Skip to content

Deal economics

Data licensing price uncertainty: how a comparable becomes a 5x range

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Data licensing price uncertainty is best reported as a range built from executed comparables, with every adjustment for rights, usable scope, preparation and buyer purpose stated as its own low and high factor. Adjustments multiply: four of plus or minus 20 percent stretch one comparable into a range whose top is five times its bottom. Investors should then ask how often signed prices landed inside the stated ranges.

Value drivers

Data licensing price

This page covers data licensing price uncertainty: how a comparable becomes a 5x range.Illustrative

Key takeaways

  • Report a data license price as a range built from executed comparables, with each adjustment's low and high factor written down.
  • Adjustment uncertainty multiplies: four factors of 0.8 to 1.2 stretch one comparable to 0.41 to 2.07 times its central value.
  • Count only executed transactions whose normalized range is no wider than two to one; keep asking prices as context, never as comparables.
  • Negotiations narrow the range only through evidence: a use statement, a rights review, an accepted sample and agreed preparation duties.
  • Test a company's ranges by coverage: with 20 signed deals, 12 or fewer prices inside stated 80 percent ranges means the ranges are too narrow.

What does data licensing price uncertainty mean for an investor?

It means a dataset's price is a range of defensible outcomes, and the width of that range is information. Before believing any figure, an investor should know which executed transactions it rests on, which assumptions convert those transactions to the dataset at hand, and how often the company's earlier ranges contained the price that was finally signed.

This article takes the negotiation side of the question. A company's internal valuation memo, the subject of our separate guide to dataset valuation confidence ranges, decides whether to explore a license at all. The question here starts once two parties are bargaining: how wide a defensible range is, which evidence narrows it, and what an investor can check after the deals have closed.

Useful evidence has three properties. It is a realized price, not an offer. Its rights, usable scope, preparation and buyer purpose are documented well enough to adjust. And it arrives with dates and denominators, so a reader can tell one deal from a pattern. A precise price per dataset that comes without any of the three is the first thing to question.

Four adjustments every comparable needs before it says anything

No two data licenses buy the same thing, so every comparable has to be converted before it is compared. Four conversions do most of the work: rights (exclusive or not, which uses, for how long), usable scope (the share of records that survive preparation and review), preparation (who cleans, links and de-identifies the records, and at whose cost) and buyer purpose (training, evaluation or a narrower internal use).

Each conversion is an estimate, so each carries its own low and high factor. A deal sold non-exclusively, with the seller doing the preparation, cannot be applied to an exclusive license of raw records without at least two factors, and neither factor is known exactly. Writing the factors down one by one lets the other side argue with a specific assumption instead of the whole price.

Even the buyer's own measurement of what the data is worth is noisy. Wang and Jia's Data Banzhaf paper (arXiv 2205.15466) reports that the noise in ordinary training is substantial enough to make different runs of the same data valuation algorithm produce inconsistent value rankings; in its ranking-stability test the average Spearman correlation between runs was about 0.038 for Shapley values and about 0.856 for Banzhaf values. The OpenDataVal benchmark (NeurIPS 2023) compared eleven valuation algorithms and concluded that no single algorithm performs uniformly best across all tasks. A price built on these measurements inherits their spread.

Why stacked adjustments multiply the range

Adjustment uncertainty compounds by multiplication, not addition. If each conversion could move a comparable anywhere from 0.8 to 1.2 times its central value, one open conversion gives a range whose top is 1.5 times its bottom. Two give 2.25, three give 3.38 and four give 5.06, a span from 0.41 to 2.07 times the central case.

That arithmetic is why a heavily adjusted comparable rarely settles a negotiation: each party can pick the end that suits it and still be defensible. One executed deal with the same rights and a single modest adjustment can carry more weight than three deals that each need four.

The table assumes the factors move independently. When one cause drives two of them, for example a buyer purpose that also changes how much preparation the records need, the real range can be narrower or wider than the multiplication suggests, and the worksheet should name the pair. The compounding, with each open adjustment between 0.8 and 1.2 (Illustrative):

Why stacked adjustments multiply the range
Open adjustmentsLow end (x central case)High end (x central case)Top-to-bottom ratio
None1.001.001.00
One0.801.201.50
Two0.641.442.25
Three0.511.733.38
Four0.412.075.06

The rights-normalized price range worksheet

The worksheet turns comparables into a negotiating range with two rules. An item counts only if it is an executed transaction, so asking prices and published ranges are listed but never counted. And it counts only if its normalized range is no wider than two to one once its adjustments are applied; wider items stay in the file as context. The negotiating range is where the counted ranges overlap.

Illustrative: in the example below, four pieces of evidence in value units (U) produce a range of 33.6 to 41.6 U, a top-to-bottom ratio of 1.24, with comparable A setting the floor and comparable D the ceiling. Comparable B, an executed deal that needs three adjustments, normalizes to 20.2 to 48.6 U and is held as context. Listing C, an asking price of 90 U, is excluded however encouraging it looks to a seller.

