This paper constructs time-varying directed news networks from a large dataset of over 1 million news articles to study cross-sectional stock returns of S&P500 stocks. The authors define a news linkage based on the position of firms in articles: the firm in the headline is the lead, and firms mentioned in the content are followers. A careful algorithm is proposed to identify firm tickers, reducing type-I errors. Using these networks, two key variables are derived: lead return (a weighted average of lead firms' returns) and network degree (a measure of attention). The empirical analysis reveals a strong comovement effect: follower stock returns move positively with contemporaneous lead returns, even after controlling for Fama-French factors and firm characteristics. Additionally, a reversal effect is found from lead returns on 1-day ahead follower returns. Portfolio tests show that sorting on lead return generates a significant daily return spread of 165.15 bps. More importantly, the network degree, as a proxy for news attention, provides robust and significant cross-sectional predictability of monthly stock returns. Decomposing the networks shows that within-sector linkages, large lead firms, and lead firms with lower stock liquidity are particularly important for this predictability. The results are robust to various controls and network formation windows, highlighting the value of news-based networks in understanding stock return dynamics.