Cognitive Embeddings: Investing Behavior to Recover Beliefs

, Joint Program in Financial Economics PhD Student

I address the inversion problem by recovering analysts’ cognitive (belief) embeddings from text. The inversion problem is to recover latent beliefs from observed behavior instead of merely predicting the behavior itself. Even when analysts share the same public information (filings, calls, news), they allocate limited attention differently, which produces systematic forecast dispersion driven by persistent differences in these embeddings. I represent each analyst’s information set s text embeddings, and the model’s attention highlights which parts of a disclosure the analyst emphasizes. Parameters are learned using forecast error, which calibrates the belief embeddings to realized outcomes. The attention layer implements a constrained optimization that reflects limited cognitive resources, linking the architecture to bounded rationality. The pooled, report level belief embedding summarizes what was taken from the call. By reconstructing beliefs and then predicting outcomes, the model inverts the forecasting process. The framework extends to settings with shared information but heterogeneous interpretation, including macroeconomic forecasting, managerial decision making, and retail investor behavior.