Skip to Main Content
We analyze portfolio creation techniques in a high frequency trading domain and randomly changing environments. We aim to create the best risk/reward portfolio based on thousands of profit histories of automated trading robots. We show that the effectiveness of standard portfolio weight calculation rules depends on the dimensionality, N, and the sample size, L, ratio. To resolve dimensionality / sample size dilemma we suggest designing a multistage feed-forward multi-agent system (MAS). At first we make simple 1/N Portfolio based expert agents. Then we use them and the regularized mean-variance framework to form a large number of more complex fusion agents. Finally we use a trained cost sensitive set of perceptrons to recognize the most successful fusion agents for making a final 1/N Portfolio based weights calculation. Experiments with 7708-dimensional 2004-2012 data confirm the effectiveness of the new approach.