Building Robust AML Models: Advanced Techniques for Effective Transaction Monitoring

Welcome to the Building Robust AML Models repository! This project showcases a comprehensive approach to optimizing Anti-Money Laundering (AML) transaction monitoring for the scenario: Large Value Deposits. The project describes the market standard tuning methodologies, explains the inefficiencies of these methodologies to tackle the generation of False Negatives (Not detecting all Suspicious alerts) and how the scenarios and tuning methodologies should be optimized for developing robust AML Compliance Models. The repository is structured to provide valuable insights for technical recruiters, interviewers, and anyone interested in data science, finance, and AML compliance.

Explore the Project

You can explore the project and its components in the following ways:

1. Project Presentation: A comprehensive presentation that covers the problem statement, methodology, results, key insights, and action items from the analysis. This presentation is ideal for understanding the project's scope and impact.

Table of Contents

• Overview
• Project Structure
• Key Steps and Methodology
• Key Findings
• Conclusion
• How to use
• Skills Demonstrated
• Author
• Connect with Me

Overview

In this project, I have applied advanced statistical techniques to optimize AML transaction monitoring for large value deposits. The objective was to develop a market-standard solution for the scenario: Large Value Deposits that enhances the detection of suspicious activities while minimizing false positives, reflecting the stringent compliance requirements of financial institutions.

Purpose

The primary goal of this project is to demonstrate my ability to interpret and analyze data to make informed financial crimes decisions. By tuning the monitoring thresholds effectively, I aim to maximize the detection of true suspicious activities, reduce generation of False Negatives and reduce operational costs associated with false positives.

Relevance

This project showcases a methodology that is widely used in the banking sector. But this method is inefficient due to large number of False Positives, and presence of False Negatives. I’ve learnt from my experiences how to develop efficient AML models - Rule-based and ML-based (Segment the customers, products etc.) models. I play an instrumental role in bridging the gap between the inefficient market standard and building a state-of-the-art in-house AML models. This presentation underscores my readiness to contribute to Morgan Stanley’s Global Financial Crimes team by leveraging my skills in data analysis and AML transaction monitoring.

Project Structure

1. Data:

   • accounts.csv: Contains account-level information, including the target variable IS_FRAUD.
   • transactions.csv: Contains transaction-level details such as sender and receiver account IDs, transaction amount (TX_AMOUNT), and timestamps.

2. Output Files:

   • DQ_report_account_data.xls: Data Quality Report for the account data.
   • DQ_report_transactions_data.xls: Data Quality Report for the transaction data.
   • threshold_summary_file.xls: Summary file containing threshold values and corresponding evaluation metrics for each percentile.

3. Scripts and Notebooks:

   • AML_Tuning.ipynb: Jupyter Notebook containing the detailed scripts for data processing, parameter tuning, and evaluation.
   • dq_report.py: Python script used to generate the Data Quality Report.

4. Presentation:

   • Building Robust AML Models Advanced Techniques for Effective Transaction Monitoring.pptx: PowerPoint presentation summarizing the project, findings, and recommendations.

Key Steps and Methodology

1. Data Quality Report (DQ Report)

• Generated a comprehensive Data Quality Report for both the account and transaction datasets.
• Analyzed missing values, unique values, data types, and distribution of numerical variables to identify outliers and skewness.

2. Parameter Tuning

• Focused on the DEPOSIT_AMT parameter due to its critical role in identifying large value transactions.
• Generated a threshold summary file encompassing percentile distributions from 1 to 99, with calculated thresholds and corresponding evaluation metrics for each percentile.

3. Threshold Evaluation Metrics

• Evaluated multiple metrics to determine the optimal threshold for DEPOSIT_AMT:
  • True Positive Rate (Recall)
  • False Positive Rate
  • Precision
  • F1 Score
  • Youden’s J Statistics

4. Optimum Threshold Selection

• Determined the best threshold using Youden's J Statistic and F1 Score.
• Visualized confusion matrices to compare the effectiveness of each threshold on both training and test datasets.
• Identified that using recall as the evaluation metric minimizes false negatives, a crucial factor in AML monitoring.

