8 Years in AML: Why Machine Learning Beats Rules in Bangladesh

Photo by New Material on Unsplash

I still remember the night we detected a structuring ring that had been flying under the radar for months, moving about BDT 50 million through bKash accounts in increments just below the BDT 100,000 MFS threshold. It was a wake-up call - our rule-based system had failed us.

So, what went wrong? The truth is, standard approaches often fail in Bangladesh due to the unique nature of our financial landscape. The sheer volume of small transactions, the preference for mobile financial services (MFS) like bKash and Nagad, and the cat-and-mouse game with money launderers mean that static rules just can't keep up.

The Hidden Problem

Rules are brittle. They're based on historical patterns and are quickly outdated by new money laundering tactics. In a country where the economy is largely cash-based and informally structured, machine learning (ML) offers a way out. By analyzing patterns in transaction data, ML can identify suspicious activity that rules might miss.

Technical Breakdown & Logic Flow

Here’s how it works: we start by collecting and preprocessing transaction data, including sender and recipient information, transaction amounts, and timestamps. We then train an ML model to recognize patterns indicative of money laundering, such as frequent small transactions or unusual patterns of activity.

One of the models we've found particularly effective is the Isolation Forest. It works by identifying data points that are farthest from the rest - in other words, the most abnormal transactions. This approach is especially useful in AML because it can detect anomalies without needing labeled data.

Python Implementation

import pandas as pd
from sklearn.ensemble import IsolationForest
# Load transaction data
data = pd.read_csv('transactions.csv')
# Preprocess data (e.g., convert timestamps to datetime format)
data['timestamp'] = pd.to_datetime(data['timestamp'])
# Train Isolation Forest model
model = IsolationForest(n_estimators=100, random_state=42)
model.fit(data[['amount', 'sender_id', 'recipient_id']])
# Predict anomalies
predictions = model.predict(data[['amount', 'sender_id', 'recipient_id']])
# Identify transactions with high anomaly scores
anomalous_transactions = data[predictions == -1]

We chose this approach over more traditional methods like decision trees or logistic regression because it handles high-dimensional data well and doesn’t require a large amount of labeled training data.

Local Application

In the context of Bangladesh, this approach aligns well with BFIU guidelines, which emphasize the importance of monitoring transactions below the BDT 100,000 threshold. By integrating ML into our AML systems, we can better detect and prevent money laundering, even when it involves small, frequent transactions.

BFIU guidelines state that all transactions, regardless of amount, should be subject to monitoring. This is where ML really shines - it can analyze vast amounts of data quickly and accurately, freeing up human analysts to focus on higher-risk cases.

Common Pitfalls & Edge Cases

One of the biggest pitfalls we've encountered is overfitting. When a model is too closely fit to the training data, it fails to generalize well to new, unseen data. This can result in a high number of false positives or, worse, false negatives.

To mitigate this, we use techniques like cross-validation and ensure that our training data is diverse and representative of real-world scenarios.

Counterintuitive Insight

One thing that’s surprised us is how often seemingly insignificant transactions are actually part of a larger money laundering scheme. By analyzing patterns and anomalies, ML can uncover these schemes in a way that rule-based systems often can’t.

Conclusion & CTA

So, what’s the takeaway? In Bangladesh, ML-based AML is the future. It offers a more effective, more efficient way to detect and prevent money laundering, even in the face of rapidly evolving tactics.

What’s the weirdest transaction pattern you’ve seen? Have you implemented ML in your AML systems? Share your experiences in the comments below. And if you’re looking for more resources on how to get started with ML in AML, check out our other articles on aitipseveryday.com.