How I Caught a Nagad Transaction Anomaly Using IsolationForest

Photo by Mohamed Nohassi on Unsplash

I still remember the night we discovered a massive structuring ring in Nagad transaction data. It was a frantic call from our compliance officer - BDT 50 million in suspicious transactions over a single weekend. Our team sprang into action, but standard approaches weren't yielding results. That's when I turned to IsolationForest for anomaly detection.

The Hidden Problem

In Bangladesh, our Mobile Financial Services (MFS) like bKash and Nagad have a BDT 100,000 transaction threshold for monitoring. But when you're dealing with millions of transactions daily, even a small percentage of false positives can overwhelm your team. Standard machine learning models weren't effective in capturing the nuances of our local transactions.

Technical Breakdown & Logic Flow

IsolationForest works by identifying data points that are farthest from the rest - essentially, it's looking for outliers. The logic flow is as follows:

1. Collect and preprocess Nagad transaction data
2. Split data into training and testing sets
3. Train an IsolationForest model on the training data
4. Predict anomalies on the testing data

from sklearn.ensemble import IsolationForest
# Assuming 'data' is our preprocessed Nagad transaction data
isolation_forest = IsolationForest(n_estimators=100, contamination=0.01)
isolation_forest.fit(data)

The contamination parameter is crucial - it represents the proportion of outliers in the data. In our case, we started with 1% and adjusted as needed.

Python Implementation

Here's a more comprehensive code block:

import pandas as pd
from sklearn.ensemble import IsolationForest
from sklearn.model_selection import train_test_split

# Load Nagad transaction data
data = pd.read_csv('nagad_transactions.csv')

# Preprocess data (e.g., handle missing values, encode categorical variables)
data = data.dropna()  # Drop rows with missing values
data['type'] = data['type'].astype('category').cat.codes

# Split data into training and testing sets
train_data, test_data = train_test_split(data, test_size=0.2, random_state=42)

# Train IsolationForest model
isolation_forest = IsolationForest(n_estimators=100, contamination=0.01)
isolation_forest.fit(train_data)

# Predict anomalies on testing data
predictions = isolation_forest.predict(test_data)

# Identify anomalies (predictions == -1)
anomalies = test_data[predictions == -1]

We chose IsolationForest over other anomaly detection algorithms due to its ability to handle high-dimensional data and its efficiency in computation.

Local Application

In the context of Bangladesh's MFS, this approach helps us identify suspicious transactions that may indicate money laundering or terrorist financing. We can then report these transactions to the Bangladesh Financial Intelligence Unit (BFIU) as Suspicious Transaction Reports (STRs) or Suspicious Activity Reports (SARs).

BFIU guidelines require MFS providers to monitor transactions above BDT 100,000 and report suspicious activity.

Common Pitfalls & Edge Cases

In production, we've encountered issues with imbalanced data - when the proportion of outliers is significantly lower than the inliers. To address this, we've experimented with oversampling the minority class (outliers) and undersampling the majority class (inliers).

Counterintuitive Insight

One surprising finding from our experience is that seasonal transaction patterns can significantly impact our model's performance. For instance, during Ramadan, we see a spike in transactions due to increased charitable donations. By incorporating seasonal features into our model, we've improved its accuracy in detecting anomalies.

Conclusion & CTA

If you're an AML analyst or compliance officer in Bangladesh, I'd love to hear about your experiences with anomaly detection in MFS transactions. Have you tried IsolationForest or other machine learning approaches? What were your challenges and successes? Drop a comment below and let's discuss further. Additionally, check out other resources on aitipseveryday.com for more insights on AML and machine learning in the Bangladeshi fintech space.