhancing their fraud detection strategies.

The Evolution of Fraud Detection in Banking

The banking sector has always been a prime target for fraudsters. Over the years, the methods used to commit fraud have evolved, becoming more complex and sophisticated.

In response, financial institutions have had to adapt their fraud detection systems. Traditional fraud detection methods relied heavily on rule-based systems and manual investigations. These systems were designed to flag transactions that met certain predefined criteria indicative of fraud.

However, as the volume of transactions increased with the advent of digital banking, these traditional systems began to show their limitations. They struggled to process the vast amounts of transaction data, leading to delays in fraud detection and prevention.

Moreover, rule-based systems were often unable to detect new types of fraud that did not fit into their predefined rules. This led to a high number of false negatives, where fraudulent transactions went undetected.

The need for a more effective solution led to the exploration of machine learning for fraud detection.

Traditional Fraud Detection vs. Machine Learning Approaches

Traditional fraud detection systems, while useful, often lacked the ability to adapt to new fraud patterns. They were rigid, relying on predefined rules that could not capture the complexity of evolving fraudulent activities.

Machine learning, on the other hand, offers a more dynamic approach. It uses algorithms that learn from historical transaction data, identifying patterns and anomalies that may indicate fraud. This ability to learn and adapt makes machine learning a powerful tool in detecting and predicting future frauds.

Moreover, machine learning can handle large volumes of data, making it ideal for the digital banking environment where millions of transactions occur daily.

Limitations of Conventional Systems in the Digital Age

In the digital age, the volume, velocity, and variety of transaction data have increased exponentially. Traditional fraud detection systems, designed for a less complex era, struggle to keep up.

These systems often generate a high number of false positives, flagging legitimate transactions as suspicious. This not only leads to unnecessary investigations but can also result in a poor customer experience.

Furthermore, conventional systems are reactive, often detecting fraud after it has occurred. In contrast, machine learning allows for proactive fraud detection, identifying potential fraud before it happens. This shift from a reactive to a proactive approach is crucial in minimising financial loss and protecting customer trust.

{{cta-first}}

Machine Learning: A Game Changer in Fraud Detection

Machine learning has emerged as a game changer in the field of fraud detection. Its ability to learn from data and adapt to new patterns makes it a powerful tool in the fight against financial fraud.

Machine learning algorithms can analyze vast amounts of transaction data in real-time. They can identify complex patterns and subtle correlations that may indicate fraudulent activity. This level of analysis is beyond the capabilities of traditional rule-based systems.

Moreover, machine learning can predict future frauds based on historical data. This predictive capability allows financial institutions to take proactive measures to prevent fraud, rather than reacting after the fact.

Machine learning also reduces the number of false positives. It can distinguish between legitimate transactions and suspicious activity with a high degree of accuracy. This not only saves resources but also improves the customer experience.

However, implementing machine learning in fraud detection is not without its challenges. It requires high-quality data, continuous model training, and a deep understanding of the underlying algorithms.

Understanding Machine Learning Algorithms in Banking

Machine learning algorithms can be broadly classified into supervised and unsupervised learning models. Supervised learning models are trained on labeled data, where the outcome of each transaction (fraudulent or legitimate) is known. These models learn to predict the outcome of new transactions based on this training.

Unsupervised learning models, on the other hand, do not require labeled data. They identify patterns and anomalies in the data, which can indicate potential fraud. These models are particularly useful in detecting new types of fraud that do not fit into known patterns.

Both supervised and unsupervised learning models have their strengths and weaknesses. The choice of model depends on the specific requirements of the financial institution and the nature of the data available.

Regardless of the type of model used, the effectiveness of machine learning in fraud detection depends largely on the quality of the data and the accuracy of the model training.

Real-Time Transaction Monitoring with Machine Learning

One of the key advantages of machine learning is its ability to process and analyse large volumes of data in real-time. This is particularly important in the context of digital banking, where transactions occur around the clock and across different channels.

Real-time transaction monitoring allows financial institutions to detect and prevent fraud as it happens. Machine learning algorithms can analyse each transaction as it occurs, flagging any suspicious activity for immediate investigation.

This real-time analysis is not limited to the transaction itself. Machine learning models can also analyze the context of the transaction, such as the customer's typical behavior, the time and location of the transaction, and other relevant factors.

This comprehensive analysis allows for more accurate fraud detection, reducing both false positives and false negatives. It also enables financial institutions to respond quickly to potential fraud, minimising financial loss and protecting customer trust.

Implementing Machine Learning Models for Fraud Detection

Implementing machine learning models for fraud detection requires a strategic approach. It's not just about choosing the right algorithms, but also about understanding the data and the business context.

The first step is to define the problem clearly. What type of fraud are you trying to detect? What are the characteristics of fraudulent transactions? What data is available for analysis? These questions will guide the choice of machine learning model and the design of the training process.

Next, the data needs to be prepared for analysis. This involves cleaning the data, handling missing values, and transforming variables as needed. The quality of the data is crucial for the performance of the machine learning model.

Once the data is ready, the machine learning model can be trained. This involves feeding the model with the training data and allowing it to learn from it. The model's performance should be evaluated and fine-tuned as necessary.

Finally, the model needs to be integrated into the existing fraud detection system. This requires careful planning and testing to ensure that the model works as expected and does not disrupt the existing processes.

Supervised vs. Unsupervised Learning in Fraud Detection

In the context of fraud detection, both supervised and unsupervised learning models have their uses. The choice between the two depends on the nature of the problem and the data available.

Supervised learning models are useful when there is a large amount of labeled data available. These models can learn from past examples of fraud and apply this knowledge to detect future frauds. However, they may not be as effective in detecting new types of fraud that do not fit into known patterns.

Unsupervised learning models, on the other hand, do not require labeled data. They can identify patterns and anomalies in the data, which can indicate potential fraud. These models are particularly useful in detecting new types of fraud that do not fit into known patterns.

Regardless of the type of model used, the effectiveness of machine learning in fraud detection depends largely on the quality of the data and the accuracy of the model training.

The Role of Data Quality and Model Training

Data quality plays a crucial role in the effectiveness of machine learning models for fraud detection. High-quality data allows the model to learn accurately and make reliable predictions.

