The optimal choice of scaling in economic agent-based models

In economics, two strategies are typically employed to reduce the size and complexity of models: (i) using representative agents by aggregating the actual entities, (ii) and downscaling, i.e. using only a sample of agents. While the first strategy has been studied in detail in mainstream economics, the implications of the second option – which is mainly used in complexity economics – are underresearched. This paper contributes to filling this gap by proposing a protocol for sensitivity analysis with respect to the scaling choice in these models. We introduce this protocol in a dual manner. First, we identify three main theoretical channels via which scaling can influence complex economic ABMs: (i) idiosyncratic shocks, (ii) information loss due to insufficient interactions, and (iii) the distribution of the characteristics of agents. Second, we analyse the implications of these mechanisms by assessing the trade-offs between three fundamental measures of model performance: precision, accuracy and running time, with different downscaling levels ranging between 0.25%–100% of the full population. We illustrate our approach using the model of Mérő et al. (2023), which is suitable for representing the housing market of Hungary at any scale in this interval (from 10,000 to 4 million agents). We show that in this model there is a non-trivial relationship between the scaling factor and the model performance. Not only does the model’s accuracy and precision depend on the model size in a non-linear manner, we also found that the evaluation of a scenario at a given level of precision takes only three to four times longer with 100 times more agents.