Individual evolutionary learning and zero-intelligence in the continuous double auction

We study behavior in a Continuous Double Auction. In this paper we report on two models, ZI and IEL, which we tested against each other using two very different data sets: a large, uncontrolled set from classroom experiments using the MobLab interface and a small, controlled set from experiments at SFU. We found that drawing subjects from a pool composed of 70% IEL agent and 30% NI agents, who randomly order from [0; 250], generated results that were a very good fit to the 2090 observations from MobLab. That mixture outperformed a pool composed only of ZI agents. We found that a pool composed of only IEL agents was a good fit to the 25 observations from SFU. With respect to the distribution of efficiency, that IEL model dominated the ZI model. But, with respect to the distribution of average prices, the ZI model does a little better than IEL. Weighing the efficiency and price fits equally, IEL seems to be a better overall fit to the SFU data than ZI.