From ABM back to real data: time series visualization and model selection in the K+S agent-based model

Agent-Based Models (ABMs) provide powerful tools for economic analysis, capturing microto- macro interactions and emergent properties. However, integration with empirical data has been a persistent challenge. To address it, we propose a protocol for integration between empirical data and ABM, building a new multidimensional similarity index that aggregates different similarity measures into a composite score, specifically designed to quantify alignment between simulated and real-world data. This metric enables a complete model ranking procedure, facilitating a streamlined model selection. The protocol is designed to be model-agnostic and flexible, allowing its application to a wide range of models beyond ABMs, including aggregate dynamical systems and any type of computational model. As an example, we apply our methodology to different configurations and model versions of the Schumpeter meeting Keynes (K+S) ABM family (Dosi, Fagiolo, and Roventini, 2010) using US data (from 1948Q1 to 2019Q1). Next, we propose a policy-informed application, attributing different weights to variables associated with policy-making decisions and technological change. The exercise is done in order to showcase the capacity of the procedure to target specific policy variables of interest, allowing for the design of empirically informed scenario analyses and projections on real-world dynamics.