I am working on applying AI/ML techniques to finance, specifically for account reconciliation. Back in 2017–2018, I developed a rule-based algorithm that used reinforcement learning to learn from previously reconciled transactions and attempt to reconcile new ones. While it worked to some extent, not all transactions were successfully reconciled, and as transaction complexity has increased over time, the system is no longer sufficient. We now need a more robust solution that can either replace the rule-based system or handle unreconciled transactions in a different way.
Currently, accountants spend many hours manually reconciling data with custodians such as Morgan Stanley and JP Morgan, which is highly time-consuming. I have already created templates for money market transactions across several banks, and those are working well. However, there are still many other transaction types to manage—such as buy, sell, dividend, and coupon payments—some of which involve foreign withholding tax (FWT), fees, or exchange rates. The current rule-based approach struggles with these scenarios, making the process slow and inefficient.
Today, I am exploring the use of AI/ML models such as XGBoost or Random Forest, either to extend the existing rule-based system or to replace it entirely. The goal is to build a more accurate, scalable, and efficient reconciliation system.