Several authors have raised the issue of non‐stationarity of security returns in empirical tests of the Arbitrage Pricing Theory (APT). This paper tests for one form of non‐stationarity, namely, conditional heteroskedasticity, in the empirical APT with observed factors. Using monthly stock returns for the period 1970 to 1988, this paper shows that conditional heteroskedasticity is a pervasive phenomenon leading to inefficient estimates of factor betas. Ignoring the problem may produce erroneous conclusions as to which risk factors require a premium. Furthermore, grouping individual securities into portfolios does not appear to diminish the presence of conditional heteroskedasticity.