Model selection in stock correlation networks

In this research we construct market networks to study correlation between the price return for all Dow Jones, NASDAQ-100 and S&P 100 indices that were traded over a period of time. We consider market networks, which have stocks as nodes and edges corresponding to correlated stocks. Specifically, a winner-take-all approach is used to determine if two nodes are adjacent. We identify that all networks based on the connecting stocks of highly correlated price returns display a scale-free degree distribution. Additionally, we use features for representing different aspects of the network. The feature includes small connected sub-graphs with three and four vertices. We use an algorithm to count frequently the number of the graphlets for our mathematical models and our constructed networks. Each network is assigned an 8-dimensional vector. We present a model selection algorithm based on supervised learning. Our algorithm classifies our market networks with the best fitting mathematical model.