State-dependent impulse responses in agent-based models: A new methodology and an economic application

The paper delves into the potential of Agent-Based Models (ABM) in analysing phenomena characterized by the non-linear propagation of shocks and system dynamics. Recognizing that state dependency can naturally emerge in complex evolving systems, we present a new methodological framework to evaluate state-dependent (or non-linear) impulse response functions in an ABM setting. Inspired by threshold time series modelling approaches, we propose analysing state-dependent impulse responses by creating alternative controlled states of the system, from which randomized impulse responses can be computed. Furthermore, a data-driven, machine- learning algorithm is proposed to endogenously identify relevant system states for the observed response. To the best of our knowledge, this is the first time such an approach is advanced. An R library implementing all the required methods is also offered to ensure applicability in diverse fields. Finally, the methodology is applied in economics to test for monetary policy shocks in a reference macro ABM, highlighting its effectiveness in mapping the system impulse response to the identified key state variables, as well as showing the importance of state dependence for policy design and systematic identification of critical system states.