Deal economics
Data supply operational margin drivers: test three on matched cohorts
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Three supply-side metrics drive data licensing margin: qualification accuracy (the share of qualified suppliers with acceptable data), preparation hours per dataset and first-review acceptance. In an illustrative model, ten points of accuracy add 10.7 margin points, against 5.6 for ten percent fewer preparation hours and 4.3 for ten points of acceptance. More applications at weaker fit cut margin 12.0 points with unlimited staff hours and 14.8 inside a fixed quarter.
Value drivers
Data supply operational
Key takeaways
- Qualification accuracy, preparation hours per dataset and first-review acceptance explain supply-side margin; application counts explain little on their own.
- A false pass consumes its full preparation and returns nothing, which is why ten points of accuracy moved illustrative margin nearly twice as much as a ten percent cut in preparation hours.
- Fifty percent more applications at weaker fit lowered margin 12.0 points with no capacity limit and 14.8 points inside a fixed 3,400-hour quarter.
- In the illustrative bridge, a shift toward easier single-system datasets earned 55.8 percent of a 13.6-point margin gain, so a pooled margin overstated process improvement.
- A metric predicts margin only if it is measured before preparation starts, within matched dataset types, and holds on a later cohort.
Which supply-side metrics actually drive data licensing margin?
Three metrics do most of the work: qualification accuracy, preparation hours per prepared dataset and first-review acceptance. Each one changes how many staff hours it takes to produce one accepted dataset, and staff time is the cost a supply operation actually controls. Application volume, supplier sign-ups and qualified counts reach margin only by moving one of those three.
The decision for an investor is which numbers to request when a data licensing company claims operating leverage, and how to tell a real driver from a metric that simply moves with the mix of work. The evidence that would settle it is specific: complete supplier cohorts traced from application through qualification, preparation and acceptance, with every staff hour counted, including the hours spent on suppliers who never produced an accepted dataset, and compared within matched dataset types.
No data licensing company, SourceX included, publishes cohort data of that kind. So this article builds the mechanism as a transparent model with every input labeled Illustrative, shows which metric moves margin most and why, tests the strongest objection, and sets out the measurement design an investor can ask a company to fill in with its own records.
The margin identity: where a supplier cohort's staff hours go
Margin after supply-side staff time is the contribution earned on accepted datasets, minus every hour spent producing them, divided by that contribution. The hours land in four piles: screening each application, preparing each qualified supplier's dataset, reworking datasets that miss first review, and the preparation that false passes consume and never repay.
False passes are the expensive pile. A supplier qualified on an answer that later proves wrong, such as an export that loses the links between records or rights that stop short of the records offered, still receives full preparation before the problem surfaces. ASQ's cost-of-quality framework names the two places such failures show up: internal failure costs arise when work misses the standard and the miss is caught before transfer to the customer, external failure costs when it is caught only afterward. A false pass found at internal review is the first kind; a dataset a buyer rejects is the second.
The model divides each quarterly cohort into two dataset types that behave differently: single-system histories drawn from one ticketing or job system, and multi-system histories that must be linked across tools. That split is the matching step. Every comparison below holds type constant, because a cohort that drifts toward easier work can look more efficient without any process changing. The table defines the five quantities an investor needs, with the denominator for each:
| Metric | Definition and denominator | When it can be observed | How it reaches margin | How it gets flattered |
|---|---|---|---|---|
| Qualification accuracy | Qualified suppliers whose dataset can pass acceptance once prepared (accepted, or rejected only for preparation defects), over all qualified suppliers, open files kept in the denominator | After preparation; earlier through a blind audit sample of qualified files | Each false pass burns its full preparation hours | Dropping withdrawn or open files from the denominator |
| Preparation hours per prepared dataset | Every preparation hour logged to the cohort, over datasets prepared, including those later rejected | During preparation | Multiplies across every qualified supplier | Shifting work onto suppliers or into a later rework step |
| First-review acceptance | Viable datasets accepted at first internal or buyer review, over viable datasets reviewed | At first review | Each miss adds rework hours and loses some datasets | Loosening the review, or reviewing easy datasets first |
| Application volume | Complete applications in the cohort, one per company | At intake | Adds screening hours at once; adds accepted datasets only at unchanged fit | Counting submissions or duplicate companies |
| Mix | Share of qualified suppliers by dataset type | At qualification | Moves average hours per dataset with no process change | Reporting one pooled margin across types |
The driver test: one step at a time, inside a fixed quarter
Qualification accuracy moved margin the most. From an illustrative baseline margin of 36.0 percent, ten more points of accuracy lifted it to 46.6 percent, while ten percent fewer preparation hours reached 41.5 percent and ten more points of first-review acceptance reached 40.3 percent.
