Language models improve economic policy using simulation feedback
Grounding Large Language Models in DSGE Simulators for Policy Generation and Forecasting
Artificial IntelligenceComputational Engineering, Finance, and ScienceMachine Learning
Summary
Language models can write economic policy advice that sounds right, but that does not prove they understand how economies really work. The authors tested a language model by putting it inside economic simulations that mimic real-world dynamics. At each step, the model makes policy decisions, sees how the economy changes, and learns from long-term rewards. This helps judge policies by their actual economic effects over time, not just by sounding plausible.
What this means in practice
- •For economic modelers: Generate and evaluate economic policies in long-term simulated environments integrating language models and DSGE frameworks.
- •For financial forecasters: Use language model-driven policies that respond to simulated economic shocks for improved scenario analysis and forecasting.
Tested on simulated data.
Authors
Aditya Dubey, Namah Gupta, Vinti Agarwal
Abstract
Large language models can produce economic policy responses that sound reasonable, but this does not show that their actions are consistent with economic dynamics. We test this by placing an instruction-tuned language model inside six Snowdrop-backed dynamic stochastic general equilibrium (DSGE) simulators. At each turn, the model observes the economy and a change in economic discourse, selects a bounded policy action, and receives the next simulated state and an economic reward. We implement a common Python interface for repeated rollouts, persistent shocks, state cloning, and rolling-horizon simulation. This setting creates a long-horizon credit-assignment problem. Policy effects may appear several quarters after an action is taken. PPO has a learned value function that can propagate delayed reward to earlier tokens through generalized advantage estimation. GRPO has no learned value function and instead assigns a group-relative advantage from complete rollout returns. It therefore cannot distinguish which earlier turn caused the outcome; if every rollout receives the same return, the normalized advantage is zero. We use PPO as the primary method and GRPO as a matched critic-free baseline. The experiments also test directional semantic signals, reward horizon, trajectory warm starts, cross-simulator transfer, and historically anchored pandemic and monetary-policy shocks. The objective is to judge policy actions by their simulated economic consequences rather than by plausible language alone.