I study how equity analysts form subjective beliefs about firms' earnings using their own written text from over 1.1 million equity research reports. Using large language models, I identify the topics discussed by analysts and represent topic-level information using textual embeddings. I introduce a novel text-instrumented Coibion-Gorodnichenko regression to examine analysts' misreactions to specific information. Using this new procedure, I find pervasive underreaction in short-term earnings forecasts across topics, whereas overreaction in long-term forecasts is concentrated in qualitative, intangible topics rather than quantitative, statistical ones. Revisions driven by qualitative information in long-term earnings forecasts strongly predict future stock returns. Finally, I use textual data to investigate the mechanisms underlying the documented misreactions. The empirical results suggest that overconfidence is an important driver of the overreaction to qualitative information, while herding appears to be important in explaining the overall underreaction seen in short-term forecasts.