Abstract
Abstract
Can an independent AI forecast correct systematic bias in human forecasts when both inform a portfolio decision? We show that disagreement between two potentially biased forecasts identifies only their relative bias; their shared bias can be learned only from realized outcomes or an external anchor. Shared bias does not affect rankings but directly affects decisions defined by an absolute target or threshold. Building on Chen and Lim (2020), we develop a dual-source calibrated fusion model in which both forecasts are calibrated on historical outcomes, disagreement predicts conditional error variance, errors may be correlated, and cross-asset common shocks enter through a factor covariance. We decompose the decision value of the second forecast into changes in the safety margin, residual common bias, and asset selection. On 37,540 Chinese firm-quarters, with analyst forecasts, an AI forecast from the same filings, and 28 out-of-sample quarters under fiscal-year-clustered inference, we find that adding the AI system as a second source lowers the negative log predictive density relative to a calibrated single-source benchmark, mainly through disagreement-dependent uncertainty rather than the AI conditional mean, but yields no detectable improvement in a target-yield portfolio. Calibration alone reduces the shortfall by about one percentage point. No specification attains nominal chance-constraint coverage, indicating material covariance misspecification. This collaborative-intelligence design separates calibration against realized outcomes from the information supplied by an independent AI view: an AI forecast can improve probabilistic accuracy without improving a level-sensitive decision, and disagreement cannot substitute for calibration against outcomes.