The problem

The compliance and audit team of a leading finance enterprise is responsible for monitoring, validating, and book-keeping of all ledger transactions every quarter. The current manual process and rule engine is a tedious and time-consuming process needing a major overhaul in the system. Considering transaction volume across verticals, the team wanted to know if Cognitive BI can come up with an innovative analytical tool to flag these non-conformances as violations and raise red flags as necessary.

The cognitive solution

The challenge to be able to carry out the detection is twofold:

1. Log analyzer

• Extract ledger transactional logs from the log file on a daily bases
• Build out a graphical structure of the processes and record tabular metrics on time taken for closure, approvals by roles per account type
• Use this metadata as a new log audit table and textual notes from ledger evaluation as unstructured data per account

2. Anomaly detector

• Using the metadata generated and sparse matrices from textual notes, a fraud classifier was developed.
• The tree-based model acts as an automated data-driven set of a rules engine

2. Business rules

• Custom business rules are used as part of the rule engine to work on top of the identified frauds and reduce false positives
• The two-stage design reduces multiple redundant review and control volume of false positives and false negatives

Key insights

The deployed solution was able to automatically detect anomalies by red-flagging fraudulent or out of sequence transactions. With an overall reduction of 30% in fraudulent misses, some interesting insights came into business allowing them to tweak certain operational procedures.