Scaling and forecasting in a data-driven agent-based model: Applications to the Italian macroeconomy

Abstract

Agent-based models typically replicate stylized facts but lack macroeconomic forecasting capabilities. Recent advancements aim to make these models data-driven, enabling predictive applications in macroeconomics. Using data primarily from Eurostat (1996–2019), we calibrate an increasingly popular data-driven model to the Italian economy and evaluate the forecasting performance of macroeconomic variables for both Austria and Italy across various model scales. Our findings show that scale has no impact on forecast accuracy. To enhance the model we test modifications to agents’ expectations and firms’ production plans, and run long-term simulations to explore model dynamics and identify areas for refinement. The results demonstrate the model’s adaptability to different country specifications, with forecasting performance comparable to basic econometric models. Scale analysis and long-term analysis reveal unexplored heterogeneity and suggest that the model should further leverage the potential of agent-based microfoundations to improve forecasting.