Data quality involves several aspects, including accuracy, completeness, consistency, and timeliness. The data should accurately represent the transactions, be complete with no missing values, be consistent across different sources, and be up-to-date.

Model training is another critical factor in the success of machine learning for fraud detection. The model needs to be trained on a representative sample of the data, with a good balance between fraudulent and legitimate transactions.

The model's performance should be evaluated and fine-tuned as necessary. This involves adjusting the model's parameters, retraining the model, and validating its performance on a separate test set.

Continuous monitoring and updating of the model is also essential to ensure that it remains effective as new patterns of fraud emerge.

Challenges in Machine Learning-Based Fraud Detection

Despite the potential of machine learning in fraud detection, there are several challenges that financial institutions need to address. One of the main challenges is the complexity of financial transactions.

Financial transactions involve numerous variables and can follow complex patterns. This complexity can make it difficult for machine learning models to accurately identify fraudulent transactions.

Another challenge is the imbalance in the data. Fraudulent transactions are relatively rare compared to legitimate transactions. This imbalance can lead to models that are biased towards predicting transactions as legitimate, resulting in a high number of false negatives.

The dynamic nature of fraud is another challenge. Fraudsters continuously adapt their tactics to evade detection. This means that machine learning models need to be regularly updated to keep up with new patterns of fraud.

Finally, there are challenges related to data privacy and security. Financial transactions involve sensitive personal information. Financial institutions need to ensure that this data is handled securely and that privacy is maintained.

Distinguishing Legitimate Transactions from Fraudulent Activity

Distinguishing between legitimate transactions and fraudulent activity such as credit card fraud is a key challenge in fraud detection. This is particularly difficult because fraudulent transactions often mimic legitimate ones.

Machine learning models can help to address this challenge by identifying patterns and anomalies in the data. However, these models need to be trained on high-quality data and need to be regularly updated to keep up with changing patterns of fraud.

False positives are another concern. These occur when legitimate transactions are incorrectly flagged as fraudulent. This can lead to unnecessary investigations and can disrupt the customer experience. Strategies to minimise false positives include refining the model's parameters and incorporating feedback from fraud investigators.

Ethical and Privacy Considerations in Data Usage

The use of machine learning in fraud detection raises several ethical and privacy considerations. One of the main concerns is the use of personal transaction data.

Financial institutions need to ensure that they are complying with data protection regulations. This includes obtaining the necessary consents for data usage and ensuring that data is stored securely.

There is also a need for transparency in the use of machine learning. Customers should be informed about how their data is being used and how decisions are being made. This can help to build trust and can also provide customers with the opportunity to correct any inaccuracies in their data.

Finally, there are ethical considerations related to the potential for bias in machine learning models. Financial institutions need to ensure that their models are fair and do not discriminate against certain groups of customers. This requires careful design and testing of the models, as well as ongoing monitoring of their performance.

Financial Institutions Winning the Fight Against Fraud

Financial institutions are increasingly turning to machine learning to combat fraud. This is not just limited to large multinational banks. Smaller banks and credit unions are also adopting these technologies, often in partnership with fintech companies.

One example is the Royal Bank of Scotland, which uses machine learning to analyze customer behaviour and identify unusual patterns. This has helped the bank to detect and prevent fraud, improving customer trust and reducing financial loss.

Another example is Danske Bank, which uses machine learning to detect money laundering. The bank's machine learning model analyses transaction data and flags suspicious activity for further investigation. This has helped the bank to comply with anti-money laundering regulations and has also reduced the cost of investigations.

These examples show that machine learning is not just a tool for the future. It is already being used today, helping financial institutions to win the fight against fraud.

{{cta-ebook}}

The Future of Fraud Detection in Banking

The future of fraud detection in banking is promising, with machine learning playing a central role. As technology continues to evolve, so too will the methods used to detect and prevent fraud.

Machine learning models will become more sophisticated, capable of analysing larger volumes of data and identifying more complex patterns of fraudulent activity. This will enable financial institutions to detect fraud more quickly and accurately, reducing financial loss and improving customer trust.

At the same time, the integration of machine learning with other technologies, such as artificial intelligence and blockchain, will enhance fraud detection capabilities. These technologies will provide additional layers of security, making it even harder for fraudsters to succeed.

The future will also see greater collaboration between financial institutions, fintech companies, and law enforcement agencies. By sharing data and insights, these organizations can work together to combat financial fraud more effectively.

Emerging Trends and Technologies

Several emerging trends and technologies are set to shape the future of fraud detection in banking. One of these is deep learning, a subset of machine learning that uses neural networks to analyse data. Deep learning can identify complex patterns and correlations in data, making it a powerful tool for detecting fraud.

Another trend is the use of behavioural biometrics, which analyses the unique ways in which individuals interact with their devices. This can help to identify fraudulent activity, as fraudsters will interact with devices in different ways to legitimate users.

Finally, the use of consortium data and shared intelligence will become more common. By pooling data from multiple sources, financial institutions can build more accurate and robust machine learning models for fraud detection.

Preparing for the Next Wave of Financial Crimes

As technology evolves, so too do the methods used by fraudsters. Financial institutions must therefore be proactive in preparing for the next wave of financial crimes. This involves staying up-to-date with the latest trends and technologies in fraud detection, and continuously updating and refining machine learning models.

Financial crime investigators will also need to develop new skills and expertise. This includes understanding how machine learning works, and how it can be applied to detect and prevent fraud. Training and professional development will therefore be crucial.

Finally, financial institutions will need to adopt a multi-layered security approach. This involves using a range of technologies and methods to detect and prevent fraud, with machine learning being just one part of the solution. By doing so, they can ensure that they are well-prepared to combat the ever-evolving threat of financial fraud.

Conclusion: Embracing Machine Learning for a Safer Banking Environment

In conclusion, as financial institutions strive to stay ahead of increasingly sophisticated fraud tactics, adopting advanced solutions like Tookitaki's FinCense becomes imperative.

With its real-time fraud prevention capabilities, FinCense empowers banks and fintechs to screen customers and transactions with remarkable 90% accuracy, ensuring robust protection against fraudulent activities. Its comprehensive risk coverage, powered by cutting-edge AI and machine learning, addresses all potential risk scenarios, providing a holistic approach to fraud detection.

