Calibration of Agent-Based Models by Means of Meta-Modeling and Nonparametric Regression

Taking agent-based models to the data is still very challenging for researchers. In this paper we propose a new method to calibrate the model parameters based on indirect inference, which consists in minimizing the distance between real and artificial data. Basically, we first introduce a nonparametric regression meta-model to approximate the relationship between model parameters and distance. Then the meta-model is estimated by local polynomial regression estimation on a small sample of parameter vectors drawn from the parameter space of the ABM. Finally, once the distance has been estimated we can pick the parameter vector minimizing it. One innovative feature of the method is the sampling scheme, based on sampling at the same time both the parameter vectors and the seed of the random numbers generator in a random fashion, which permits to average out the effect of randomness without resorting to Monte Carlo simulations. A battery of simple calibration exercises performed on an agent-based macro model shows that the method allows to minimize the distance with good precision using relatively few simulations of the model.