Reading a Competitor's Talent Outflow From Public Departures

You need to judge whether a competitor is quietly losing the people who matter, using only public signals, before you brief leadership or plan a poach. This guide is for strategy and research teams, talent-intelligence analysts, and operators sizing a rival's exposure. It carries one worked target end to end so a solo analyst without a paid talent-insights subscription can reproduce the read and defend every number in it.

The library already has guides that reconstruct a competitor's current org chart or read its roadmap from open roles. This one does something different: it traces people leaving over time. The output is a dated departure list, a public-signal attrition rate for one function, and a pattern label with a stated confidence tier. That is a finding you can put in front of leadership, not a vibe.

What "talent outflow" actually means here, and why the public read is inverted

Talent outflow is the rate and shape of people leaving a company over a defined window, read from dated public departures rather than from an HR system. The job is not to count exits. It is to turn exits into a rate, then classify the shape of that rate.

The formula is not in dispute. Separations divided by average headcount, times 100. Average headcount is the count at the start of the period plus the count at the end, divided by two. A rolling 12-month calculation is the most accurate frame because it smooths seasonal anomalies and short-term fluctuations. The paid tool that dominates this space defines attrition identically: departures in the past 12 months divided by the average number of employees during that period.

Here is the trap that separates a public read from an HR read. An HR system knows headcount and counts exits directly. You see the exits from public profiles, but you have to estimate the headcount yourself. The numerator is public. The denominator is not. That inversion is the whole difficulty of the job, and it is why a sloppy analyst reports a raw count of eleven departures as if it were meaningful. Eleven out of forty is a crisis. Eleven out of four hundred is Tuesday.

Throughout this guide I will carry one target: the backend engineering function at a mid-size payments company, read over a rolling 12 months. Substitute your own target as you go. The queries, the intermediate counts, and the forks are the same regardless of which company you point them at.

Estimating the denominator you cannot see

The denominator is a function-level headcount estimate, and it is the single most fragile input in the read. Build it from at least two independent signals and state your uncertainty out loud.

For the worked case, backend engineering, I have three ways to bound the number. First, count current profiles that list the target as employer and carry a backend or infrastructure title. Second, cross-check against a total company headcount signal and apply a plausible engineering share. Third, calibrate against pool sizes for the same skill in a known geography. Refolk's index is useful for that third calibration because it gives you an absolute scale for a skill, not a percentage.

Dataset A - Refolk index pool sizes (denominator reference for a departure read)

These counts come from Refolk's index of professional profiles. The ratios are mine. Note the caveat that matters: the third row uses a title and seniority filter, not a skill filter, so the 65% is directional only. Do not treat it as a clean subset. The point of the table is scale. A US JavaScript engineering pool near 97,000 tells you that a single target company's backend function is a small, countable slice, which means your headcount estimate can be built profile by profile rather than guessed.

When you cannot count profiles cleanly, bracket the denominator with a low and a high estimate and carry both through the arithmetic. A rate of "somewhere between 9% and 16%" is honest. A rate of "13.4%" built on a headcount you invented is not.

Reading the benchmark: what "elevated" means for engineering

Elevated means materially above the tech software benchmark of roughly 13%, not above the 10.9% global all-company average. Engineering churns faster than the broad economy by design, so the benchmark you anchor to decides whether your finding is real or manufactured.

Dataset B - Annual attrition benchmarks (comparison for the "elevated" threshold)

The tech software sector came in at 13.2% in an analysis of more than half a billion member profiles, against a 10.9% global average. US voluntary attrition averaged 13% in Mercer's 2025 survey of 2,617 organizations, down from a Great Resignation peak of 24.7% in 2022. For a monthly frame, the JOLTS voluntary quits rate for October 2024 was 2.1% across all industries, with the highest at 4.3% in food and accommodation and the lowest 0.5% in federal government.

The reason to anchor to 13% and not 10.9% is tenure. Tech tenure often runs 2 to 3 years against 4.1 years economy-wide. The same absolute exit count that alarms in a stable sector is routine in engineering. If you benchmark a payments company's backend team against the global average, you will report an exodus that is simply the sector's normal metabolism.

