Job Posting Data As A Growth Signal

When to Use

Use this skill when building fundamental growth features for equity factor models from web-scraped job posting datasets (LinkUp, Coresignal, Greenhouse-sourced panels). Corporate hiring is a costly, revealed-preference capital allocation decision, and the peer-reviewed evidence supports it as a forward-looking disclosure: Gutiérrez, Lourie, Nekrasov and Shevlin (Management Science, 2020) find changes in online job postings are positively associated with future firm performance — one-year-ahead growth in headcount, sales and earnings — with a stronger relation when the postings represent growth rather than replacement hiring. That growth-vs-replacement distinction is what the role-mix weighting in this module approximates.

This module computes, for one company and one observation period: QoQ active-posting growth, a role-mix multiplier that up-weights Engineering/R&D and Sales openings, a configurable stale-listing haircut that shrinks the score toward neutral when postings look like ghost listings, a small-base gate, and a normalized Corporate Expansion Score in $[-1.0, +1.0]$.

When NOT to Use

• Not an expected-return forecast. EXPANSION_BULLISH describes the firm's trajectory, not the stock's. The accounting-based hiring-rate factor points the other way in the cross section: Belo, Lin and Bazdresch (JPE, 2014) find high hiring rates predict lower subsequent returns (roughly -1.5pp annual risk premium per +10pp hiring rate). Combine this score with valuation and risk factors; never wire it straight to an order router.
• Not point-in-time safe on its own. The engine takes two counts and has no knowledge-time axis. Feeding it vendor data stamped with the observation date rather than the delivery date leaks look-ahead into any backtest. Wrap the input with backtesting-alt-data-strategies-with-realistic-availability-lag first.
• Not a licensing or compliance control. Scraped posting data raises vendor contract, web-scraping and MNPI questions this module does not touch. See alternative-data-vendor-due-diligence-checklist and insider-trading-controls-for-alternative-data-usage.
• Not a cross-vendor comparator. Aggregator-sourced feeds carry duplicate listings that direct-from-career-site feeds do not. Two vendors report different counts for the same company, so a series spliced across vendors — or across a panel-coverage change — measures the vendor, not the company.
• Not valid on thin coverage. Sub-min_previous_postings bases produce arithmetic, not signal: 2 to 10 postings is +400%. Those snapshots return INSUFFICIENT_DATA.
• Not seasonally adjusted. See the pitfall below; QoQ on a seasonal hirer measures the calendar.

Prerequisites

• Company job posting snapshot (ticker, company_name, current_active_postings_count, previous_active_postings_count, engineering_postings_pct, sales_postings_pct, avg_posting_duration_days).
• Both counts drawn from the same vendor panel under the same point-in-time convention, with the vendor's publication lag already applied upstream.
• Role shares expressed as fractions in $[0.0, 1.0]$ (0.5 means 50%), summing to at most 1.0 — they are shares of the same posting count and cannot overlap.
• Calibration decisions for the four engine parameters. The defaults (120-day stale threshold, 0.5 haircut, 10-posting floor, 0.25 classification band) are this module's conventions, not an industry or regulatory standard; no external source prescribes them.

Workflow

1. Validate the Snapshot Before Scoring: calculate_growth_score calls snapshot.validate() first and raises JobPostingSignalError on negative counts, NaN/infinite values, shares outside $[0,1]$, role shares summing above 100%, or a blank ticker. This is deliberate: an unchecked NaN count clamps to -1.0 and is reported as a confident CONTRACTION_BEARISH, and nan > 120 is False, so a NaN duration would silently escape the ghost penalty.
2. Gate the Small Base: if previous_active_postings_count < min_previous_postings (default 10), classification is INSUFFICIENT_DATA and the score is 0.0. A zero previous count has no defined growth rate at all — the engine reports 0.0 rather than substituting a denominator.
3. QoQ Active Hiring Velocity: compute $\text{Growth}_{\text{pct}} = \frac{\text{Current} - \text{Previous}}{\text{Previous}} \times 100$. This value is always reported for audit, even when the base gate suppresses the score.
4. Strategic Role Mix Weighting: apply $W_{\text{role}} = 1.0 + (\text{Engineering share} \times 0.5) + (\text{Sales share} \times 0.3)$, so a 100%-engineering book carries $W = 1.5$.
5. Stale Listing Haircut: if avg_posting_duration_days is strictly greater than the threshold, multiply by $(1 - \text{haircut factor})$. Note the direction: this shrinks the score toward neutral in both directions, so a contraction of $-0.54$ becomes $-0.27$ and a $-0.40$ reading is demoted from bearish to neutral. Stale postings make the count less informative — they do not make the firm more bearish.
6. Score, Clamp and Classify: $S_{\text{growth}}$ is the clamp of the raw product into $[-1.0, +1.0]$.
   • $S_{\text{growth}} \ge +0.25 \implies$ EXPANSION_BULLISH.
   • $S_{\text{growth}} \le -0.25 \implies$ CONTRACTION_BEARISH.
   • Otherwise $\implies$ STABLE_NEUTRAL.
7. Rank on the Unclamped Score: the clamp saturates above roughly 75% QoQ growth, collapsing every fast grower onto exactly $+1.0$. For cross-sectional ranking use raw_growth_score (unclamped) or qoq_postings_growth_pct; $|\text{raw}| > 1$ marks a saturated reading and the audit note carries [SATURATED at clamp bound].
8. Audit Report Generation: output structured JobPostingSignalReport.

