We study retail investors' topic-specific attention and its stock-return predictability. From 340 million posts on China's largest stock forum, we build firm-day attention measures for 200 interpretable topics. Using LASSO, we identify the most predictive topic: a one-standard-deviation rise in its attention predicts 0.44–0.72 percentage points lower next-month returns across different keyword-intensity thresholds, with partial reversal within one to two months. Predictions are stronger among smaller, non-main-board firms, and post-2015. The results remain robust to orthogonalization against recent returns, aggregate post sentiment controls, and alternative information channels.