Introduction

The ultimate test of a theory is its empirical validity, so the question whether a model ‘fits the data well’ is crucial. In the last chapter, we introduced some of the many issues involved in model evaluation. Here, we dig into the problem of tuning the values of the parameters. Moreover, we are also interested in the values of the estimated parameters for interpreting the model behaviour and to perform what-if type counterfactual (e.g., policy) evaluation exercises. Agent-based (AB) models are in general complex non-linear models, and can therefore display many different behaviours depending on the region of the parameter space being sampled. Assessing the performances of the model in the right region of the parameter space is therefore important for model evaluation. Once this region has been identified and the model deemed appropriate for its scopes, lessons can be learned about what might happen in the real world if some of the parameters changed, either as a consequence of some unforeseen developments (scenario analysis) or due to some specific actions purposefully implemented (policy analysis).

Our goal, broadly speaking, is comparing (possibly an infinite number of) instances of the model with different parameter values and select those that fits the data better.

Before going on, a first remark is necessary. Generally, we do not aim at calibrating or estimating a model by getting to a single optimal choice for all the parameters. In a frequentist approach, we rather look at confidence intervals – that is, ranges where the ‘true’ value of the parameters, assuming the model is correctly specified, is likely to lie – while in a Bayesian approach we focus on the posterior probability distributions for the parameters – reflecting our uncertainty about the parameters values given our prior knowledge and the information contained in the data. In this chapter we will provide examples of both approaches.