Full procedure: see references/workflows.md. Standards reference: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Reading EXPANSION_BULLISH as a Buy: the label describes hiring, not expected return, and the hiring-rate literature finds the opposite sign for returns. A backtest that longs the top expansion decile without a valuation control is testing a documented negative return predictor.
• A Mean Duration That Hides the Ghost Tail: 90% fresh postings plus 10% two-year-old ghosts averages about 100 days and never trips a 120-day threshold. The average is a blunt detector; where the vendor exposes a stale share or per-posting creation/deletion dates, prefer those and set the threshold accordingly.
• Haircutting a Hard-to-Fill Senior Req as a Ghost Job: Chen and Li (Review of Accounting Studies, 2023) find longer vacancy duration for high-skill roles is associated with higher future profitability, while fast fills signal strength for low-skill roles. A flat haircut on an engineering-heavy firm penalises exactly the case the evidence says is benign — the same firm this module's role weighting just up-weighted.
• The Haircut Quietly Rescuing a Shrinking Firm: because it shrinks toward neutral symmetrically, a firm cutting postings 40% with a stale book scores $-0.20$ and reads STABLE_NEUTRAL, not bearish. Check has_ghost_postings_penalty before trusting a neutral reading.
• Ranking on a Saturated Score: two firms at $+1.34$ and $+3.00$ raw both report $+1.0$. Any cross-sectional sort on corporate_growth_score silently ties the entire top of the distribution.
• Small-Base Arithmetic: a company going from 2 to 10 postings is +400% growth and would dominate any long book. Hence the base gate — do not lower it below the vendor's coverage floor for micro-caps.
• Unadjusted Seasonal Hiring Noise: QoQ comparison treats a retailer's Q4 holiday requisition spike as corporate growth. The engine is period-agnostic: pass the same quarter one year prior as previous_active_postings_count for seasonal hirers, or de-seasonalize upstream. Do not mix YoY and QoQ bases within one cross-section.
• Splicing Vendors or Panels: a vendor adding a company's regional career sites mid-history creates a posting-count jump indistinguishable from real hiring.
• Trading the Observation Date: posting counts are collected, deduplicated and delivered with a lag. Scoring on the observation date rather than the delivery date is look-ahead bias, not alpha.

Verification

• Audit Tech Expansion (current=300, previous=150 $\implies +100%$ growth, Engineering $= 50%$, Sales $= 30%$, Duration $= 30$ days): role factor $= 1 + 0.25 + 0.09 = 1.34$, raw score $= 1.34$, so the reported $S_{\text{growth}}$ is the clamp bound $+1.0$ (saturated, not $1.34$) and the classification is EXPANSION_BULLISH.
• Audit Stale Layoff Company (current=50, previous=100, Engineering $= 10%$, Sales $= 10%$, Duration $= 150$ days): role factor $= 1.08$, raw $= -0.50 \times 1.08 \times 0.5 = -0.27 \implies$ CONTRACTION_BEARISH with has_ghost_postings_penalty=True. Re-run at Duration $= 90$ days and confirm the score doubles to $-0.54$ — the haircut is symmetric.
• Assert the threshold is strict: Duration $= 120.0$ carries no penalty, $120.01$ does.
• Assert previous=0 and previous=2 both return INSUFFICIENT_DATA with a zero score, and that previous=10 is scored.
• Assert NaN counts, NaN durations, negative counts, a share passed as 50, and role shares summing above 1.0 all raise JobPostingSignalError rather than producing a score.
• Assert two saturating firms with raw scores $1.34$ and $3.00$ are separable by raw_growth_score while both report $+1.0$.
• Run python -m unittest discover -s skills/job-posting-data-as-a-growth-signal/scripts and confirm a 100% pass rate.

Related Skills

• app-download-and-usage-data-for-consumer-companies
• patent-filing-data-for-innovation-signal-research
• backtesting-alt-data-strategies-with-realistic-availability-lag
• alternative-data-vendor-due-diligence-checklist
• insider-trading-controls-for-alternative-data-usage