The baseline cohort is 200 complete applications in a quarter, 60 qualified suppliers split evenly between the two types, 1.5 screening hours per application and 3,400 staff hours of capacity. Single-system datasets start at 60 percent accuracy, 30 preparation hours and 75 percent first-review acceptance, and each accepted one contributes the equivalent of 150 staff hours. Multi-system datasets start at 40 percent, 60 hours and 60 percent, and contribute 260. A viable dataset that misses first review gets 12 hours of rework and then passes half the time.
Accuracy wins because of where the waste sits. Raising it turns false passes into accepted datasets without adding a single preparation hour: total hours rise only by a little extra rework, while accepted datasets rise by about a fifth. Cutting preparation hours helps every file, including the false passes it still prepares; acceptance gains only touch viable files that would have needed rework.
Per step, each point of accuracy returned 1.27 margin points, each one percent of preparation hours removed returned 0.56, and each point of first-review acceptance returned 0.45. The units differ, so the useful comparison is price: what a point of accuracy costs a company to buy, usually reviewer time spent before preparation starts, against what a percent of preparation time costs in tooling or training. Changes are computed before rounding, so a change can differ by 0.1 from the gap between two rounded margins.
Each run below was computed by the script with the capacity limit applied (Illustrative):
| Run (Illustrative) | Accepted datasets | Staff hours | Hours per accepted dataset | Margin after staff time (change) |
|---|---|---|---|---|
| Baseline | 25.4 | 3,111.6 | 122.7 | 36.0% |
| Qualification accuracy +10 points | 30.4 | 3,135.0 | 103.2 | 46.6% (+10.7 points) |
| Preparation hours -10% | 25.4 | 2,841.6 | 112.1 | 41.5% (+5.6 points) |
| First-review acceptance +10 points | 26.9 | 3,075.6 | 114.5 | 40.3% (+4.3 points) |
| Applications +50%, weaker fit | 21.5 | 3,400.0 | 158.5 | 21.1% (-14.8 points) |
| Applications +50%, same fit | 26.6 | 3,400.0 | 127.8 | 33.3% (-2.7 points) |
Do more applications raise accepted supply or only expense?
More applications raise expense first and raise accepted supply only when fit holds and capacity exists. In the model, 100 extra applications at weaker fit add 20 qualified suppliers, all multi-system and only a quarter of them viable, and margin falls 12.0 points to 24.0 percent even with no limit on staff hours.
Inside a fixed quarter the result is worse. Every application is screened first, so the extra 150 screening hours and the weaker files' preparation crowd out work on good files: the equivalent of 21.5 qualified suppliers carry over unworked, accepted datasets fall from 25.4 to 21.5, and margin drops 14.8 points to 21.1 percent. The company is busier and produces less.
The control run separates volume from fit. The same 100 extra applications at the base cohort's fit leave hours per accepted dataset at 122.7 when capacity can stretch, and cost 2.7 points within the fixed quarter, because 27 qualified suppliers wait while every application is still screened. Volume alone is not the problem; volume at weaker fit, arriving ahead of capacity, is.
When a company reports growth in applications, ask for qualification accuracy by intake channel and month, and for the hours spent on suppliers who qualified but never reached acceptance. If accuracy fell in the months volume rose, the growth bought cost rather than supply.
Is the margin gain operational, or did the mix get easier?
Often the mix got easier. In the model's bridge, margin rose 13.6 points between a first and a fourth quarter, and a shift toward single-system datasets explains 7.6 of them, 55.8 percent of the gain; accuracy, preparation and acceptance together explain the other 6.0.
In the fourth quarter, single-system suppliers rise from 30 to 42 of the 60 qualified, and on those files accuracy reaches 65 percent, preparation falls to 27 hours and first-review acceptance climbs to 80 percent. Multi-system files keep their first-quarter values. Applications, screening hours and capacity stay the same.
Any bridge has to split one change among several causes, and the split depends on the order in which causes are switched on, because they interact: switched first, mix is worth 6.8 points; switched last, 8.3. The table uses the Shapley value, which averages each factor's contribution over all 24 possible orders, so the four parts sum exactly to the total. Shorrocks (2013) set out this approach as a unified framework for decomposition in distributional analysis.
The matched view tells the same story. Within single-system files, margin rose from 53.3 to 61.8 percent; within multi-system files it stayed at 19.6 percent. The process improved only on the easier work, which matters before anyone extrapolates the pooled gain to a pipeline that could tilt back toward harder datasets.
