Normalising Counts Across Entities of Different Size

A company with 400,000 employees mentions "hiring" in job postings more often, in absolute terms, than a company with 4,000 employees, and that difference has nothing to do with which company is actually growing faster. Any raw count — job postings, app downloads, patent filings, store visits — is contaminated by the sheer size of the entity it is counted for, and normalizing that count against a sensible denominator is usually the single most consequential step in turning it into a usable cross-sectional signal.

Picking a denominator

A raw count $n_c$ for company $c$ becomes a comparable rate once divided by an appropriate scale variable $s_c$:

$$r_c = \frac{n_c}{s_c}$$

In words: the rate is the count divided by something that captures the company's underlying size on the same dimension the count measures — job postings per existing employee, app downloads per existing user, patent filings per R&D dollar spent. The wrong denominator (say, dividing job postings by market cap instead of headcount) can distort the ranking just as badly as using no denominator at all.

Worked example

Company A has 250,000 employees and 3,000 open job postings this month. Company B has 8,000 employees and 200 open postings. On raw count, A looks like the more active hirer, 3,000 versus 200. Normalized by headcount, A's rate is $3{,}000 / 250{,}000 = 0.012$, or 1.2% of its workforce; B's rate is $200 / 8{,}000 = 0.025$, or 2.5% — more than double A's rate. B is the company hiring intensively relative to its own size, which is closer to what a signal chasing "companies about to grow headcount fast" actually wants to capture.

What this means in practice

Test more than one candidate denominator before committing — headcount, revenue, and R&D spend each answer a slightly different question, and the "best" one depends on what economic story the signal is meant to capture, not on which produces the prettiest backtest.