5. Methodology Optimization

• Optimized the tuning methodology using recall to ensure maximum detection of suspicious activities.
• Applied the identified thresholds to both training and test datasets to validate their effectiveness.

Key Findings

• J Statistics: Balances sensitivity and specificity but may not always prioritize the highest recall.
• F1 Score: Balances precision and recall, providing a comprehensive view of performance.
• Recall: Maximizes detection of suspicious activities but may lead to a higher number of false positives.

Conclusion

This project demonstrates my expertise in developing and tuning AML models, both rule-based and machine learning-based. By leveraging advanced data analysis techniques, I have optimized transaction monitoring for large value deposits, ensuring high detection efficiency while minimizing operational costs. Different scenario designs, tuning methodologies, and metrics must be used and should be tailor-made according to each use case, bank product, etc. Additionally, building a good post-detection solution is crucial to further reduce false positives. I have the skillsets to achieve all this.

How to Use

1. Clone the repository:

   git clone https://github.com/ds-souvik/Threshold-Tuning-for-Anti-Money-Laundering-AML-Scenario.git
   cd Threshold-Tuning-for-Anti-Money-Laundering-AML-Scenario

2. Set up the environment:

   • Ensure you have Python and Jupyter Notebook installed. You can use the following commands to set up a virtual environment:

   python -m venv aml_env
   source aml_env/bin/activate  # On Windows use `aml_env\Scripts\activate`
   pip install -r requirements.txt

3. Run the Jupyter Notebook:

   jupyter notebook AML_Tuning.ipynb

4. Generate Data Quality Reports:

   • Execute the dq_report.py script to generate data quality reports for the account and transaction datasets:

   python dq_report.py

5. Explore the Output Files:

   • Review the generated data quality reports and threshold summary file in the Output_files directory.

6. Analyze and Visualize Results:

   • Follow the steps in the Jupyter Notebook to analyze and visualize the results of the parameter tuning and threshold evaluation metrics.

7. Review the Presentation:

   • Open Building Robust AML Models Advanced Techniques for Effective Transaction Monitoring.pptx to understand the project summary, methodology, and findings.

8. Customize and Extend:

   • Modify the scripts and notebook as needed to adapt the methodology to different AML scenarios or datasets.

Skills Demonstrated

Technical Skills:

• Data Analysis: Proficient in using Python for data preprocessing, cleaning, and analysis. Implemented advanced statistical techniques to optimize AML transaction monitoring.
• AML Threshold Tuning and Statistical Analysis: Applied various evaluation metrics including True Positive Rate (Recall), False Positive Rate, Precision, F1 Score, and Youden’s J Statistics to determine optimal thresholds.
• Data Quality Assessment: Generated comprehensive Data Quality Reports to identify missing values, unique values, outliers, and skewness in the datasets.
• Data Visualization: Created and visualized confusion matrices to compare the effectiveness of different thresholds on both training and test datasets.
• Programming: Developed reusable scripts for data processing, threshold tuning, and evaluation. Proficient in Python, including libraries such as pandas, numpy, matplotlib, seaborn, and scikit-learn.
• Jupyter Notebooks: Documented the entire analysis process, including data preprocessing, parameter tuning, and evaluation, in a clear and reproducible manner using Jupyter Notebooks.

Non-Technical Skills:

• Problem Solving: Identified inefficiencies in market-standard AML tuning methodologies and developed optimized solutions tailored to specific AML scenarios.
• Attention to Detail: Ensured accuracy in data analysis and model tuning by meticulously evaluating multiple metrics and thresholds.
• Communication: Effectively documented and presented project findings, methodologies, and recommendations in a professional manner.
• Project Management: Managed the project workflow, from data acquisition and preprocessing to model tuning and evaluation, ensuring timely completion and high-quality deliverables.
• Industry Knowledge: Demonstrated understanding of AML compliance requirements and financial crime prevention strategies. Leveraged industry experience to develop robust AML models.
• Adaptability: Showcased ability to quickly learn and implement new techniques and methodologies to improve AML transaction monitoring.

Author

Souvik Ganguly

---

Connect with me: If you would like to connect with me, feel free to reach out via LinkedIn or email me at souvik.ganguly.ds@gmail.com.

---

If you liked the project, consider putting a star?