Machine learning helps reduce cash reserves by $40M | Fusion Alliance

Managing bank and credit union reserve cash is a complex exercise: manage it too tightly and your institution may be subject to high-interest Federal Reserve borrowing fees. Manage it too loosely and your firm may lose out on substantial interest revenue from parked cash.

Our client, a wholesale financial services provider to hundreds of credit unions in the U.S., traditionally kept a large volume of cash in reserve to account for member credit union activity. Since these credit unions conduct business autonomously, the organization was constantly challenged to predict members’ cash reserves without any direct control or visibility. It was time to explore opportunities to apply advanced analytics to predict member activity and drive better returns on reserve cash, and that’s what led to their partnership with the Fusion Alliance team.

ON-DEMAND BANKING WEBINAR

Learn how to turn data into insights that drive cross-sell revenue

While this client was unable to directly influence credit union spending and borrowing, they possessed one critical element – decades of financial transaction data to support the cash reserve engagement. Company leaders understood there were patterns in the member credit union data based on calendar milestones (payroll activity, mortgage pay activity, etc.) but needed help identifying these regularities in the noise across hundreds of credit unions and billions in cash.

This project explored 18 years of historical cash data to predict the next 60 business days of member activity, in aggregate and by cash account. The initiative additionally provided a discrete view for the investment desk to simulate cash and borrowing needs to effectively partner with finance.

Ultimately, the machine learning algorithm that needed to be selected would need to favor recency of history but still account for the entire body of transactions. To accomplish this, our team:

• Landed and cleaned data in the company’s Azure Cloud
• Accumulated success metrics on a variety of algorithms to achieve the desired liquidity aims for the organization
• Ultimately, selected a long short-term memory recurrent neural network

Once we achieved the desired metrics for cash management, our team:

• Developed an analytical website solution that:
  • Allowed the company’s finance team to feed new data
  • Exposed long-term analytics with the liquidity for the investment team to effectively manage bank cash in the big picture
  • Secured the environment according to bank best practices
• Developed a weekly retraining process to keep LSTM models current
• Integrated the solution with a machine learning web service hosted in Azure