Mobile Transaction Forecasting and Anomaly Detection
"Sparkling Water combines the fast, scalable ML algorithms of H2O, the H2O flow UI, Scala, and Python, with the capabilities of Apache Spark. This allowed for really rapid prototyping and ad-hoc experimentation."Rahul Gupta Data Scientist, Capital One
- Advanced Analytics
- Explainable AI
Overview of the Challenge
To keep its mobile app up and running at all times, Capital One has a dedicated technology operations group, which monitors all of the bank’s critical systems and platforms. Based on company policies, the group configures a number of alerts that are triggered at specific thresholds. While some alerts are straightforward and easy to set up, such as when a certain number of failures occur within a specified timeframe, others, including volume alerts, are notoriously tricky to calculate. A drop in the volume of transactions, or a higher than expected volume can be indicative of a problem, however finding an effective method of creating an alert is far from simple. “Volume is hard to detect, measure, and alert on, says Donald Gennetten, a data engineer at Capital One. “You’ve got volumes that change overtime; you have factors such as the time of day, day of week, and other seasonal elements. When you try to do calculations on volume anomalies, you quickly realize that you have too many distinct thresholds to calculate and maintain.” Gennetten and his team needed a solution that could scale and didn’t require a lot of coding, development, and oversight to manage. “This was the perfect situation for us to leverage machine learning,” he concludes.