Pooled rates can mislead badly when groups differ. Bickel, Hammel and O'Connell's 1975 study of Berkeley graduate admissions in Science found that an aggregate gap between groups of applicants was misleading, because applicants were spread unevenly across departments that were harder or easier to enter, and few departments showed the gap on their own. Dataset types play the role of departments here.
The bridge, computed by script (Illustrative), with the credit each factor would get if it were switched first or last:
| Factor switched from Q1 to Q4 values | Shapley contribution | If switched first | If switched last | Share of the 13.6-point gain |
|---|---|---|---|---|
| Mix toward single-system datasets | 7.6 points | 6.8 points | 8.3 points | 55.8% |
| Qualification accuracy | 2.6 points | 2.4 points | 2.8 points | 19.3% |
| Preparation hours | 2.2 points | 1.9 points | 2.4 points | 15.9% |
| First-review acceptance | 1.2 points | 1.1 points | 1.3 points | 9.0% |
How to tell a metric that predicts margin from one that tracks it
A metric predicts margin when it is measured before the hours are spent, compared within a matched dataset type, and still related to margin on a later cohort that played no part in choosing it. First-review acceptance arrives too late to steer a cohort; accuracy can be made an early signal with an audit sample. The design below fills that requirement and works for any company's records:
At a glance
How to tell a metric that predicts margin from one that tracks it
Checklist · 6 items
- Cohort: every complete application in a fixed intake window, dated by completion, with withdrawn, rejected and still-open files kept; open files are censored at the measurement date, never dropped.
- Strata fixed in advance: number of source systems, years of history and requested record type, with the count of files in each stratum reported beside every rate.
- Accuracy before preparation: a blind second review of a random sample of qualified files, scored later against acceptance, plus a smaller sample of rejected files to find good suppliers turned away.
- Hours by stage: every logged hour on the cohort, including hours on suppliers who never produced an accepted dataset, split into screening, preparation, rework and support.
- Out-of-time check: fix each metric's definition on one cohort, then test its relationship with margin on the next cohort within the same strata.
- Gaming check: when a metric becomes a team target, track the measure it could be traded against, such as accuracy rising while qualified counts in hard strata collapse.
0 of 6 ticked. Tick each item as you go.
What could make the margin-driver thesis wrong
Five things could break it: value per dataset may matter more than hours, capacity may stretch cheaply, mix may be strategy rather than noise, accuracy may be unobservable in practice, and the ranking itself rests on chosen inputs. Each is a reason to ask for evidence, not to dismiss operating metrics.
- Value can swamp cost. The model fixes contribution per accepted dataset by type. If buyer demand or the rights a deal grants move value per dataset by more than operations move hours, operating metrics become secondary, so ask for value per accepted dataset by type next to hours.
- Capacity may stretch. If a company can add trained reviewers quickly at the same cost per hour, the within-quarter penalty on extra volume shrinks toward the same-fit case.
- Mix may be a deliberate choice. A company that favors single-system supply because buyers ask for it is not hiding anything; the issue is only whether a pooled margin is presented as process improvement.
- Accuracy is hard to see. Rejected suppliers are never prepared, so good suppliers turned away stay invisible without an audit sample, and a company can report high accuracy by qualifying only safe cases.
- The ranking follows the inputs. A cohort with cheap preparation and costly screening could reorder the drivers, which is why the script and its inputs are published with the method.
Margin-driver questions for any data licensing company, SourceX included
Ask for cohort evidence in the company's own definitions, then restate it in the definitions above. SourceX belongs on the same list. Its how-it-works page says that during qualification "SourceX checks fit, completes the inventory with you, runs the rights check and sets the processing route", and that qualifying is not a promise of an offer; the gap between qualification and a dataset accepted in writing at Closing is where its accuracy would show. Its application is a short fit check that asks for no files and transfers nothing, and its screening guide lists rights to license and an authorized sponsor among the traits of a good candidate, so most of the accuracy question is settled after intake rather than at it.
The same page names the events a cohort ledger would need: a signed Dataset Deal Confirmation that fixes the exact dataset for each deal, preparation and delivery, written acceptance at Closing, or a no-deal notice when a deal does not close. SourceX does not publish cohort, hours or margin figures; according to its investors page, company and commercial metrics are shared privately with investors. SourceX is raising from accredited investors under Rule 506(c), and its investor materials are available on request through that page. The questions below work for SourceX and for any other data licensing company:
- What share of qualified suppliers in each quarterly cohort reached an accepted dataset, by dataset type, with open files counted?
- How many preparation hours went to suppliers who qualified but never reached acceptance, and what failed in those files?
- Do you audit a sample of qualified and rejected files before preparation starts, and how did audit verdicts compare with later acceptance?
- How did the mix of dataset types change over the last four cohorts, and what was margin within each type?