Moreover, FinCense's seamless integration with existing systems enhances operational efficiency, allowing compliance teams to concentrate on the most significant threats. By choosing Tookitaki's FinCense, financial institutions can safeguard their operations and foster a secure environment for their customers, paving the way for a future where fraud is effectively mitigated.

Talk to an Expert

Ready to Streamline Your Anti-Financial Crime Compliance?

Our Thought Leadership Guides

Blogs
14 May 2026
6 min
read

AML Compliance for Remittance and Money Transfer Companies: An APAC Guide

It is a Thursday afternoon. Your firm is processing remittances on the Singapore–Philippines corridor — six thousand transactions before the weekend. You are licensed under MAS as a Major Payment Institution and registered as a Remittance and Transfer Company with the BSP in Manila. MAS published updated PSN02 guidance last month. This morning, the BSP examination schedule landed in your inbox. Two regulators. Two compliance programmes. One compliance team of four people. That is the daily operating reality for most APAC-licensed remittance operators, and it is the starting point for every AML programme design conversation.

This guide covers what money transfer AML compliance APAC-wide actually requires — by jurisdiction, by obligation, and by what good operational execution looks like.

Talk to an Expert

Why Remittance Companies Carry Higher AML Risk

FATF has consistently identified remittance and money transfer as a high-risk sector. Not because remittance operators are bad actors, but because of the transaction patterns that characterise the business.

Remittance is cash-intensive in many corridors. Some jurisdictions allow senders to pay in cash at agent locations with limited identification requirements. High-volume, low-value transactions create conditions where structuring — the practice of breaking amounts to stay below reporting thresholds — is easier to conceal than in lower-volume banking environments. A customer sending MYR 500 twice a week looks almost identical to a customer structuring around MYR 25,000 CTR thresholds.

FATF Recommendation 16 — the Travel Rule — applies specifically to wire transfers. Remittance companies are wire transfer originators. They must collect, transmit, and retain originator and beneficiary information with every qualifying transfer. This is not the same obligation as KYC. It is a data transmission requirement that sits on top of the CDD framework.

The cross-border nature of remittance creates bilateral exposure. A transfer from Singapore to Manila passes through both MAS and BSP oversight. A compliance failure — a missed STR, an inadequate CDD record, a Travel Rule data gap — does not stay in one jurisdiction. Both regulators can examine the same transaction.

The APAC corridors under heaviest examination scrutiny are among the highest-volume remittance corridors in the world: Singapore–Philippines, Malaysia–Bangladesh, Australia–India, and Philippines–Middle East. High volume does not reduce examiner focus. It increases it.

APAC Regulatory Obligations by Jurisdiction

Singapore (MAS)

Cross-border money transfer above SGD 3 million per month requires a Major Payment Institution licence under the Payment Services Act. The MAS PSA AML obligations for payment institutions are set out in PSN02, which covers CDD, ongoing monitoring, and STR and CTR filing requirements.

The FATF Travel Rule applies at SGD 1,500. For every transfer at or above that threshold, the MPS must transmit originator name, account number, and address or national identity number — plus beneficiary name and account number — to the receiving institution with the payment. The obligation to transmit sits with the sender regardless of whether the beneficiary institution can receive the data in structured form.

STR filing must occur within five business days of the determination that the transaction is suspicious. MAS examiners in 2024 specifically cited STR quality — not volume — as an examination focus area. An STR that describes the suspicious transaction in one sentence without analysis of the pattern does not meet the standard.

Australia (AUSTRAC)

All remittance dealers must register with AUSTRAC before commencing operations. Unregistered remittance dealing is a criminal offence under the AML/CTF Act 2006. This is not a technicality — AUSTRAC has prosecuted unlicensed remittance dealing, and its enforcement record includes actions against informal value transfer networks operating in parallel to registered dealers.

Registered remittance dealers carry the same AML/CTF programme obligations as banks under Chapter 16 of the AML/CTF Rules, without the same IT infrastructure to support them. Threshold Transaction Reports apply to cash transactions above AUD 10,000. Suspicious Matter Reports must be filed for qualifying transactions without a fixed deadline, but AUSTRAC expects prompt filing — delays beyond a few days are examined.

Malaysia (BNM)

Remittance operators require a Money Services Business licence under the MSB Act 2011. The AMLATFPUAA framework applies — the same statutory framework as banks — imposing CDD, ongoing monitoring, and STR and CTR obligations.

CTR threshold is MYR 25,000 for cash transactions. STR filing is required within three business days of the determination. BNM's most recent national risk assessment specifically identifies hawala-style informal remittance networks operating alongside licensed MSBs as a risk vector. That finding has translated directly into elevated examination scrutiny for licensed operators, who face more frequent and detailed examinations as regulators attempt to map the boundary between formal and informal channels.

Philippines (BSP)

Remittance operators require a Remittance and Transfer Company licence from the BSP. The AML programme obligations are set by AMLA and BSP Circular 950 — the same framework that governs banks, applied in full to RTCs.

CTR threshold is PHP 500,000. STR filing is required within five business days. The Philippines exited the FATF grey list in January 2023, but exit has not reduced examination pressure — BSP has increased examination frequency for RTCs since 2023, consistent with post-grey-list monitoring by both the BSP and AMLC.

New Zealand (DIA)

Remittance operators are Phase 2 reporting entities under the AML/CFT Act 2009, supervised by the Department of Internal Affairs. The same CDD, ongoing monitoring, and SAR and PTR obligations that apply to banks apply in full to remittance operators. The DIA's supervisory approach includes sector-wide audits and thematic reviews — it does not reserve examination resources only for larger entities.

The FATF Travel Rule in Practice for APAC Remittance Operators

FATF Recommendation 16 requires the originating institution to transmit originator and beneficiary information with every wire transfer above the applicable threshold. Across APAC, the operative thresholds are SGD 1,500 under MAS, AUD 1,000 under AUSTRAC, and USD 1,000 equivalent as the FATF baseline for jurisdictions without a lower domestic threshold.

The data that must travel with the payment: originator name, account number, address or national identity number; beneficiary name and beneficiary account number. These fields must populate the payment message — they cannot be retained on file at the sending institution and supplied only on request.

