MacroIC: A Foundation Model for In-Context Macroeconomic Policy

Macroeconomic policy analysis relies on many structural models, each with its own preferences, frictions, and calibration. Each model must be solved or estimated separately, and none can adapt its decision rule to a new economy from a few observations, so knowledge about policy rules does not accumulate across models. We introduce , a foundation policy model that addresses this gap by pretraining once on simulated trajectories from heterogeneous structural models and adapting through in-context learning. A shared vocabulary of economic roles allows one causal transformer to predict actions from an economy's history without a model identifier or test-time weight updates. Its advantage over explicit rule estimation is concentrated in short demonstrations: on held-out monetary models, including additional models evaluated after the training recipe was fixed, a short context gives the frozen policy higher prediction skill than local rule estimation, a learned-prior estimator, and a pretrained tabular regressor, while explicit estimation catches up with longer histories. Replacing the added structural trajectories with reduced-form simulations lowers short-context skill under the same continuation recipe. Separately, the shared policy retains low consumption-equivalent regret on a suite of planner economies. These results delimit the regime in which structural policy pretraining is effective: predicting unfamiliar decision rules from limited demonstrations while retaining precision on previously learned control tasks.