Fiscal Policy Simulation Using AI And Big Data: Improving Government Financial Planning

Abstract

In almost every country, fiscal policy simulation design is crucial and challenging, since finance authorities need to prepare plans for government revenues and expenditures on a set of financial variables over time. Conventionally, such complex policy design problems are handled by experts, whose goal is to achieve the desirable control performance under a set of regulations. However, when dealing with sophisticated contingent regulations, constraints, and massive state spaces, expert knowledge may not be enough for optimal plans. This paper demonstrates a new approach that offers helpful computational tools for policy design. This novel approach views the government policy design process as a complex multi-agent deep reinforcement learning problem, where the AI Economist employs interpretable policy functions, and focuses on the design of the future AI-based policy options generation engines. To better capture policy design expressed in the control space, this multi-agent framework directly models the fiscal policy options generation process, developing both regressive and generative AI-based engines for the design of policy trajectories. Furthermore, they develop training from both on-policy samples from fiscal simulation execution and offline samples. Based on the AI Economist, this approach can hold a better chance seeking better solutions to address even those most difficult and sophisticated finance design questions.