The operational problem is well-documented. Many beneficiary institutions in the corridors where APAC remittance volumes are highest — particularly in developing-market corridors — do not have systems capable of receiving structured Travel Rule data. The sending institution's obligation does not dissolve because the receiving institution lacks the infrastructure. Compliance requires transmitting the data within whatever message structure the payment uses: MT103 field population for SWIFT transactions, or the equivalent structured fields in ISO 20022 message formats.

Travel Rule technology solutions — TRISA, VerifyVASP, and Sygna Bridge are the most widely deployed in APAC for virtual asset transfers — are increasingly being applied to fiat remittance payment flows as well. For most APAC remittance operators on real-time domestic rails, the Travel Rule data obligation sits inside the payment message design, not in a separate data transmission layer.

ChatGPT Image May 14, 2026, 03_38_23 PM

Transaction Monitoring Requirements Specific to Remittance

High-volume, low-value transaction environments cannot be monitored with the dollar-threshold rules designed for retail banking. A rule that fires above USD 5,000 will miss the dominant remittance pattern entirely — hundreds of transactions at USD 200 to USD 500 per customer per month — and generate alert noise on the routine flows that constitute most of the business.

For an overview of how automated transaction monitoring works, the underlying detection logic matters more than the threshold level. Remittance monitoring is a typology problem, not a threshold problem.

Velocity monitoring is the primary detection method for mule accounts in remittance networks. The pattern is not a single large transfer — it is twenty transactions in forty-eight hours across multiple corridors from the same account or beneficial owner. A system calibrated only to flag high-value single transactions will not detect this.

Corridor-specific scenario calibration is not optional. The Singapore–Philippines corridor has different fraud typologies from the Malaysia–Bangladesh corridor. Monitoring scenarios applied generically across all corridors without tuning for the specific patterns in each one will produce both false positives on legitimate traffic and false negatives on actual suspicious activity.

Round-number structuring is the simplest pattern and the one most often missed by single-threshold rules. Transactions consistently placed just below the CTR threshold — MYR 24,500, AUD 9,800, PHP 499,000 — are a textbook structuring indicator. A rule with a single threshold at the CTR level will not catch this. The detection logic must look at the cluster of transactions below the threshold, not just the individual transaction value.

Beneficiary account reuse is a mule indicator: multiple unrelated customers sending to the same unfamiliar beneficiary account. This pattern requires a system capable of cross-customer analysis, not just single-customer transaction review. Rules-based systems that process each customer's alerts in isolation cannot detect it.

For remittance operators evaluating their technology choices, the same detection architecture issues apply as those covered in TM for payment companies and e-wallets — the product and customer profiles are different, but the architectural requirements for cross-customer scenario coverage are the same.

What Good Looks Like for a Multi-Jurisdiction Remittance Operator

A compliance officer managing two or three APAC licences simultaneously with a small team is not running a bank compliance programme at reduced scale. The operational structure is different.

A single TM platform across all jurisdictions is operationally necessary, not aspirational. Compliance officers in multi-jurisdiction firms who reconcile alerts from separate system instances — one per market — spend time on logistics that should go into analysis. The same transaction, flagged differently in two systems because the rule calibrations differ, creates reconciliation work that multiplies with volume.

Pre-settlement processing on real-time rails is required where payment is irrevocable on settlement. On PayNow, DuitNow, NPP, and InstaPay, a payment that clears cannot be recalled. Batch monitoring that runs after settlement has already processed the payment before the alert fires. The monitoring must run against the payment instruction before settlement, not the settled record.

Travel Rule data workflow integrated into the payment process eliminates the manual population of originator and beneficiary data as a separate step. When Travel Rule data handling is separated from payment processing and managed by different team members, the data quality degrades and the audit trail becomes inconsistent.

STR and CTR filing workflows built per jurisdiction address the material operational differences between regulatory regimes: different templates, different filing portals, different time windows, different field requirements. A case management system that requires the analyst to manually navigate those differences for each jurisdiction adds material risk. The workflows should enforce the right template for the jurisdiction of the filing, triggered by the currency of the transaction.

Selecting the right platform requires working through a structured evaluation. The Transaction Monitoring Software Buyer's Guide covers the criteria relevant to multi-jurisdiction operators, including how to assess vendor coverage across APAC regulatory regimes.

FinCense for APAC Remittance Operators

FinCense is deployed at remittance and payment operators across APAC — not only at banks. The platform is configured for the transaction patterns, corridor structures, and regulatory filing requirements that remittance operators encounter, not adapted from a banking deployment.

The scenario library includes more than fifty financial crime typologies covering the patterns most prevalent in remittance: mule account networks identified by cross-customer beneficiary account reuse, APP scam indicators in outbound payment flows, velocity structuring across corridors, and cross-border layering patterns. These are pre-built scenarios, not configurations that require the compliance team to write detection logic from scratch.

Pre-settlement processing is available across PayNow, DuitNow, NPP, InstaPay, and FAST — covering the real-time rails in Singapore, Malaysia, Australia, and the Philippines where irrevocable payment risk requires monitoring before settlement, not after.

Multi-jurisdiction STR and CTR filing workflows are built into the case management interface. Filing to AUSTRAC, BNM, AMLC, or MAS FIU from a single case triggers the correct jurisdiction-specific template, with the applicable time window displayed for the analyst at the case level.

In production deployments, FinCense has reduced false positive rates by up to 50% compared to legacy rules-based systems. For a remittance operator managing three hundred thousand transactions per month with a compliance team of four, a 50% reduction in false positive volume is not a performance metric — it is the difference between a workable alert queue and one that structurally cannot be cleared before the next batch arrives.

Book a demo to see FinCense configured for APAC remittance compliance — with corridor-specific scenarios already calibrated and multi-jurisdiction filing workflows built in.

For the full vendor evaluation framework, see the Transaction Monitoring Software Buyer's Guide.

AML Compliance for Remittance and Money Transfer Companies: An APAC Guide
Blogs
14 May 2026
6 min
read

Transaction Monitoring in Malaysia: BNM Requirements and Best Practices

Bank Negara Malaysia shifted from prescriptive to risk-based supervision several years ago. For transaction monitoring, that shift has specific consequences. Institutions that run static threshold-only systems — rules set at go-live and unchanged since — are increasingly out of step with what BNM examiners expect to see.

