This thesis examines whether different machine learning models, when used to forecast firm-level equity returns, imply materially different stochastic discount factors, and whether those differences translate into economically meaningful variation in firm-level risk exposures and pricing errors. The analysis compares linear and nonlinear models including ordinary least squares regression, least absolute shrinkage and selection operator (LASSO) regression, principal components regression, gradient boosted decision trees, feedforward neural networks, and long short-term memory (LSTM) networks under a unified empirical design. Using monthly excess returns for S&P 500 constituent stocks from the Center of Research in Security Prices over the period 2000 to 2024, with predictors drawn from standard factor models and return-based anomalies, expected returns are first estimated at the firm level. These forecasts are then mapped into traded, time-varying stochastic discount factors viamean–variance efficient portfolio constructions derived from unconditional moment conditions. Model performance is evaluated using out-of-sample cross-sectional R2 and average absolute pricing errors. Across all specifications, the estimated stochastic discount factors are well behaved and economically sensible. Nonlinear models generate more state-contingent and heavy-tailed pricing kernels than their linear counterparts, reflecting richer dynamics in state pricing. However, these differences do not translate into substantial improvements in firm-level pricingaccuracy, with average pricing errors remaining similar across model classes. This evidence suggests that model choice primarily affects the dynamics and distributional properties of the pricing kernel rather than overall cross-sectional explanatory power. Overall, the findings indicate that while machine learning methods meaningfully alter the implied structure of the stochastic discount factor, gains from nonlinear modeling are limited in terms of average firm-level pricing performance. A key limitation is the reliance on mean–variance efficiency in stochastic discount factor construction, as a nontrivial subset of firms exhibit idiosyncratic behavior that is difficult to price accurately within this framework. These results highlight the distinction between modeling state-dependent pricing dynamics and improving cross-sectional pricing accuracy, and point to directions for future work that relax portfolio efficiency assumptions or incorporate alternative pricing restrictions.