We reassess the predictive power of risk-neutral excess-of-market stock variance (Martin and Wagner (2019, MW)) for stock returns. After correcting two look-ahead biases that influence evidence supporting an average predictive coefficient of 0.5 reported in prior works, we find the data are too noisy to reject the null hypothesis of an average coefficient of zero. However, this insignificant average predictive coefficient conceals the predictability’s strong covariance with market volatility, as well as its large variation across characteristics-sorted subsamples. Out-of-sample analysis confirms that while the MW model does not significantly outperform benchmark models on average, it significantly outperforms during high-volatility periods.