I study the relationship between sell-side analysts’ forecast bias and stock returns by comparing three forecast families—price targets (PTG), earnings per share (EPS), and long-term growth (LTG)—to ex-ante machine-learning (ML) benchmarks. I show that bias predict returns in a systematic and nonlinear manner. Across horizons, forecast bias predicts returns positively in the short term but negatively in the long term. Across the distribution of bias, both extreme optimism and extreme pessimism are followed by higher subsequent returns relative to moderately biased forecasts, generating a robust U-shaped relationship between bias and returns. These patterns are consistent across forecast families, and bias in return expectations co-moves positively with bias in cash-flow expectations, suggesting a common source of belief distortions. I show that conformism to the consensus provides a common denominator, and that anchoring offers a powerful reduced-form model of analysts’ belief formation. An asset-pricing model with asymmetric information and positively skewed fundamental shocks accounts for these dynamics: analysts anchor on consensus and misperceive the informativeness of signals, producing the observed bias-return relation.