Machine learning in banking: 5x reduction in fraud risk through transaction anomaly detection

Web

Innowise has developed an ML-powered system that monitors digital transactions and detects suspicious or fraudulent behavior.

Customer

Industry

Banking, Finance

Region

Client since

2021

Our customer is a large commercial bank with a network of branches across the country, offering deposits, loans, and other services.

Detailed information about the client cannot be disclosed under the provisions of the NDA.

Challenge: Increased risk of financial fraud and account takeovers that a key American bank encountered

The primary focus of every bank or financial institution is the satisfaction and security of its account holders. As part of their daily operations, these institutions handle client accounts, oversee investments, maintain adequate liquidity, and perform other functions.

Unfortunately, the banking industry is currently facing a significant threat from suspicious and malicious activities that not only endanger customers but also the industry as a whole. Until recently, banks have used mostly manual rules-based systems, but with fraudsters turning more sophisticated, these systems are rapidly becoming inefficient.

One of the key American banks approached Innowise, seeking an effective machine learning in banking solution to detect and combat financial fraud. As the client expanded and the number of transactions increased, the bank periodically faced malicious activities that endangered its safety and reputation. Surely, our customer had an anti-money laundering system that prevented criminal profits from being camouflaged and incorporated into the financial system. However, it lacked accuracy, showing a high number of false positives and leaving room for account takeovers and payments fraud.

Solution: ML-based solution to detect transaction anomalies and prevent fraudulent activities

Innowise recommended integrating an ML-powered extension into the banking ecosystem to analyze big data volumes and safeguard funds from malicious activities. Account holders' transactions are analyzed and alerted if any uncharacteristic, suspicious, or fraudulent behavior is detected. Using deep learning in fintech algorithms, our project team analyzed vast amounts of data to detect abnormalities that could indicate fraud.

Data aggregation

As the first step, Innowise gathered and consolidated all banking-related data, including the user’s identities, locations, payment methods, transaction histories, and other relevant factors.

Abnormal patterns elaboration

We derived distinctive patterns, such as unusually high transaction amounts or splitting transactions to avoid automatic tax reporting. These patterns enable ML algorithms to distinguish fraudulent activities from regular banking operations and trigger appropriate actions when a risky pattern emerges. Based on that, transactions are categorized as either “good” (legitimate) or “bad” (fraudulent).

Overall, Innowise accessed a vast dataset (e.g., dozens of millions samples based on neural network, transactional and historical data), highly effective in identifying patterns and spotting abnormal behaviour that deviates from the norm. We selected the most critical features by comparing expectations with actual data and recursive feature elimination techniques. Also, our team identified missing data labels, providing techniques for better fraud detection.

Model training

Since ruled-based patterns highlight explicit fraud cases, our ML specialists developed algorithms detecting unusual or unfamiliar circumstances where conventional algorithms fail. As a result, the extension can make predictions even without sufficient data, relying on machine learning training techniques. Thus, our solution utilizes embedded representations instead of classical aggregated features to process transactions.

Full-fledged ML model

When a threat is identified, the system transmits this data in real time to the administrator, who can freeze or cancel operations until further investigation. Depending on fraud probability, there are three possible outcomes:

If the probability of fraud is below 5%, the transaction is approved;
If the probability of fraud is between 6% and 70%, an extra check such as a one-time SMS code, fingerprint, or secret question is required;
If the probability of fraud exceeds 80%, the transaction is rejected, requiring manual processing and analysis.

Furthermore, we ensured comprehensive ML model explainability tools, helping to understand prediction results and providing a seamless user experience.

Technologies

Back-end

Python, Scala

Front-end

HTML, CSS, React, MUI

Data engineering

Apache Flink, Redis Feast, Apache Hive, Apache Airflow, Apache Kafka, Apache Spark

Machine Learning/Data Science

Apache Spark MLLib, Scikit-learn, LightGBM, XGBoost, Hyperopt, PySpark, Numpy, Pandas, Scipy

MLOps

DVC, MLFlow, Comet

Environment

Docker, Docker Compose, Kubernetes, Jenkins

Process

Initially, our project team developed business and technical requirements to meet customer expectations. Throughout the project, our business analyst maintained close contact with the client’s banking consultants to gain a deeper understanding of the client’s business and take full advantage of machine learning in financial services.

