Learning Predictive Models for Financial Time Series by Using Agent Based Simulations

In this work, we discuss a computational technique to model financial time series combining a learning component with a simulation one. An agent based model of the financial market is used to simulate how the market will evolve in the short term while the learning component based on evolutionary computation is used to optimize the simulation parameters. Our experimentations on the DJIA and SP500 time series show the effectiveness of our learning simulation system in their modeling. Also we test its robustness under several experimental conditions and we compare the predictions made by our system to those obtained by other approaches. Our results show that our system is as good as, if not better than, alternative approaches to modeling financial time series. Moreover we show that our approach requires a simple input, the time series for which a model has to be learned, versus the complex and feature rich input to be given to other systems thanks to the ability of our system to adjust its parameters by learning.