2026working paper
European sovereign debt control through reinforcement learning
Tato Khundadze, Willi Semmler
Frontiers in Artificial Intelligence · AI for Human Learning and Behavior Change
The resilience of economic systems depends mainly on coordination among key stakeholders during macroeconomic or external shocks, while a lack of coordination can lead to financial and economic crises. The paper builds on the experience of global and regional shocks, such as the Eurozone crises of 2009–2012 and the economic disruption resulting from COVID-19, starting in 2020. The paper demonstrates the importance of cooperation in monetary and fiscal policies during emergencies to address macroeconomic non-resilience, particularly focusing on public debt management. The Euro area is chosen as the sample for testing the models presented in the paper, given that its resilience is heavily dependent on cooperation among different actors within the region. The shocks affecting nations within the European Union are asymmetric, and the responses to these shocks require coordination, considering heterogeneous economic structures, levels of economic development, and policies. We develop a macroeconomic modeling framework to simulate fiscal and monetary policy interactions under a cooperative regime. The approach builds on earlier nonlinear control models and incorporates modern reinforcement learning techniques. Specifically, we implement the Soft Actor-Critic algorithm to optimize policy responses across key variables including inflation, interest rates, output gaps, public debt, and government net lending. We demonstrate that the Soft Actor-Critic algorithm provides comparable or, in some cases, better solutions to multi-objective macroeconomic optimization problems, in comparison to Nonlinear Model Predictive Control (NMPC) algorithm.
2026working paper
Learning Macroeconomic Dynamics from Data: Applications of Universal Differential Equations and SINDy
Tato Khundadze
SSRN
This paper introduces scientific machine learning methods to macroeconomic model discovery, focusing on Sparse Identification of Nonlinear Dynamics (SINDy), Universal Differential Equations (UDEs), and a hybrid UDE-SINDy pipeline in which a neural network embedded within a differential equation learns unknown dynamics that are subsequently distilled into interpretable symbolic equations via sparse regression. The Goodwin growth cycle, a canonical predator-prey formulation of distributional conflict between labour and capital, serves as the testing ground throughout. We survey the relevant methodology, implement each approach in a unified computational framework, and benchmark them on problems calibrated to macroeconomic scales, where data are scarce, noisy, and nonstationary. On stationary synthetic Goodwin data, the hybrid pipeline recovers the bilinear interaction coefficients to four significant figures, with the vu term emerging as the sole active component of the sparse representation. On nonstationary synthetic data, consisting of a four-regime piecewise system with forced oscillations, abrupt parameter shifts, and transition damping, both direct SINDy and UDE-SINDy recover the predator-prey structure with correct structural identification. These experiments establish that the methods can operate under conditions characteristic of macroeconomic time series. When the validated pipeline is applied to 77 years of U.S. quarterly data (1948:Q1-2025:Q3) on employment, capacity utilisation, and the labour share, neither method recovers the vu interaction at any regularisation strength. Direct SINDy returns dense polynomial models; the UDE-SINDy pipeline finds near-zero neural network corrections and no dominant interaction term. The result is consistent across both methods and both empirical specifications, classical (v, u) and structuralist (TCU, u). The Lotka-Volterra functional form that defines the Goodwin model is not present in the data. Several caveats apply, including the restriction to secondorder polynomial libraries and the treatment of the system as a two-dimensional autonomous ODE. The methods developed here are not specific to the Goodwin model and can be applied to other macroeconomic dynamical systems for which candidate functional forms can be specified.