Quantitative Estimation of Market Sentiment: a discussion of two a lternatives

Modeling and estimating the time series (the value in time) of assets traded in real financial markets is an intriguing challenge that has attracted researchers from many fields including Economics, Statistics, and more recently Computer Science thanks to the common availability of computational power that made possible the investigation of computational or simulative methods for modeling financial time series. In this paper, we discuss a computational simulation technique based on agent based modeling and learning to closely approximate the SP500 and DJIA indexes over many periods and under several experimental set ups. According to our modeling approach, the value in time of a financial asset emerges as an aggregate result of several independent investment decisions (to buy, to sell, or to hold) during a short period of time. We can therefore reproduce the process of value formation by computationally simulating the community of agents-investors. We will compare our system's performances with respect to other approaches on the same time series to provide empirical data about the effectiveness of different computational techniques. The main finding emerging from our work is that a simple architecture for a simulator combining agent based modeling and learning produces close approximations for the SP500 and DJIA time series. The approximation results are comparable to those observed when evaluating prediction rules learned by neural networks or particle swarm optimization. An additional characteristic of our modeling approach is that it can provide insights about the contribution of each agent to the process of value formation for a financial asset.