Domain-Driven Local Exceptional Pattern Mining for Detecting Stock Price Manipulation

Abstract. Recently, a new data mining methodology, Domain Driven Data Mining (D3M), has been developed. On top of data-centered pattern mining, D3M generally targets the actionable knowledge discovery under domain-specific circumstances. It strongly appreciates the involvement of domain intel-ligence in the whole process of data mining, and consequently leads to the deliverables that can satisfy business user needs and decision-making. Follow-ing the methodology of D3M, this paper investigates local exceptional patterns in real-life microstructure stock data for detecting stock price manipulations. Different from existing pattern analysis mainly on interday data, we deal with tick-by-tick data. Our approach proposes new mechanisms for constructing mi-crostructure order sequences by involving domain factors and business logics, and for measuring the interestingness of patterns from business concern pers-pective. Real-life data experiments on an exchange data demonstrate that the outcomes generated by following D3M can satisfy business expectations and support business users to take actions for market surveillance.