How should investors value firms without return histories? In practice, investors typically proxy the cost of equity in discounted cash flow valuations using peer-based betas. Using IPOs as a natural laboratory, I show that disclosed business risks provide an informative basis for beta estimation. I introduce Aggregated Cluster Embeddings (ACE), a context-aware framework that encodes disclosed business risks into economically meaningful numerical representations. Using these representations as inputs to machine learning models, I obtain beta estimates that are up to 31 percent more accurate than peer-based benchmarks. Despite this improvement, investors do not appear to fully incorporate the information contained in disclosed business risks at issuance. Portfolios formed on predictions derived from ACE representations exhibit cross-sectional return predictability. The resulting long–short portfolio earns a significant six-factor alpha of 101 basis points per month during the first year after the IPO. These abnormal returns vanish as return histories accumulate, consistent with markets gradually learning firms’ true systematic risk exposures.