Can agent-based models probe market microstructure?

We provide evidence that the use of realistic order matching procedures in agent-based models that attempt to represent continuous double auction markets at an intraday time scale introduces nuanced difficulties for model calibration, even when the calibration techniques employed perform well on simpler, closed-form models. We find that the method of simulated moments, though able to determine a number of parameters rooted in market microstructure with relative confidence and recover important features of real financial markets such as order flow correlation, is only able to determine an ambiguous link between data and parameters related to agent behavioral rules and population dynamics. We argue that this may either result from limitations of the calibration techniques employed, suggesting that more sophisticated approaches would need to be considered, or may alternatively point to the possibility that the structure of the niches that agents exploit in real financial markets may be more important determinants of measurable dynamics than the behaviors they engage in to exploit those niches.