Optimizing Securities Lending Transactions for a Multinational Bank
To manage the uncertainty in demand and supply of securities required to cover short sales, a leading multinational bank implemented C3.ai Securities Lending Optimization.
Banks face significant uncertainty when assessing a client’s request to borrow securities. The bank must quickly evaluate both the quantity of securities the borrower is expected to trade and whether sufficient securities are available to lend. Unfortunately, neither quantity is knowable.
Traditional rules-based approaches for approving client requests are only sufficient to automatically approve the most simple, low-margin lending opportunities. Members of the Securities Lending sales desk are forced to make rapid decisions based on little information and in the face of high uncertainty. As a result, the bank failed to approve many highly profitable, executable opportunities on a daily basis. Further, reliance on manual review resulted in slow inquiry response times and a distracted sales desk, making the bank less competitive with their hedge fund clientele.
About the Multinational Bank
- $110 billion Annual Revenue
- 46 million customers
- Over 200,000 employees
- $2.3 trillion Assets Under Management
- 16 weeks from project start to production-ready application
- 2 years of historical data from 25 data sources
- 400+ time-based machine learning features
- 100+ million predictions across 2 years
- 98% reduction in manual review
- 40% increase in profitable “hard-to-borrow” requests approved
- $25 million increase in annual revenue
Accurately Predicting Variability
Over the course of the 16-week project, C3.ai worked with the customer to integrate relevant data and configure scalable, production-ready machine learning algorithms to predict upper bounds of short sale quantity for each client request.
Data from 25 sources were unified and federated into more than 50 C3.ai models representing timed relationships among customers, securities, requests, transactions, inventory, exchange trades, earnings, and corporate actions.
The unified data image was used to configure over 400 time-based machine learning features. These features represent signals from customers, securities, and market behavior that are predictive of short sale quantity for each request.
An interpretable machine learning model was configured and trained for each client to predict the upper and lower bounds of execution quantity given the client request. The model was used to generate predictions for more than 100 million client requests.
Finally, the team back tested the predictions against historical locates and measured a 40% increase in locate approvals with an expected revenue increase of $25 million per year.
Revenue increase in hard-to-borrow securities
Increase in annual revenue driven by more transactions