Christina Ann Lacomb

A genetic algorithm approach to detecting temporal patterns indicative of financial statement fraud

This study presents a genetic algorithm approach to detecting financial statement fraud. The study uses a sample comprising a target class of 51 companies accused by the Securities and Exchange Commission of improperly recognizing revenue and a peer class of 339 companies matched on industry and size (revenue). Variables include 76 comparative metrics, based on specific financial metrics and ratios that capture company performance in the context of historical and industry performance, and nine company characteristics. Time-based patterns detected by the genetic algorithm accurately classify 63% of the target class companies and 95% of the peer class companies. Copyright

Agent Learning to Manage Costs for Event Detection

Recent scandals around manipulated financial filings have caused investors and analysts to search for alternative ways to study the financial health of companies. The use of news events such as CEO or auditor changes has proven valuable at providing insights into the status of a companys financial health. However, this information can be extremely difficult and expensive to gather in practice. An intelligent multi-agent system was designed and developed to simulate the collection of news events in an efficient, cost-effective manner. Results show that a multi-agent system is an effective tool for collecting critical business intelligence while minimizing cost

Monitoring key company events through deliberative learning

Recent scandals concerning the discovery of fraud committed by a few high profile companies has reinforced a need for innovative approaches to detecting fraudulent company behavior. Fraud detection experts agree that many of the critical clues to fraud, such as frequent management and auditor changes, can be found in qualitative sources such as news articles, press releases, and footnotes accompanying financial statements. This paper presents a simulated multi-agent system that learns how to collect valuable events from textual sources with pinpoint precision, utilizing the best content providers for each event type while minimizing the overall cost.