My work studies the information boundary of counterfactual policy analysis.
I ask when observed histories reveal enough about hidden beliefs, adaptive capacity, policy support, generated exposures, and learned laws of motion to support a counterfactual path; when policy or deployment changes the future information environment; and how credibility can be restored through additional variation, proxy information, protected experimentation, or more limited policy claims.
Solo-authored work
Counterfactual Blindness
This paper studies when observed histories fail to preserve the future experimental capacity needed to evaluate a counterfactual policy query. The archive can be complete while the latent continuation states or off-path experiments relevant for a new policy comparison are no longer supported. I characterize when this creates decision-relevant blindness, when targeted experimentation can restore credibility, and when the counterfactual boundary is binding under the available experimental opportunities.
Reward-Free Reinforcement Learning Before Knowing What Will Matter
This paper studies pre-objective information acquisition when a controlled transition record may expire before the future decision criterion is known. I characterize when reward-free exploration creates reusable evidence for later tasks, when waiting is preferable, and how task arrival and evidence expiration determine the value of collecting information before knowing what will matter.
Inference with Generated Shifts in Shift-Share Designs
This paper studies inference when the shifts entering exposure designs are estimated rather than primitive. I characterize how first-stage uncertainty propagates through exposure shares and develop inference that reproduces the relevant cross-stage sampling relationship.
Collaborative work
Prediction Is Not Control
With Noah Williams
An estimated law of motion can predict accurately under historical feedback yet fail when used for counterfactual control. We develop identification, occupancy, value, and support conditions for reliable recursive policy optimization.
Scheduled for the 2026 Chapel Hill-Copenhagen Conference on Macroeconomics and Deep Learning.
Adaptivity and Vulnerability
With Noah Williams
We study adaptive capacity—the ability to translate macro information into consumption-saving adjustment—and show how it changes future vulnerability paths and liquidity-based household rankings.
Presented at SEM 2026 in St. Louis.
Diagnosing and Correcting Model-Based Counterfactual Paths
With Chengwei Wang
We study when residual state contains recoverable policy-contingent information in model-based counterfactual paths. Conditional diffusion is used as a correction layer for path forecasts, with the identifying information set and failure cases stated explicitly.
Scheduled for SEA 2026.
Homepage - Xueqing Yang 