We uncover financial professionals’ mental models—the reasoning they use to explain their quantitative forecasts. We organize our analysis around a framework of top-down attention, where analysts endogenously choose both a valuation method and how to allocate attention across variables, based on each variable’s relevance for valuation and the cost of acquiring information about it. Using the near-universe of 1.6 million equity analyst reports, we collect the valuation methods analysts adopt to compute their price targets. We then prompt large language models (LLMs) on a subset of 110,000 reports to extract 4.8 million lines of reasoning—each combining a topic, valuation channel, time horizon, and sentiment. To validate the reliability of our output, we introduce a multi-step LLM prompting strategy and new diagnostic tools. We document four main findings. (1) Analysts exhibit sparse mental representations, focusing on a limited set of topics, that are primarily related to top-line items, and forward-looking. (2) The choice of valuation methods and topic focus is closely linked. (3) There is substantial disagreement among analysts, and differences in attention weights to firm-specific variables are a bigger source of disagreement than differences in valuation weights on those variables. (4) Lastly, variation in mental models aligns with key asset pricing patterns both in the time-series and in the cross-section.