Malaysia's FATF Mutual Evaluation, conducted in 2021 and published in 2022, rated the country as partially compliant or non-compliant across several technical recommendations, including Recommendation 10 (customer due diligence) and Recommendation 16 (wire transfers). The evaluation flagged weaknesses in ongoing monitoring and STR quality at reporting institutions. BNM's supervisory response has been direct: examinations since 2022 have placed transaction monitoring programmes under considerably more scrutiny than before the assessment.

This article covers what BNM specifically requires from a transaction monitoring programme, the reporting thresholds institutions must meet, what examiners look for in practice, and where FinCense addresses the framework.

For background on Malaysia's full AML/CFT regulatory framework, see our overview of Malaysia's AML/CFT obligations under AMLATFPUAA and the BNM Policy Document.

Talk to an Expert

Malaysia's AML/CFT Regulatory Framework — the TM Foundation

Transaction monitoring in Malaysia sits on two legal instruments.

AMLATFPUAA 2001 (as amended) is the primary legislation. The Anti-Money Laundering, Anti-Terrorism Financing and Proceeds of Unlawful Activities Act 2001 establishes the obligations of Reporting Institutions — who they are, what they must do, and what penalties apply when they fail. The 2014 and 2020 amendments expanded the predicate offence list, brought Designated Non-Financial Businesses and Professions (DNFBPs) into scope, and raised maximum penalties to MYR 3 million per offence.

BNM's AML/CFT/CPF/TFS Policy Document (2023) is the operational standard. This is where BNM translates the Act's obligations into programme requirements — including the specific requirements for transaction monitoring systems, alert investigation processes, and calibration governance. When a BNM examiner cites a deficiency, the reference is almost always to the Policy Document, not to the Act itself.

Reporting Institutions under AMLATFPUAA cover a wide range of entities: licensed banks, Islamic banks, development financial institutions, insurance companies, capital market intermediaries, money services businesses, e-money issuers, digital banks, and — since the Phase 2 expansion in 2020 — lawyers, accountants, and real estate agents.

BNM supervises financial institutions. The Securities Commission supervises capital market intermediaries. The Companies Commission oversees designated company service providers. Each supervisor applies the AMLATFPUAA framework to its regulated population. For BNM-supervised institutions, the Policy Document is the day-to-day compliance standard.

What BNM's Policy Document Requires for Transaction Monitoring

Section 14 of the Policy Document covers ongoing monitoring and record-keeping. The requirements are specific.

Automated systems are mandatory. Institutions must implement an automated transaction monitoring system adequate for the nature, scale, and complexity of their business. Manual review of sampled transactions does not satisfy this requirement. The system must be capable of detecting patterns across the full transaction population, not a sample.

Calibration must reflect the institution's own risk profile. This is the element that static threshold systems most commonly fail on. BNM does not prescribe specific thresholds. It requires that the thresholds and scenarios in use reflect the institution's customer risk assessment — the output of the enterprise-wide risk assessment, not the vendor's default configuration. A rural cooperative bank and a digital bank processing international remittances have materially different customer risk profiles. The same rule library cannot serve both, and BNM's Policy Document makes clear that it is the institution's responsibility to demonstrate that calibration is appropriate to their specific population.

Monitoring must be continuous. BNM's ongoing monitoring language mirrors FATF Recommendation 10 — monitoring must operate across the full course of the customer relationship, not as a periodic batch process that reviews a subset of transactions once a month. For real-time payment channels, this has practical implications: batch processing that catches a transaction two days after settlement is not equivalent to monitoring at the point of transaction.

Every alert must be assessed and documented. BNM expects a documented investigation workflow. Each alert must be assessed, the assessment must be recorded, and the disposition — whether the alert is closed with rationale or escalated to STR review — must be traceable. An alert queue that shows "reviewed" with no supporting investigation record does not satisfy the Policy Document's requirements.

Calibration must be reviewed periodically. At minimum, BNM expects annual calibration reviews. Reviews are also required when the customer base or product profile changes materially — new product launch, significant customer segment growth, entry into a new geographic market. The review and any resulting threshold adjustments must be documented with dated sign-off from a senior compliance officer.

Section 11 of the Policy Document, which covers customer due diligence, is directly relevant to transaction monitoring design. The CDD risk classification assigned to each customer — standard, medium, or high risk — should determine the intensity of monitoring applied to that customer's transactions. An institution that applies identical monitoring rules to all customers regardless of CDD risk classification is not meeting the risk-based requirement.

ChatGPT Image May 14, 2026, 03_18_55 PM

Reporting Thresholds and STR Obligations

Cash Transaction Reports (CTRs). Transactions in cash or cash equivalents above MYR 25,000 must be reported to BNM's Financial Intelligence and Enforcement Department (FIED) within 3 business days of the transaction.

Suspicious Transaction Reports (STRs). There is no threshold for STR filings. The obligation is triggered by suspicion — when a compliance officer, having reviewed available information, determines that a transaction or pattern of transactions is suspicious. Once that determination is made, the STR must be filed with BNM/FIED within 3 business days.

The 3-business-day clock on STR filings is a common source of examination findings. Where the investigation workflow requires multiple sequential sign-offs before filing, the clock can expire before the report reaches the MLRO. Institutions whose internal escalation processes consistently result in filings on day 3 or later are at risk.

Tipping off prohibition. Institutions must not inform the customer — directly or indirectly — that an STR has been or will be filed. This prohibition extends to staff below compliance officer level and applies during the alert investigation process, not only at the point of filing.

Record retention. All transaction records and CDD documentation must be retained for 6 years from the end of the business relationship. BNM examiners reviewing a programme may request records from any point within that 6-year window. Institutions whose systems do not retain complete alert investigation records for the full retention period will be unable to demonstrate compliance for the period not covered.

Digital Banks and E-Money Issuers — Specific TM Considerations

BNM issued the Digital Bank licensing framework in 2022. Five digital banks have been licensed under that framework. They are subject to the same AMLATFPUAA obligations as conventional licensed banks — including the full Policy Document requirements for transaction monitoring systems, calibration, alert investigation, and reporting.

