Using LLMs to Write Financial Research Reports: Capability Evaluation and Prompt Engineering Practices

Over the past two years, the capability boundaries of large language models (LLMs) have kept advancing into professional domains, and financial research is among the first fields to feel the impact. The work behind a research report on a publicly listed company can be decomposed into three steps: extracting key information from a large volume of announcements, financial statements, and industry materials; performing causal reasoning and logical argumentation; and producing a well-structured, rigorously written report. This maps almost exactly onto the capability chain LLMs are best at: information compression, logical reasoning, and structured writing. This article takes “having an LLM write a research report on a listed company” as the target task, walks through the full process—from capability evaluation and prompt design to output structure and quality validation—and provides a prompt template you can reuse directly.

1. Evaluate Before You Use: Does the Model Really “Understand Finance”?

Before letting a model write a report, the first question to answer is: how good is it, really, at financial tasks? A high score on a general benchmark does not mean the model passes in a vertical domain—strong general conversational ability says nothing about the accuracy of its financial interpretation. The prudent approach is to build a small in-house evaluation around five dimensions:

1. Information extraction: Give the model a raw, unedited earnings summary or announcement text and check whether it accurately extracts key metrics such as revenue, gross margin, and operating cash flow, and correctly maps them to the right business segments.
2. Financial interpretation: Ask the model to explain why gross margin declined. Does it stay at the surface of the numbers, or can it distinguish among factors of different natures, such as cost-structure changes, intensifying competition, and one-off gains or losses?
3. Logical reasoning: Provide a set of public industry information and examine whether its argumentation on the competitive landscape and growth drivers forms a complete chain—or whether it contains circular reasoning or a disconnect between evidence and conclusion.
4. Methodology mastery: Test whether the model properly understands the underlying assumptions and applicable boundaries of mainstream methods such as discounted cash flow (DCF) and comparable-company valuation.
5. Expression standards: Is its terminology accurate? Is the structure clear? Can it consciously separate “statements of fact” from “subjective inference”?

For scoring, use multiple generations per task, have several reviewers score independently, drop the extreme scores, and take the average. This reduces score variance caused by randomness and makes different models comparable. One point, however, must be clear: scores are for initial screening only. Before any model is put into formal use, it should first pass a real, end-to-end task validation loop—which is exactly what the next section does.

2. The Five-Element Prompt: Shaping the Model into a “Senior Analyst”

The most common mistake in practice is tossing a one-liner like “write me a research report on company X” at the model. What comes back is usually a templated, generic piece with no informational value. To reach usable quality, a prompt should contain at least five elements: role setting, capability list, task description, hard constraints, and output structure. Here is a template validated in practice:

# Role
You are an industry researcher with more than ten years of experience,
focused on the technology and internet sectors, with expertise in
financial statement analysis, industry landscape assessment, and
valuation modeling.
# Capability requirements
- Skilled at extracting key information from raw source materials and
  strictly separating facts from inferences;
- Familiar with the assumptions and applicable boundaries of mainstream
  valuation methods;
- Rigorous in logic and accurate in terminology, balancing the reading
  efficiency of professional readers and management.
# Task
Based only on the public materials I provide, write a research brief on
a Hong Kong-listed internet company.
# Output structure
1. Investment highlights: three core theses, each consisting of
   "one factual statement + one inference";
2. Company overview and business model: revenue sources, customer
   structure, competitive moats;
3. Industry analysis: market size, competitive landscape, key growth
   drivers;
4. Financial analysis: three-year trends in revenue, margins, and cash
   flow, with attribution;
5. Valuation discussion: method used, key assumptions, and sensitivity;
6. Risk factors: at least three, each directly tied to the argumentation
   in the body of the report.
# Hard constraints
- Use only data contained in the provided materials; for undisclosed
  metrics, state "not disclosed in the provided materials";
- Every inference must come with a complete reasoning chain;
- The entire document is a research draft and does not constitute
  investment advice.

The design intent behind this template has four layers. First, the role setting injects domain priors into the model, which noticeably raises terminology density and the professionalism of the analytical perspective. Second, the capability list effectively declares the passing bar up front, and the model aligns toward that standard during generation. Third, the hard constraints are the key to suppressing hallucination—the rules “use only the provided materials” and “flag undisclosed metrics” pull the model back from free improvisation to evidence-based writing. Fourth, the fixed output structure makes the result engineering-ready: easy to post-process programmatically, iterate section by section, and divide among multiple collaborators.

3. Validating the Output: What to Check in Each of the Six Sections

Once the model produces output according to the template, human review cannot stop at “read it once and ship it.” Review section by section, because each section carries a different type of risk:

1. Investment highlights: Verify that every core thesis finds support in the body; watch out for “clickbait”-style conclusions that sound bold but say nothing.
2. Company overview: Check that the business model description is consistent with the source materials, and that the revenue structure has not been attributed to the wrong segments.
3. Industry analysis: This is the section where models most easily “fabricate plausible-looking numbers.” Every market figure must be traced to a source; anything untraceable must be deleted or explicitly flagged as uncertain.
4. Financial analysis: Every number must be checked against the original disclosure documents. Pay particular attention to consistency in units, accounting periods, and the basis on which each metric is calculated.
5. Valuation discussion: The passing standard is not “it produced a target price” but “it clearly laid out the assumptions and sensitivities.” The real value of a valuation is the logical framework it presents—not a seemingly precise number.
6. Risk factors: Reject boilerplate such as “macroeconomic volatility” or “policy changes.” Every risk must be directly relevant to the argumentation in the body and be trackable and verifiable.

4. Quality Risks: Three Reasons Humans Must Stay in the Loop

First, hallucination risk. An LLM can generate numbers and causal chains that look perfectly reasonable—the more fluent the text, the more vigilant you should be. Checking every claim against the source materials is the only reliable line of defense. Second, timeliness risk. A model’s parametric knowledge has a cutoff date, while industry and company fundamentals change quickly. In production, use retrieval-augmented generation (RAG) to feed the model the latest announcements and news rather than relying on the model’s “memory”—otherwise it may draw conclusions from outdated information. Third, accountability. Research reports directly influence investment decisions, so model output must be positioned, in every scenario, as a draft or auxiliary material. Whoever signs off on the report is responsible for all of its content. This is both a compliance requirement and the professional baseline of the research industry.

5. Engineering It: From One-Off Conversations to a Research Pipeline

Single conversations are fine for validating feasibility. To scale up, consolidate the process into a stable pipeline: collect announcements and financial reports into a repository on a regular schedule (backed by object storage and document-parsing services); chunk the text and build a retrievable knowledge base; at the report-generation stage, use retrieval augmentation to give the model the latest context; generate each section separately, then assemble and polish the full draft; finally, close the loop with a human review step for quality gatekeeping and audit trails. On an enterprise cloud platform such as Baidu AI Cloud, these capabilities can be orchestrated into reusable low-code workflows, and unified permission management and auditing mechanisms can meet the compliance requirements of financial institutions. Once the workflow is engineered, the efficiency gains shift from “individual experimentation” to “team capacity,” freeing researchers to concentrate on high-value work such as tracing data sources, challenging the logic, and making the final judgment.

6. Conclusion

The real value of having an LLM write research reports is not replacing the analyst, but freeing people from information shuffling and first-draft writing. The practical lessons can be condensed into one sentence: use evaluation to screen models, use constraints to suppress hallucination, use structure to guarantee quality, and use human review to hold the line. Get these four right, and an LLM moves from “sounds smart in a chat” to “reliable in actual use”—a research assistant that delivers stable, controllable, and sustainable value in financial research.