We use the long-term Capital Market Assumptions of major asset managers and institutional investor consultants from 1987 to 2022 to provide three stylized facts about their subjective risk and return expectations on 19 asset classes. First, there is a strong and positive subjective risk-return tradeoff, with most of the variability in subjective expected returns due to variability in subjective risk premia (compensation for market beta) as opposed to subjective alphas. Second, belief variation and the positive risk-return tradeoff are both stronger across asset classes than across institutions. And third, the subjective expected returns of these institutions predict subsequent realized returns across asset classes and over time. Taken together, our findings imply that models with subjective beliefs should reflect a risk-return tradeoff. Additionally, accounting for this subjective risk-return tradeoff when modeling multiple asset classes is even more important than incorporating average belief distortions or belief heterogeneity in our setting.

We propose a structural model of constant gain learning about future earnings growth that incorporates preferences for the timing of cash flows. As implied by the model, a cross-sectional decomposition using survey forecasts shows that high price-earnings ratios are accounted for by both low expected returns and overly high expected earnings growth. The model quantitatively matches a number of asset pricing moments, as learning about growth interacts strongly with the preference for the timing of cash flows, and provides insights on the roles of risk premia and mispricing in the cross-section of stocks. The magnitudes and timing of the comovement between prices, earnings growth surprises, and anomaly returns are all consistent with a gradual learning process rather than expectations being highly sensitive to the most recent realization. Large earnings growth surprises do not immediately translate into large one-period returns, but instead are gradually reflected in future returns over time.

Traditional asset pricing models like the CAPM explain realized returns worse than newer asset pricing models like Fama-French-5, but why? I show that traditional models explain subjective risk and return expectations well but also predict return-subtracting mispricing. Newer models, by contrast, explain subjective risk and return expectations poorly but predict return-enhancing mispricing. These results imply that the CAPM is a good model of risk but fails to explain realized returns because risk is correlated with mispricing. I explain this disconnect by a model in which all the CAPM assumptions hold, except that some investors have an optimism bias.