The assumption that digital banks operate under a lighter compliance perimeter than conventional banks is incorrect. BNM's licensing documentation is explicit: digital banks must meet equivalent standards, adapted for their operating model and customer base.

E-money issuers licensed under the Financial Services Act 2013 have tiered account structures. Tier 1 accounts carry a MYR 5,000 cumulative balance limit and are treated as lower-risk. That lower-risk designation reduces CDD intensity — it does not eliminate transaction monitoring obligations. E-money issuers must monitor for anomalies within the Tier 1 population, including patterns that would not be unusual in isolation but become suspicious in aggregate.

BNM's financial crime risk assessments have specifically identified typologies associated with digital banking and e-wallet channels:

  • Mule account layering through e-wallets, where proceeds move through multiple accounts in rapid succession before withdrawal
  • Rapid in-out velocity patterns — high-value inflows immediately followed by bulk transfers or withdrawals, with no plausible commercial purpose
  • Account takeover followed by bulk transfers, where the transaction pattern changes sharply after a suspected credential compromise

These typologies require specific monitoring rules. Generic monitoring scenarios designed for conventional banking products will not detect them reliably.

BNM has signalled through its 2025 e-money AML/CFT exposure draft that CDD and monitoring requirements for e-money issuers will be tightened if enacted — with specific requirements for transaction monitoring aligned to each institution's customer risk assessment rather than applied at the product level. Institutions that currently apply product-level defaults should treat this as a forward indicator of examination direction.

For BNM's specific KYC and CDD requirements for digital banks and e-money issuers, see our guide to BNM's digital bank and e-money KYC requirements.

Six Criteria for an Effective TM Programme Under BNM

These criteria are derived from BNM's Policy Document requirements and recurring examination findings.

1. Risk-based calibration. Alert thresholds and scenarios must reflect the institution's specific customer risk profile — the output of the enterprise-wide risk assessment, reviewed and updated when the population changes. Vendor defaults are a starting point, not a destination. BNM's examination record shows that institutions running unmodified vendor configurations are routinely cited.

2. Coverage of Malaysian financial crime typologies. BNM's financial crime risk assessments identify specific patterns relevant to the Malaysian market: cross-border trade-based money laundering, corporate account structuring, e-wallet mule networks, and instant payment fraud. These typologies must be in the active rule library, not on a watch list for future implementation.

3. Pre-settlement screening for instant payments. Malaysia's Real-time Retail Payments Platform — RPP, operating as DuitNow — processes irrevocable instant payments. Batch monitoring that reviews DuitNow transactions after settlement cannot intercept a suspicious payment. Pre-settlement evaluation logic, equivalent to what Singapore's PayNow and Australia's NPP require, is necessary for institutions with material DuitNow volumes.

4. Alert quality over alert volume. BNM examination findings have consistently cited alert investigation backlogs — queues with unreviewed alerts older than 30 days — as evidence of inadequate programme maintenance. A system that generates high alert volumes at low accuracy does not demonstrate active monitoring. It demonstrates an overwhelmed compliance function. Reducing false positive rates is not a nice-to-have; it is a programme governance requirement.

5. Explainable alert logic. Compliance analysts must understand why an alert was raised in order to make a quality investigation decision. A model that outputs a suspicion score without an explanation of which behaviours contributed to it puts the analyst in the position of making a filing decision based on a number rather than evidence. BNM examiners reviewing investigation records will ask the analyst what they found and why they made their disposition decision. "The system flagged it" is not an answer.

6. Documented calibration. BNM expects evidence that thresholds are reviewed and adjusted over time. A rule set deployed at system go-live and unchanged for two or three years — with no documentation of reviews, no record of what was considered and rejected, and no sign-off from senior compliance — is a finding in waiting. The documentation requirement exists regardless of whether the thresholds themselves are appropriate.

For a broader overview of how transaction monitoring works and what an effective programme requires, see our introduction to transaction monitoring.

Common BNM Examination Findings in Transaction Monitoring

Based on publicly available supervisory guidance and BNM examination themes, the following findings recur across reporting institutions:

Alert investigation backlogs. Queues with alerts unreviewed for more than 30 days are treated as a red flag. BNM examiners will ask how long the backlog has existed and what steps the compliance function took to address it.

Insufficient typology coverage for digital banking products. Institutions with e-wallet or digital banking products that apply conventional banking monitoring rules without product-specific scenarios are consistently cited for typology gaps.

No evidence of calibration review. Institutions that cannot produce documentation of when thresholds were last reviewed, what data informed the review, and who approved the outcome have a governance failure regardless of whether their thresholds happen to be appropriate.

STR filing delays. Investigation workflows with multiple sequential sign-offs that consistently result in filings on day 3 or later — or that have produced late filings — generate findings. BNM treats the 3-business-day requirement as a firm deadline, not a target.

Inadequate alert disposition documentation. An examiner reviewing a closed alert needs to understand the analyst's rationale. A disposition record that shows the alert was reviewed without documenting what was found, what was considered, and why the decision was made does not meet the Policy Document standard.

How FinCense Addresses the BNM Framework

FinCense is pre-configured with BNM-aligned typologies. The rule library includes DuitNow-specific scenarios — pre-settlement screening logic for instant payments — and e-wallet fraud patterns documented in BNM's financial crime risk assessments.

Alert thresholds are calibrated to each institution's customer risk assessment during implementation. Generic vendor defaults are not applied. The calibration rationale is documented and retained for examination review.

CTR and STR workflows are built into the case management module, with filing deadline tracking. Compliance officers see the filing deadline at the point of alert escalation, not after the 3-business-day window has passed.

In production deployments, FinCense has reduced false positive rates by up to 50% compared to legacy rule-based systems. For a compliance team managing 300 daily alerts, that reduction represents approximately 150 fewer dead-end investigations per day — which directly addresses the backlog problem that BNM examination findings most commonly cite.

Audit trail exports are structured for BNM examination review. Every alert record includes the rule or scenario that triggered it, the investigation timeline, the analyst's documented rationale, and the disposition outcome.

