We develop and estimate a dynamic model of financial-market learning in which traders can endogenously switch between correct and misspecified models. Unlike frameworks that impose behavioral biases exogenously, our traders rationally choose to deviate from the correct model when short-term gains outweigh informational accuracy - a mechanism we term greed-driven model choice. This switching behavior generates persistent wedges between subjective and objective beliefs, producing volatility smiles, return predictability, and pricing-kernel distortions as equilibrium outcomes. Using option-price data and a structurally estimated binomial tree model with path dependence and informed traders, we identify belief-switching thresholds that quantify how greed and fear jointly shape learning and risk premia. We document economically significant welfare and performance costs of misspecification, expressed in certainty-equivalent losses and Sharpe-ratio reductions. Our results reveal how heterogeneous information processing and learning frictions jointly shape equilibrium pricing kernels, offering new empirical evidence on misspecified learning in dynamic models and the behavioral foundations of rational asset pricing.
