Trading-limit prediction & lock-out prevention

It all started when

Citi Equities clients had trading limits — in USD, at both an individual and an account level. Once a limit was breached, the client was locked out and had to go through a dedicated customer-service channel for manual intervention. The volume of these lock-outs was steadily increasing, and each one meant friction for the client and cost for the desk.

What Monica built

A predictive ensemble that flagged at-risk clients before they breached a limit — but getting there ran through a root-cause analysis and a first recommendation that leadership rejected.

The root-cause analysis surfaced two things: the same ~20% of clients repeatedly breached their limits, and the real driver was the account-level limit — which clients weren't proactively made aware of — rather than the conservative individual limit, a deliberate internal control.

Monica's initial recommendation was to raise awareness of both limit types and replace individual limits with threshold-based notifications at 60% and 80% of the account limit. Leadership rejected it: high-frequency, low-latency trading made check-based notification insufficient, and the individual limits were a deliberate, risk-averse control they wanted to keep.

So she pivoted to a predictive approach. After trialling a series of candidate models, she settled on an ensemble of LSTM, Random Forest, and regression. The LSTM was trained as a per-client model for individualised pattern detection, while Random Forest and regression were trained on more general, cross-client data.

Monica ran the workstream as both product manager and developer:

• Led the primary analysis, pitched the solution, and used client-servicing complaint data to define the stakeholder problem.
• Kept senior Equities leadership informed throughout via direct discussions.
• Managed a team of 5 interns and personally developed the Random Forest model instance.

Impact

• Prevented 80% of client lock-out instances.
• Reduced client complaints requiring servicing by 55%.

Major Takeaways

• A rejected recommendation can be productive — leadership pushback on the awareness-based fix forced a shift to a genuinely predictive solution that turned out to be stronger.
• Owning a production ML model (Random Forest) while running the PM workstream meant bridging both roles on one project.
• Trial-and-error model selection showed a per-client vs. general-data split (LSTM vs. Random Forest/regression) captured individual behaviour better than any single general model.
• The real constraint was balancing a risk-averse, low-latency trading environment against the ideal customer experience.