Fiscal Policy Towards Optimizing Macroeconomic Indicators by Integrating FRB/US with Reinforcement Learning

Fiscal policy optimization in the context of competing macroeconomic objectives poses significant challenges for economic policymakers. Although the Federal Reserve’s FRB/US model provides sophisticated forecasts, its reliance on predefined scenarios constrains exploration of the full policy space. This research introduces the RL-FRB/US model which integrates the FRB/US model and Proximal Policy Optimization (PPO) reinforcement learning (RL) model with an active enhancement of relocation mechanism for fiscal policy optimization. The RL-FRB/US model demonstrates significant performance improvements over baseline FRB/US simulations in the period 2000–2024. By 2024Q2, the RLFRB/US model achieved higher real GDP (RL-FRB/US model: 23,407 trillion $ vs. FRB/US model: 23,218 trillion $), lower unemployment (3.23% vs. 3.96%), and more effective inflation management (PCPI RL-FRB/US model: 317.9 vs. FRB/US model: 312.3). During recessions, the model consistently delivered superior counter-cyclical responses, with unemployment peaks significantly reduced during major downturns—during the 1982 recession, peak unemployment reached only 9.9% compared to 10.9% in traditional simulations. While the RL-FRB/US model showed similar federal budget deficits by 2024 (RL-FRB/US model: -1,767 trillion $ vs. FRB/US model: -1,758 trillion $), it achieved substantially lower debt-to-GDP ratios (RL-FRB/US model: 26,535 trillion $ vs. FRB/US model: 30,186 trillion $) through more strategic debt management during expansionary periods. The output indicates that a combination of reinforcement learning and macroeconomic modeling introduces more reliable outputs than the traditional model, which provides policymakers with powerful decision-support instruments to balance inflation control, targeted unemployment rate and fiscal sustainability.