The same exit count that alarms in a stable sector is routine in engineering; anchor to tech, not to the world.

So for the worked case: if backend engineering shows an annualized rate near 13%, that is normal churn. Around 18% to 20% is elevated and worth a diffuse-versus-targeted test. Above the 24.7% peak territory, something structural is happening and you should already be hunting for a shared destination.

The dating problem: turning imprecise profiles into a timeline

Dating an exit means converting an imprecise or missing end date into a bounded departure window with a confidence flag. There is no published error-rate study for dating a single profile, so treat precision claims with suspicion and record uncertainty explicitly.

What is documented is that profiles lag. People are coached to update their profile about two weeks after starting a new job, and often far later. One professional body notes there is no rule on when to record a departure date, and that six months is approaching the limits of that consideration. Worse, candidates are coached to keep the last role listed as "Present" and put the actual end date after the job title, which produces a profile that reads as still-employed when the person left months ago.

Use this dating procedure, in order:

• Month-only end date: assume mid-month.
• Year-only end date: assume mid-year and widen the confidence window to plus or minus six months.
• No end date but a new employer present: use the new role's start date as the departure ceiling.
• No end date and no new employer: flag low-confidence and do not date it. It is not evidence yet.

The decay data tells you how much rot to expect in the underlying profiles.

Dataset C - Public-signal dating and decay inputs

B2B contact databases decay at 2.1% per month, compounding to 22.5% per year. Platform market intelligence itself lags several weeks to a few months behind the current state of the market. Both facts push the same way: the most recent quarter is systematically undercounted, because recent exits have not yet surfaced in profiles. A "quiet" recent quarter can be the noisiest one. Leave the most recent 8 to 12 weeks marked provisional in every read.

This is where the friction is worst for a solo analyst, and where a sourcing tool earns its place. Instead of scrolling profiles one at a time to find who left and where they landed, you can ask for the cohort directly and get dated, destination-tagged results back.

Refolk removes the per-profile grind of the departure list and the dating pass, so the analyst spends time on confounder clearance and classification, which is where judgment actually lives.

The procedure, end to end

This is the full read, from a named target to a one-page finding. Each step names who does it and what "done" looks like so you can hand it off or resume it later.

Applied to the worked payments case, step two might return 22 backend departures inside 12 months. Step three drops 4 to low-confidence for missing dates. Step four removes 3 internal relabels and 2 stale profiles, leaving 13 corroborated exits. Against an estimated average backend headcount of 85, that is a 15.3% rate. Above the 13% tech benchmark, below panic territory, and now the classification step decides what it means.

Classifying the pattern: churn, diffuse loss, or collapse

The pattern is decided by three variables: clustering in time, shared destination, and a leader link. Normal churn scatters; a targeted collapse clusters around a departed leader heading to one or two places.

No public source defines a numeric threshold for a collapse, so the following is an analyst-derived working rule, not a sourced standard. Treat 3 or more people from one team or function, joining the same or two destination companies, inside a 90-day window, as a candidate collapse. Scattered destinations over a full 12 months read as diffuse churn. The acqui-hire literature supports the shape: an acqui-hire secures a cluster of scarce talent, especially when the team is already committed to working together, and when founders leave the acquirer's organization, other team members tend to follow them out.

That last fact is the key one. Collapse detection hinges on the leader, not the count. A shared destination plus a departed manager is more diagnostic than raw exit volume. Two engineers independently joining the same large employer is noise. Two engineers joining the same 15-person startup that their former staff engineer founded is a signal.

For the worked case, the 13 corroborated backend exits scatter across nine destination companies with no departed lead engineer. That is diffuse loss, elevated but not targeted. If instead 5 of them had landed at one seed-stage startup founded by the former backend lead, the same 13 exits would be a targeted collapse and a very different brief.

The three-label output is the deliverable:

• Normal churn: at or below the tech benchmark, scattered. No action beyond monitoring.
• Elevated diffuse loss: above benchmark, scattered, no leader link. Worth understanding the cause, weak poach signal.
• Targeted collapse: clustered, shared destination, leader link. Strong signal for both a competitive read and a poach plan.

How this read goes wrong

Every failure mode here manufactures a finding that is not there, or hides one that is. This section is the most valuable part of the guide because the errors are subtle and they all look plausible in a spreadsheet.

