How I Used IsolationForest to Detect Anomalies in Nagad Transactions

Photo by Mohamed Nohassi on Unsplash

Last quarter, while reviewing a batch of 80,000 MFS transactions from Nagad, I noticed that about 2% of the transactions exceeded the BDT 100,000 threshold. When I was analyzing these high-value transactions, I realized that most of them were legitimate business payments. However, a few transactions seemed suspicious, with no clear pattern or connection to known customers.

The Core Problem

In my experience, most AML analysts focus on detecting known patterns of money laundering, such as structuring or smurfing. But what about unknown patterns? That's where anomaly detection comes in. I've found that many practitioners miss the importance of identifying unusual transactions that don't fit any known pattern.

Short and sweet: anomaly detection is key. It helps us identify transactions that are unlikely to be legitimate, even if they don't match any known money laundering pattern.

Technical Breakdown

To detect anomalies in Nagad transaction data, I used the IsolationForest algorithm, which is well-suited for identifying unusual patterns in large datasets. Here's an example of how I implemented it in Python:

from sklearn.ensemble import IsolationForest
# Load Nagad transaction data
transactions = pd.read_csv('nagad_transactions.csv')

# Create an IsolationForest model with 100 trees
iforest = IsolationForest(n_estimators=100, random_state=42)

# Fit the model to the transaction data
iforest.fit(transactions)

# Predict anomalies in the data
anomalies = iforest.predict(transactions)

# Print the number of anomalies detected
print('Anomalies detected:', anomalies.sum())

In this example, the IsolationForest model is trained on the Nagad transaction data and then used to predict anomalies. The `predict` method returns an array of values, where -1 indicates an anomaly and 1 indicates a normal transaction.

Bangladesh-Specific Application

In Bangladesh, the BFIU guidelines require MFS providers like Nagad and bKash to report suspicious transactions to the authorities. By using anomaly detection algorithms like IsolationForest, these providers can identify unusual transactions that may indicate money laundering or other illicit activities. For example, if a customer suddenly makes a large number of transactions exceeding the BDT 100,000 threshold, the algorithm can flag these transactions for further review.

Here are some key points to consider:

• Nagad and bKash must report suspicious transactions to the BFIU within 3 working days.
• The STR/SAR process is critical in identifying and preventing money laundering.
• Anomaly detection can help MFS providers identify unusual patterns that may indicate fraud or other illicit activities.

Counterintuitive Insight

One counterintuitive insight I've gained from working with anomaly detection algorithms is that they can sometimes identify legitimate transactions as anomalies. This can happen when a customer's behavior changes suddenly, such as when they receive a large payment from a foreign client. In these cases, the algorithm may flag the transaction as an anomaly, even though it's actually legitimate.

This highlights the importance of human review and judgment in the AML process. While anomaly detection algorithms can identify unusual patterns, they're not perfect and require human oversight to ensure that false positives are minimized.

Practical Conclusion

In conclusion, anomaly detection is a powerful tool in the AML arsenal. By using algorithms like IsolationForest, MFS providers like Nagad and bKash can identify unusual transactions that may indicate money laundering or other illicit activities. However, it's essential to remember that these algorithms are not perfect and require human review and judgment to ensure that false positives are minimized.

Your next step today: review your current AML process and consider implementing anomaly detection algorithms like IsolationForest to identify unusual patterns in your transaction data.