Introduction
The rising threat of ATO Fraud
In the financial sector, a major challenge is the increasing threat of Account Takeover (ATO) fraud: identity theft falling under third-party fraud. This type of fraud typically remains undetected until after financial damage has been inflicted, emphasizing the industry’s urgent need for advanced detection methods.
Institutions traditionally rely on reactive models that only flag fraudulent activities post-loss, affecting customer trust and satisfaction. A proactive approach using machine learning models is reshaping fraud management strategies while prioritizing asset protection and customer confidence.
Challenge
Proactive fraud prevention
A leading U.S. bank ranked among the top 20 faced severe losses due to ATO fraud. Without a predictive system, the Fraud team could only react after losses were incurred, impacting customer satisfaction due to the frequency of such incidents.
Reactive fraud detection
The bank’s existing approach allowed ATO fraud detection only after monetary losses had occurred. This reactive model led to significant financial losses and affected customer trust and satisfaction.
Near real-time fraud prediction
There was no predictive mechanism to identify potential ATO fraud cases promptly. The Fraud team urgently required a near real-time solution to anticipate and prevent monetary losses.
Solution
Data-driven discovery and predictive modeling
We discovered that patterns in non-monetary batch and streaming data from various sources are vital for predicting ATO fraud. This data, fed into a predictive engine, generated a likelihood score for each customer’s risk of experiencing an account takeover. Our collaborative sessions with the Fraud team and an extensive data discovery phase led to identifying historical ATO cases, forming the basis for the model’s target variable.
Using machine learning, the model analyzed account activities like login frequencies, device and IP usage, and changes in security details. This comprehensive approach enabled the Fraud team to devise targeted, preemptive strategies against ATO fraud, significantly enhancing fraud mitigation efforts.
What we provided:A near real-time fraud detection system
Our advanced model tackles ATO fraud using a highly imbalanced dataset and features near real-time scoring on the Google Cloud Platform (GCP). This generates customer level ATO probability scores to indicate a potential account take over scenario to the Fraud team. A Looker dashboard was also developed for the Data Science team to continuously monitor data and model performance.
We built a scalable and robust solution on GCP. GCP services like BigQuery, Dataflow, Cloud Composer, Cloud Storage, Cloud build, Vertex AI Custom Training, Source Code Repository, Jupyter Notebooks & Looker are used in the solution. Both the model training and scoring pipelines were implemented along that enables near real time scoring of the customers activity to predict risk of account take over. We also implemented a Looker dashboard to monitor both data and model drift.
Outcome
Long-term strategic advantages
The immediate impact:
| Near real-time fraud predictions | Blueprint for future projects | Empowerment through documentation |
Enabled predictions of account takeover frauds in near real-time from non-monetary activities. The final solution implementation was near real-time (every 5 minutes batch). |
The project is a model for future initiatives, with end-to-end implementation on GCP and a solid foundation for MLOps infrastructure. |
Comprehensive model risk management and data lineage documents were provided, complete with a codebase, facilitating ongoing maintenance and future retraining of the model by the client. |
The long-term benefits:
Once the Fraud team fully integrates the solution, its outputs should enable the client to implement targeted interventions at the individual customer level, substantially mitigating financial and reputational risks. This shift from a reactive post-loss investigation to proactive fraud detection is expected to realize significant savings.