Taking the Next Step

For the complete vendor evaluation framework — including the seven questions to ask any transaction monitoring vendor — see our Transaction Monitoring Software Buyer's Guide.

Book a demo to see FinCense running against BNM-specific Malaysian financial crime scenarios, including DuitNow pre-settlement screening and e-wallet mule detection.

Transaction Monitoring in Malaysia: BNM Requirements and Best Practices
Blogs
14 May 2026
6 min
read

What Is PEP Screening? A Complete Guide for Banks and Fintechs

In 2016, the Monetary Authority of Singapore revoked the banking licences of Falcon Private Bank and BSI Bank — both in the same year. The proximate cause was their handling of 1MDB-linked funds. At the centre of that scandal stood Najib Razak, then Prime Minister of Malaysia and, by every applicable definition, a politically exposed person.

Here is what made 1MDB so instructive: those banks did not fail to identify Najib Razak as a PEP. His status was not hidden. He was the head of government of a sovereign nation. The failure was what came after identification — no meaningful source of wealth verification, no senior management scrutiny calibrated to the risk, and no ongoing monitoring that could have caught the pattern of transfers as they accumulated. USD 4.5 billion moved through the system. The problem was not that PEP screening did not exist. The problem was that PEP screening stopped at the checkbox.

That distinction between identifying a PEP and actually managing the risk that designation carries, is what this guide covers.

Talk to an Expert

What Is a Politically Exposed Person (PEP)?

FATF Recommendation 12 defines a PEP as a natural person who is or has been entrusted with a prominent public function. That definition is broader than most practitioners assume.

There are three categories:

Domestic PEPs hold senior positions within their own country. Government ministers, senior legislators, senior military officers, executives of state-owned enterprises, and senior judiciary members all qualify. A sitting Malaysian minister is a domestic PEP. A Philippine senator is a domestic PEP. A member of the BSP board is a domestic PEP.

Foreign PEPs hold equivalent positions in another country. An Indonesian government official is a foreign PEP from the perspective of a Singapore bank onboarding them as a client.

International organisation PEPs are senior executives of bodies such as the UN, World Bank, and IMF.

Relatives and Close Associates

This category is where most PEP screening programmes fail quietly. FATF Recommendation 12 explicitly extends the elevated risk designation to relatives and close associates (RCAs) — family members and known business associates of a PEP.

The Indonesian government official's spouse is an RCA. A business partner who shares ownership of a company with a Philippine senator is an RCA. An account held by an RCA, with no direct PEP name on it, carries the same risk elevation as the PEP's own account. A screening programme that only looks at the account holder's name will miss this entirely.

How Long Does PEP Status Last?

FATF does not set a sunset period. A former prime minister who left office last year does not automatically cease to be a PEP risk.

MAS and BNM guidance both indicate a risk-based approach with no automatic de-listing. Many APAC jurisdictions require treating former PEPs as high-risk for at least 12 months after leaving office. In practice, the risk-based approach means continuing EDD until the institution can demonstrate — and document — that the elevated risk has materially diminished.

Why PEPs Are High-Risk: The Regulatory Rationale

PEPs have access to state resources, procurement decisions, and regulatory influence. That access creates both the opportunity and, in environments with weak governance, the structural conditions for corruption-linked money laundering.

The 1MDB case demonstrated this precisely. Najib Razak's position as Prime Minister gave him effective control over a sovereign wealth fund. Funds were extracted through a network of transactions routed through accounts at Falcon Private Bank Singapore, BSI Bank Singapore, and 1MDB-linked accounts at multiple Malaysian banks. The mechanism was not sophisticated in isolation — large transfers between entities with opaque ownership, wire patterns inconsistent with stated business purpose, and inadequate documentation of source of funds. What made it possible was the combination of PEP access and institutional failure to apply the monitoring that FATF Recommendation 12 requires.

MAS revoked Falcon's licence in October 2016. BSI's licence was revoked in May of the same year. Both had processed transactions that, under any functioning ongoing monitoring programme, should have generated alerts long before the funds were moved.

FATF Recommendation 12 requires all FATF member jurisdictions to apply enhanced due diligence to PEPs. Across APAC, every major financial regulator has implemented this through binding instruments: more rigorous identification, source of funds and wealth verification, senior management or board approval, and — critically — ongoing monitoring, not just onboarding review.

The PEP Screening Process: Step by Step

Step 1: Identification at onboarding. Screen the customer's name against PEP databases at account opening. This is the minimum. It is also, for many institutions, where the process ends — which is not compliant.

Step 2: Selecting list sources. No single global PEP register exists. Governments do not publish a unified, machine-readable list of their own officials. Commercial PEP databases — World-Check, Dow Jones Risk & Compliance, ComplyAdvantage, and others — aggregate from public sources: government gazettes, parliament records, regulatory filings, and adverse media. The quality of the database determines the quality of the screening. Not all databases are equal on APAC coverage.

Step 3: Fuzzy and phonetic matching. PEP names in APAC are routinely transliterated from Arabic, Mandarin, Malay, Tagalog, or Bahasa Indonesia into Latin script. "Muhammad" has over 30 common English transliterations documented in screening literature. A system doing exact string matching will miss a match on "Mohamed" when the database entry reads "Muhammad." The minimum standard is fuzzy matching with configurable similarity thresholds — the compliance team sets the sensitivity, trading off false positives against false negatives based on the institution's risk appetite.

Step 4: Alias and AKA coverage. A single PEP entry in a quality commercial database may carry 10 to 30 aliases — formal name, preferred name, name in original script, transliterations, common abbreviations. Screening must cover all aliases, not only the primary entry.

Step 5: RCA screening. The institution must screen known family members and business associates in addition to the PEP themselves. This requires a database that explicitly links RCA relationships to PEP entries, and screening logic that applies that linkage at the match stage.

Step 6: Risk scoring. A binary PEP flag — PEP or not PEP — is not sufficient for a risk-based programme. A senior minister in a country with a Corruption Perceptions Index score in the bottom quartile presents materially different risk than a local government official in a high-CPI jurisdiction. Screening output should produce a risk score based on the PEP's role, the jurisdiction's CPI, and the nature of the relationship (direct PEP or RCA) — not just a match indicator.