- In quarters when applications rose, what happened to accuracy and to hours per accepted dataset in the same cohort?
- Which metric carries a team target, and what stops it from being met by narrowing what you qualify?
Methods note: model rules, inputs and what was not measured
The model, the driver test, the bridge and the measurement design were built for this article on 2026-10-28 as an editorial model. A short Python script (xs0200_margin_drivers.py, kept with its output in the writer's working folder) computes every figure with exact fractions, rounds once, half up, at output and uses no randomness. All inputs are examples chosen to show the arithmetic; no SourceX number enters the model.
Rules as the script applies them. For each dataset type, viable datasets (those that can pass acceptance once prepared) equal qualified suppliers times accuracy, and every qualified supplier is prepared. A viable dataset passes first review at the acceptance rate; otherwise it gets 12 rework hours and passes with probability one half. False passes receive no rework and are lost after full preparation. Every complete application takes 1.5 screening hours. Margin is contribution (150 staff-hour equivalents per accepted single-system dataset, 260 per multi-system dataset) minus all hours, divided by contribution. Under the 3,400-hour capacity, screening comes first and the hours left cover preparation and rework for the same fraction of every type's qualified suppliers; the rest wait and add no hours or datasets that quarter. Every run uses this limit except the two volume runs labeled as having none.
Each driver step applies to both types at once, and the per-step sensitivities change one input by one point, or by one percent of preparation hours, for both types. The weaker-fit run adds 100 applications and 20 qualified multi-system suppliers at 25 percent accuracy with multi-system preparation hours and acceptance; the same-fit run adds the same 100 applications and raises each type to 45 qualified. The bridge switches four factors from first-quarter to fourth-quarter values (qualified counts by type, accuracy, preparation hours and acceptance) and averages each factor's change in margin over all 24 orders; the matched view assigns screening hours to each type in proportion to its qualified suppliers.
Public evidence: ASQ's cost-of-quality page, Shorrocks (2013) in the Journal of Economic Inequality and Bickel, Hammel and O'Connell (1975) in Science, each checked through a search excerpt on the publisher's own site, and SourceX's how-it-works, application, screening-guide and investors pages, fetched directly on 2026-10-28. Not measured: any real company's accuracy, hours, acceptance, value per dataset or margin; good suppliers wrongly turned away; withdrawals while suppliers wait; and buyer rejections after delivery.
Frequently asked questions
What is qualification accuracy for a data supplier funnel?
It is the share of qualified suppliers whose dataset proves acceptable once prepared, with rejections caused by preparation defects counted against preparation rather than qualification, and open files kept in the denominator. It differs from a qualification rate, which only counts how many applicants passed. A company can raise its qualification rate while accuracy falls, and that costs margin because every false pass is prepared in full.
Why can more supplier applications lower margin?
Every application has to be screened, and applicants added by a broader campaign often fit worse. In this article's illustrative model, 50 percent more applications at weaker fit lowered margin 12.0 points, and 14.8 points within a fixed quarter of staff hours, while the same volume at unchanged fit left hours per accepted dataset where they were.
What is a matched cohort in supply economics?
It is a comparison of suppliers who entered in different periods but share the traits that drive effort, such as the number of source systems and years of history. Comparing within those groups stops a change in the mix of work from looking like a process improvement.
How do you split a margin change among several operating metrics?
Switch each metric from its old value to its new one and record the change in margin. Because the answer depends on the order of switching, a Shapley attribution averages each metric's contribution over every possible order, so the parts add up exactly to the total change.
Does SourceX publish its supply-side margins?
No. SourceX keeps cohort, hours and margin figures off its public pages, and its investors page offers company and commercial metrics to investors on request. The definitions in this article can frame the same questions for any data licensing company.
Sources
- ASQ defines internal failure costs as arising when work fails to reach quality standards and is detected before transfer to the customer, and external failure costs as detected only after transfer. Source
- Shorrocks (2013), Journal of Economic Inequality 11, 99-126, sets out decomposition procedures for distributional analysis in a unified framework based on the Shapley value. Source
- Bickel, Hammel and O'Connell (Science, 1975) found aggregate Berkeley graduate admissions data showed a clear but misleading pattern that few individual departments showed. Source
- SourceX's how-it-works page describes qualification (fit check, inventory, rights check, processing route), the Dataset Deal Confirmation, written acceptance at Closing and the no-deal notice. Source
- SourceX's application page describes a short fit check that needs no files, passwords or API tokens and licenses or transfers nothing. Source
- SourceX's screening guide lists rights to license and an authorized sponsor among the profile of a good candidate. Source
- SourceX's investors page says it shares company and commercial metrics with investors directly and offers investor materials on request. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.