Panel Data Models for Stock Returns: the Importance of Industries

We perform specification tests of unbalanced panel data models for stock return prediction using an estimate of the coefficient covariance matrix robust to spatial heteroskedasticity. We apply estimators of individual effects and time effects in unbalanced panel models. For a panel of 1243 US firms for the period 1985–2002 we find that industry effects are significant and interact with firm characteristics like size and momentum. Parameters on these characteristics are significantly different across industries. We therefore conclude that panel models are industry specific. This conclusion is robust to the estimation method and the data. High expected returns are mostly related to the cash flow-to-price ratio and the analyst earnings revisions, somewhat to the momentum, but hardly to valuation ratios like the book-to-price ratio. In-sample simulations of portfolio construction result in long-short portfolios with substantial abnormal returns.