We isolate key economic assumptions that cause misspecification and identification failures in conventional factor risk premia estimation. To address these issues, we introduce Tradable Factor Risk Premia (TFRP), defined by the negative covariance between the factors and the projection of the Stochastic Discount Factor onto returns. TFRP estimators are point-identified, independent of other risk factors, and have a natural interpretation through mimicking portfolios. They also enable Oracle inference, performing as if weak or irrelevant factors were known. Empirically, we present the first large-scale study of the identification problem across the entire factor zoo, showing that it is a widespread issue.

Are there risk factors that are pervasive across major classes of corporate securities: stocks, bonds, and options? We employ a novel econometric procedure that relies on asset characteristics to estimate a conditional latent factor model. A common risk factor structure prominently emerges across asset classes. Several common factors explain a substantial amount of time-series variation of individual asset returns across all three asset classes, and have sizable Sharpe ratios. Several of our factors are highly correlated with some of asset-class-specific factors as well as macroeconomic and financial variables. While a small set of common factors does not fully capture the cross-section of average returns, imposing the factor structure is useful in practice, especially in out-of-sample analysis. A mean-variance efficient portfolio that utilizes asset characteristics achieves a high Sharpe ratio as different asset classes hedge each other's exposures to the common factors.

We show that a nonlinear factor model is closer to the mean-variance frontier the larger the Sharpe ratio of its mimicking portfolio. A linear factor model is a special case for which the mimicking portfolio is the tangency portfolio of the factors. Across a wide range of models, nonlinearities of the factors are priced as they significantly increase the Sharpe ratio metric. The preferred model depends on the test assets, which are relevant for model comparison as they are needed to mimic factor nonlinearities. Momentum anomalies are the most important in the mimicking portfolios, which helps explain why they command a premium.