Multi-agent system based on oscillating agents for portfolio design

Improved cellular-automaton-based models of excitable media were employed to mimic economic and financial units in rapidly changing environments. Depending on the model parameters we can obtain chaotic, spiral, vibrant, or regular agents´ behaviors. Sums of the outputs of groups of integrated agents are fluctuating in time, similar to observed oscillations in the financial time series. The spectral representation of the output time series allows classification of the agent groups and determination of the model parameters. This inspired the creation of spectra-based clustering of financial time series and the development of a two-stage multi-agent system (MAS). With specific trading agents, the latest training data interval is less useful than previous agents for selecting the best well-timed detected data history segments. The adaptive MAS-based schema notably outperforms three benchmark methods in situations in which we have very complex portfolios with several thousand investment profiles.