ChatGPT Image May 14, 2026, 02_42_29 PM

Enhanced Due Diligence for PEPs: What Regulators Require

The table below summarises EDD requirements for PEPs across the five APAC jurisdictions where Tookitaki clients operate most frequently.

ChatGPT Image May 14, 2026, 02_57_55 PM

The common thread across all five: source of funds and wealth documentation, senior management or board approval, and enhanced ongoing monitoring. Not just enhanced onboarding. The onboarding review and the ongoing monitoring obligation are distinct requirements, and both are mandatory.

For institutions operating in the Philippines specifically, BSP Circular 706 sits alongside the country's AMLA framework. The sanctions screening obligations in the Philippines carry their own separate requirements that must be addressed in parallel with PEP screening — the two programmes are related but not interchangeable.

Ongoing Monitoring of PEPs: Where Most Programmes Break Down

PEP status is not static. A politician loses office. A state enterprise executive is newly appointed to a board. A businessman is awarded a government contract, making him an RCA of a minister. A company linked to a PEP is nationalised. Every one of those events changes the risk profile of an account, sometimes immediately.

The ongoing monitoring obligation means the institution must catch those changes — not only at annual review, but as close to real-time as the database update frequency permits.

List update frequency matters. Commercial PEP databases update continuously, adding new entries and modifying existing ones as source information changes. A batch re-screening process running on a 30-day cycle will miss PEP status changes that occurred in the intervening period. The institution that processes a transaction for a newly appointed government minister in week two of the month, having last screened at the start of the month, has a gap it cannot explain to an examiner.

Transaction monitoring is the second layer. PEP account status should be an input into the transaction monitoring system, not a separate silo. PEP accounts need calibrated scenarios — elevated sensitivity thresholds for large cash transactions, unusual international wire patterns, structuring activity. Identifying a customer as a PEP at onboarding, then running standard monitoring scenarios against their account, defeats much of the purpose of the classification. For an overview of how transaction monitoring and customer risk profiles interact, see our complete guide to transaction monitoring.

Adverse media screening is mandatory, not optional. MAS and BNM guidance both require ongoing adverse media monitoring as a component of the EDD programme for PEPs. News coverage linking a PEP to corruption allegations, enforcement action, or financial crime investigations is material information that changes the risk assessment — and must be picked up between formal review cycles, not only when the annual review is triggered.

Common Failures in PEP Screening Programmes

Six patterns appear consistently in examiner findings and enforcement actions across APAC.

Screening only at onboarding. The institution ran the check when the account was opened. Nobody re-screened when the PEP database was updated, when the customer's circumstances changed, or at any subsequent interval. This is the most common finding.

No RCA screening. The PEP's spouse holds an account. The PEP's business partner is a beneficial owner of a corporate client. Neither was linked to the PEP entry in the screening logic. The RCA relationship was not in the database configuration or was not applied consistently.

Binary flag without risk scoring. Every PEP received the same treatment — a flag, a notation, and no differentiated response based on role, jurisdiction, or exposure level. A senior minister in a country rated 20 on the CPI was processed the same way as a retired local councillor from a G7 country.

Manual re-screening processes. Someone downloaded the updated database, manually ran names against it, and filed the results in a spreadsheet. At scale, this cannot keep pace with the update frequency of commercial databases and creates an audit trail that examiners will question.

No audit trail. Examiners want to see that every customer was screened, when the screening occurred, against which version of the database, what matches were returned, and what the analyst's disposition decision was for each match. Institutions that cannot produce this log face significant difficulties in examination.

Treating identification as the endpoint. The purpose of identifying a PEP is not to decide whether to accept or reject the relationship — although that is one possible outcome. The purpose is to apply EDD and ongoing monitoring calibrated to the risk. Refusing a relationship without applying the EDD process, or accepting it without doing so, both represent programme failures.

Technology Requirements for Effective PEP Screening

A manual or partially manual PEP screening programme cannot meet the operational requirements of FATF Recommendation 12 at scale. The technology stack must address each component of the process.

Automated database ingestion. The system pulls updated PEP data directly from commercial database providers. No manual upload, no batch delay beyond what the provider's feed supports.

Fuzzy and phonetic matching with configurable thresholds. The compliance team sets the similarity threshold — not a fixed value baked into the system by the vendor. Institutions serving APAC clients need matching logic calibrated for Southeast Asian name transliterations, which present different challenges than Western name matching.

RCA relationship mapping. The match logic applies RCA linkages from the database to customers who are not themselves PEPs, flagging accounts where a beneficial owner, signatory, or counterparty is an RCA of a listed PEP.

Risk scoring output. The screening event produces a risk score, not just a match indicator. The score reflects the PEP's role, the jurisdiction's CPI ranking, and the relationship type (direct PEP, family member, or business associate).

Full audit trail. Every screening event is logged with a timestamp, the database version used, the match score, the analyst's decision, and the rationale documented in the system. This log is the institution's primary defence in an examination or enforcement inquiry.

Integration with transaction monitoring. PEP status feeds into the transaction monitoring configuration. A match on a counterparty in an international wire transfer triggers both a screening alert and a monitoring review. PEP account flags elevate the sensitivity of transaction monitoring scenarios. The two systems operate as components of a single risk management programme, not independent tools producing separate outputs. The Transaction Monitoring Software Buyer's Guide covers the evaluation criteria for the broader platform, including how screening and monitoring integration should be assessed.

PEP Screening in FinCense

FinCense covers PEP screening as part of its integrated AML platform. It is not a standalone screening module bolted to a separate transaction monitoring system — the PEP identification, risk scoring, and monitoring inputs operate together within the same platform.

The system comes pre-configured with APAC-relevant PEP databases, with fuzzy matching calibrated for the transliteration patterns common in Southeast Asian names. Every screening event is logged in a format that MAS, BNM, BSP, and AUSTRAC examiners can follow — timestamp, database version, match score, disposition, rationale.

When a customer's PEP status changes — a new appointment, a newly documented RCA relationship, an adverse media hit — the platform reflects that change in the monitoring configuration, not only in the customer record.

Book a demo to see FinCense's PEP screening running against APAC-specific scenarios.

 What Is PEP Screening? A Complete Guide for Banks and Fintechs