This paper investigates whether the cross-sectional variance (CSV) of stock returns and its asymmetric components contain incremental information to predict stock market volatility under a high-frequency, heterogeneous autoregressive (HAR) model framework. We present novel evidence that CSV is a powerful predictor of future realized volatility, both in- and out-of-sample, even after controlling for the well-established predictors obtained from intraday data. Further analysis suggests that distinguishing between positive and negative CSV components in the forecasting model enhances the predictive capability of volatility models at all out-of-sample forecasting horizons, with the asymmetric HAR-type-ACSV model consistently outperforming all alternative HAR-type variations. We argue that the asymmetries in the predictive relation between CSV and volatility are largely driven by the disagreement among market participants that spikes during bad times. Finally, economic analysis shows that incorporating CSV in the forecasting model can generate sizeable economic gains for a mean–variance investor, suggesting that out-of-sample predictive ability of CSV can be exploited in forward looking investment strategies to enhance investment returns.