As for the ML solution, the most challenging aspect was achieving optimal metrics for users with different transaction histories. Our model was effective for account holders with substantial transactional history but ineffective for new users with a lack of historical data. Such users were treated as inactive accounts containing only identity information and no transaction history. Though this assumption eliminates the benefit of having complete user data, it, nonetheless, yields reasonably stable training results for the ML model.

After discussing the issue, we investigated “few-shot learning” methods that could enhance our metrics. We conducted a proof of concept but it did not result in the substantial improvements we anticipated. Therefore, our project team continued to enhance the platform and delve deeper into our client’s business domain. This enabled us to design features that significantly impacted the “few-shot learning” model, ensuring accurate prediction results for the banking machine learning project.

Our team followed the Scrum methodology, with three-week sprints throughout the project. We held regular meet-ups with the team via Microsoft Teams to monitor the progress of the project and accommodate any changes to the scope. As of now, the project has been successfully completed.

Team

Project Manager

Business Analyst

Front-End Developers

Back-End Developers

ML Engineers

Data Engineers

UI/UX Designer

QA Engineer

Results: x2.4 faster processing speed with fewer false positives and a reduced risk of unspotted fraud

Innowise has developed an advanced ML-powered extension to detect suspicious or fraudulent activities and take proactive measures based on that. We ensured impeccable security and eliminated the risk of breaches and financial crimes. Overall, machine learning in banking and finance implementation brought the following benefits:

higher speed

Machine learning algorithms rapidly analyze vast amounts of data. With the pace and volume of banking transactions growing, our machine-learning platform scrutinizes new information continuously.

improved efficiency

Machine learning algorithms can execute repetitive operations and instantly spot subtle alterations in patterns. Our ML solution examines hundreds of thousands of payments per second, streamlining the entire transaction process.

accuracy

In this project, we used machine learning algorithms that can be trained to identify patterns within seemingly trivial data. They recognize subtle or non-intuitive patterns that would be challenging, if not impossible, for humans to discern. This enhances the precision of fraud detection, resulting in fewer false positives and a reduced risk of unspotted fraud.

Project duration

June 2021 - December 2022

99.3%

accuracy in reducing fraud

x2.4

faster processing speed

Related cases

Contact us!

Book a call or fill out the form below and we’ll get back to you once we’ve processed your request.

What happens next?

Having received and processed your request, we will get back to you shortly to detail your project needs and sign an NDA to ensure the confidentiality of information.

After examining requirements, our analysts and developers devise a project proposal with the scope of works, team size, time, and cost estimates.

We arrange a meeting with you to discuss the offer and come to an agreement.

We sign a contract and start working on your project as quickly as possible.

Machine learning in banking: 5x reduction in fraud risk through transaction anomaly detection

Customer

Challenge: Increased risk of financial fraud and account takeovers that a key American bank encountered

Solution: ML-based solution to detect transaction anomalies and prevent fraudulent activities

Technologies

Process

Team

Results: x2.4 faster processing speed with fewer false positives and a reduced risk of unspotted fraud

Related cases

Contact us!

What happens next?

Let^●s develop
software together!

Make another request

Close

Machine learning in banking: 5x reduction in fraud risk through transaction anomaly detection

Customer

Challenge: Increased risk of financial fraud and account takeovers that a key American bank encountered

Solution: ML-based solution to detect transaction anomalies and prevent fraudulent activities

Technologies

Process

Team

Results: x2.4 faster processing speed with fewer false positives and a reduced risk of unspotted fraud

Related cases

Contact us!

What happens next?

Let●s develop software together!

Subscribe to our newsletter

Make another request

Close

Let^●s develop
software together!

Subscribe
to our newsletter