Evolving software traders and detecting community structure in financial markets

A trophic network, commonly referred to as a food web, describes the feeding relationships between different groups of species in an ecosystem. Ecologists construct trophic networks to aid their understanding of ecosystems. In the realm of financial markets, trophic networks can serve an analogous role. Their use could potentially illuminate underlying dynamics responsible for commonly observed macro-level phenomena. For example, one might hypothesize that periods of market volatility occur after a keystone trader species becomes inactive and the trophic network subsequently restructures itself. The primary topic in this research investigates whether it is possible to detect trophic structure within real-world financial markets. Secondarily, the efficacy of using genetic programming to evolve software traders in a simulated stock market (continuous double auction) is examined.;The research to follow is split into three parts. In Part I, new tools for detecting community structure in complex networks are developed. First, a two-phase macro-strategy for community detection is introduced. The approach is unique in that it can be used in combination with any existing community detection algorithm to provide high-yield, robust results. Second, the resolution limit inherent to the community structure measurement known as modularity is illustrated experimentally. To overcome this limitation, a fine-granularity community structure measure called divisionality is developed. Third, a dual-assortative measure (DAMM) of community structure is established. DAMM extends the domain of networks that can be analyzed for community structure to include those with negatively weighted edges.;Part II focuses on the evolution of software agents that compete in an artificial financial market. The evolutionary framework is based on a stack-based language (Staq) that was developed for genetic programming (GP). The genetic programs of two evolved agents, each based on a different fitness function, are examined. One of these evolved traders, known as clear & hoist (CH), reveals a limitation of the simulated market: a lack of fundamentalism. Two value-based strategies are developed to address this shortcoming. The effect of each strategy on the CH trader is independently examined.;In Part III, the community structure tools developed in Part I are used to detect trophic species in financial market data. After introducing the trophic detection algorithm, a methodology for assessing the significance of detected structure is described. The efficacy of the approach is demonstrated using simulated data. Finally, real-world data from the London Stock Exchange (LSE) is examined using the trophic detection framework. Although significant structure is detected in subsets of the real-world data, the results are inconsistent. However, given limitations of the LSE data, the lack of consistent detection is not surprising. Most notably, each trader in this data represents an entity acting on the behalf of many individuals and lx institutions having different strategies. Due to this aggregation, the trading actions of individuals are obfuscated and thus the trophic structure is as well. Examination of real-world data with greater specificity---detailing trades at the level of individuals---is warranted.