Evaluating Firm-Level Expected Return Proxies: Implications for Estimating Treatment Effects