Industry Research and Benchmarking
Building a cited peer-set benchmark from public filings while keeping facts and inference strictly separated. Defining a peer set and metric definitions, extracting figures from filings, tagging every claim as fact (cited) or inference (reasoned), and validating that every source resolves. The discipline that keeps research memos from quietly fabricating a benchmark.
120 min
6
16
5
Learning Objectives
By the end of this chapter you should be able to:
- 1Select a defensible peer set and a single set of metric definitions, so gross margin, revenue growth, and R&D intensity are computed the same way across each company in the comparison.
- 2Extract each figure from the provided filings and cite the source line it came from, so no number in the benchmark is left unsupported.
- 3Separate fact from inference by tagging each claim as either a cited fact or a labeled, reasoned inference.
- 4Apply the comparability caveat where a disclosure differs, flagging a metric as not directly comparable rather than estimating a figure a company did not disclose.
- 5Verify that each citation resolves by opening the source and confirming the figure is actually there before the benchmark is relied on.
- 6Frame a cited benchmark for a reader, leading with the takeaway and surfacing the comparability caveat rather than handing over a raw table.
Part One: The Benchmark Request and What a Cited Benchmark Means. Section 1 of 6.
Part One · The Benchmark Request and What a Cited Benchmark Means
The Benchmark Request and What a Cited Benchmark Means
Part One
The Benchmark Request and What a Cited Benchmark Means
A partner wants to know how the company you cover stacks up against its peers, on one page, by Friday. The research is mostly reading and arithmetic, which makes it a strong candidate for AI. It is also where a fluent tool can quietly invent a number, so this module starts by defining the standard the memo has to meet.
The Friday benchmark
You are an equity associate covering the industrial components sector. A partner asks for a one-page benchmark of Meridian Components against three close peers: how their gross margins, revenue growth, and research spending compare, with a short read on what stands out. The filings are public, the metrics are standard, and most of the job is pulling figures out of documents and lining them up. On its face this looks like exactly the kind of task an AI tool should shorten from an afternoon to twenty minutes.
It usually can, but the risk profile is specific. The failure that matters in research is not a slow draft; it is a confident, well-formatted figure that is not actually in any filing, or a citation that points to a page where the number is not there. A benchmark that looks authoritative and contains one fabricated data point is worse than no benchmark, because a reader will act on it. So the value of AI here depends less on the draft and more on the discipline wrapped around it: extract each figure, verify each citation, and keep facts separate from the analyst's own inferences.
What good looks like: a cited benchmark
Before pointing a tool at anything, define the finish line. A cited benchmark has a few properties that a raw table does not. It rests on a clearly chosen peer set, companies close enough in business model and size that comparing them is meaningful. It applies one shared definition for each metric, so gross margin means the same thing for each company. Each figure is traceable to a source line, so a reviewer can check it. Each claim is tagged as either a fact or an inference, following the discipline of fact versus inference: a cited number is a fact, while a reasoned statement about what the numbers imply is labeled as the analyst's inference rather than presented as disclosed.
The workhorse behind the whole approach is a pattern worth naming: extract-verify-cite. You extract each figure from a source, you verify that the citation resolves by opening the source and finding the number, and you cite it so nothing floats free of evidence. That order matters. Extraction without verification is where invented figures survive, and a citation attached to an unverified number is worse than no citation, because it borrows credibility the figure has not earned.
Two properties depend on judgment you set before drafting. The peer set is a choice: pick companies that are genuinely comparable, and be ready to say why. The definitions are a choice too: R&D intensity, for example, only compares cleanly when each company discloses R&D the same way. Deciding both up front turns "benchmark these companies" into a task the workflow can actually check.
Check Your Understanding
Knowledge Check 1
Industry Research
An analyst is asked to benchmark a company against its competitors on gross margin, revenue growth, and research spending. Before extracting any figures, which pair of decisions most improves the quality of the comparison?