Internal relabel read as exit. The title changed but the company did not. A promotion gets counted as a departure. Since 25 to 35% of job titles change annually in a typical database, an analyst keying on title changes rather than employer changes can fabricate a 25-plus percent exodus that never happened. Always check the employer field, not the title.

Divestiture or M&A mass move. A whole cohort "leaves" on the same date to one new entity. A spin-off gets scored as a collapse. Between 5 and 10% of companies undergo major structural change annually, so this is not rare. Cross-check any tight cluster against dated deal news before you call it.

Stale profile, still employed. A person left but never updated, or updated late. This understates your numerator, and the reverse, people who left but never updated, inflates a phantom "still there" headcount. Corroborate exits with a new-employer start date.

Over-annualizing a short window. Multiplying a hot quarter by four projects a normal function into an exodus. One quarter does not make a trend; small samples are volatile. Track at least two consecutive quarters before concluding you have a systemic pattern.

Denominator guessing. No headcount estimate, so raw counts masquerade as a rate. Confirm you divided by average headcount rather than reporting a count.

Coincidental shared destination. Two engineers independently join the same large employer and you call an acqui-hire on noise. Require 3 or more from one team in a tight window plus a leader link.

Profile-lag blind spot. Reading the current month as complete when data lags several weeks to a few months. Leave the most recent 8 to 12 weeks provisional.

The funnel is the discipline made visible. The gap between 22 and 13 is the difference between a defensible finding and an inflated one. If you skip the middle stages, you report 22 over 85 as a 26% exodus and brief leadership on a crisis that is really a 15% elevated-but-normal read.

Confidence tiers and the pre-report checklist

Confidence is a stated tier, not a feeling, and it reflects corroboration depth and profile-lag exposure. State it explicitly so leadership knows how hard to lean on the read.

Use three tiers. High means every exit has a second corroborating signal, the denominator is built from two independent estimates, no confounder is unresolved, and the window excludes the provisional recent quarter. Medium means most exits are corroborated but the denominator is a single-source bracket, or the recent-quarter provisionality is material. Low means the departure list leans on profile edits alone, the denominator is a guess, or a confounder could not be ruled out.

There is no formal published standard for the minimum evidence before you report, so this checklist is assembled from the data-quality and attrition literature. Run every item before the read leaves your desk.

TARGET: [company], [function]
WINDOW: [start date] to [end date], rolling 12 months
RATE: [numerator] exits / [denominator] avg headcount = [X]% ([annualized? yes/no, caveat])
BENCHMARK: tech software ~13.2%; this read is [at / above / well above]
PATTERN: [normal churn / elevated diffuse loss / targeted collapse]
LEADER LINK: [none / named departed lead and destination]
CONFOUNDERS CLEARED: [relabels removed], [M&A checked], [stale flagged]
PROVISIONAL: last [8-12] weeks incomplete due to profile lag
CONFIDENCE: [high / medium / low] because [reason]
SO WHAT: [poach signal / monitor / no action]

Keeping the read current

A talent outflow read is a snapshot that decays, so schedule a re-check rather than treating one pass as permanent. Because profiles lag weeks to months and contact data decays around 2.1% per month, a read that is six months old is describing a company that no longer exists in the same shape.

Re-run the full procedure quarterly for any target on an active watchlist, and confirm the earlier finding held: one quarter does not make a trend, so a single elevated read becomes a systemic conclusion only after two consecutive quarters point the same way. On the re-run, pay attention to the previously provisional recent quarter, because the exits that were invisible last time will now have surfaced, and that back-fill is often where a diffuse read tips into a targeted one. Ask the departure cohort directly on each pass, grouped by new employer, so a forming collapse shows up as a shared destination before it is obvious to anyone reading headlines.

Sources genuinely disagree on whether to annualize a short window. Some warn it over-projects small samples into false exoduses; others accept it as a standard normalization. Carry both views in the finding rather than picking one silently, and let the reader see the raw window number alongside the annualized one. That honesty is what makes the read defensible when leadership pushes back, and it is what separates this standard from the paid-dashboard answer that hands you a single number with no visible seams.