We simulate data from well-known economic frameworks in which the true conditional SDF can be calculated. We examine multiple factor construction methodologies in these simulated economies, estimating the SDF by linear regression on past factor returns. Creating many factors using nonlinear combinations of characteristics as portfolio weights and then using shrinkage in the regression (complexity) works well. However, a latent factor model with fewer factors than characteristics (instrumented principal components) also works well. Classical methods do not perform well.