If the counted ranges did not overlap, the worksheet would report a conflict rather than average its way past it. A conflict is a finding: the comparables disagree about something the adjustments missed, and the next step is to find out what. The worksheet, with every value computed by the script described in the methods note:

The rights-normalized price range worksheet
Evidence (Illustrative)Stated value and typeAdjustments, low to high factorNormalized rangeTreatment
Comparable A42 U; executed license with the same exclusive training rightsUsable scope 0.8 to 1.033.6 to 42.0 U (1.25 to 1)Counts; sets the floor
Comparable B30 U; executed, non-exclusive, seller prepared the recordsExclusivity 1.2 to 1.8; term 0.8 to 1.0; preparation 0.7 to 0.920.2 to 48.6 U (2.41 to 1)Context only: wider than 2 to 1
Listing C90 U; asking price on a public listingNone applied90.0 U as listedExcluded: not a transaction
Comparable D26 U; executed, evaluation-only useBuyer purpose 1.15 to 1.6; usable scope 0.9 to 1.026.9 to 41.6 U (1.55 to 1)Counts; sets the ceiling
Negotiating rangeOverlap of A and DTwo counted items33.6 to 41.6 U (1.24 to 1)Midpoint 37.6 U; no conflict

How the range narrows as a negotiation produces evidence

A negotiation narrows the range only when it produces evidence, and each kind of evidence closes one assumption. A written statement of the buyer's use settles purpose, a completed rights review settles scope, a sample that passes the buyer's acceptance test settles the usable share, and an agreed split of preparation work settles that cost. Concessions without evidence move the price inside the range and leave its width alone.

In the illustrative ladder below, four open assumptions of 0.8 to 1.2 each start the talks at 0.41 to 2.07 times the central case, and the modeled band closes as each is settled. The signed price then picks one number. That number is an agreement, not a measurement: a deal that signs near the bottom of its range does not prove the range was wrong, though a company whose deals keep signing outside their ranges has a calibration problem.

SourceX's published process has the same shape without disclosing how it prices. Its how-it-works page says an indicative estimate comes first and that pricing stays indicative until a Dataset Deal Confirmation is signed; its FAQ says the exact payout and terms are fixed in that Confirmation, which a supplier can sign, question or decline.

Each step, the assumption it settles and the modeled band (Illustrative):

How the range narrows as a negotiation produces evidence
Negotiation stepAssumption settledEvidence that settles itModeled band (x central case)
Indicative estimateNone yet; four remain openInventory description and the comparables file0.41 to 2.07 (5.06 to 1)
Buyer states its use in writingBuyer purposeA signed use statement or a draft permitted-use clause0.51 to 1.73 (3.38 to 1)
Rights review completedRights scopeA review log naming what may be licensed and what is excluded0.64 to 1.44 (2.25 to 1)
Sample passes the acceptance testUsable shareThe buyer's written result on an agreed sample0.80 to 1.20 (1.50 to 1)
Preparation duties agreedPreparation costAn agreed split of cleaning, linking and de-identification1.00; the modeled band closes
A funnel chart of an illustrative price band narrowing across five negotiation steps, from 0.41 to 2.07 times the central case at the indicative estimate to a single value once preparation duties are agreed.
How evidence narrows a modeled price band. Illustrative: each open assumption spans 0.8 to 1.2 of its central case.

Are asking prices and published ranges a shortcut? Testing the alternative

No. An asking price records what a seller hoped for, and a published range records what a marketplace wants applicants to expect; neither records what a buyer paid after negotiating rights, scope and preparation. Official price statistics draw the same line. In its Handbook of Methods, the Bureau of Labor Statistics says the Producer Price Index asks for transaction prices net of discounts, premiums, rebates and allowances rather than list prices (paraphrased).

The strongest form of the alternative is that asking prices at least cap the range from above. They can cap the seller's expectations, which is why the worksheet lists them. They cannot cap the buyer's side, because no buyer agreed to them, and a published range for a whole category says nothing about the rights or preparation behind any single dataset.

An investor who sees list prices or category ranges presented as evidence of value should ask for the executed equivalents and for the adjustments that connect them to the dataset in question. If the company has none, the honest range is wide, and the honest pitch says so.

How investors can test whether a company's price ranges are honest

Ask for the company's history of stated ranges against signed prices, then test calibration the way forecasters do. The Federal Reserve's Summary of Economic Projections (June 2026) says the confidence interval around its median projections is based on root mean squared errors of private and government forecasts made over the previous 20 years: the width comes from measured past error, not preference. GAO's Cost Estimating and Assessment Guide likewise makes quantifying risk and uncertainty a formal step whose purpose is a stated level of confidence for the point estimate (paraphrased).

Applied to data licenses, the core metric is coverage: the share of signed prices that fell inside the range stated at a fixed earlier step. If a company calls its ranges 80 percent ranges, coverage should sit near 80 percent across many deals. A one-sided binomial test at the 5 percent level gives a floor: with 10 signed deals, ranges are too narrow if 5 or fewer prices landed inside; with 20 deals, 12 or fewer; with 40 deals, 27 or fewer. Coverage far above the stated level is the opposite problem, ranges too wide to guide anyone.

SourceX publishes no figures of this kind; its investor page reserves company and commercial metrics for investors, who ask for them through a materials request form. Five metrics to request from any company, each with its denominator:

How investors can test whether a company's price ranges are honest
MetricDefinitionDenominatorWhat a weak answer looks like
Range coverageSigned prices inside the range stated at the indicative stepEvery deal signed in the period, including those far from the estimateCoverage reported only for deals that closed near the midpoint
Realization ratioSigned price divided by the indicative midpointThe same signed deals, with median and spreadAn average with no spread and no count
Width by stepTop-to-bottom ratio at each negotiation stepDeals that reached each stepNo narrowing between first estimate and signature
Evidence mixShare of range inputs that are executed rather than askedEvery evidence item in the fileAsking prices or category ranges counted as comparables
AbandonmentDeals that ended without a signature, by stepEvery deal that received an indicative estimateLost deals dropped from the denominator

What could make range-based pricing wrong

Ranges carry their own failure modes, and an investor should test for each before trusting a company that reports them.

  • Ranges as an excuse: a range so wide that every outcome fits is a refusal to estimate. The coverage test catches ranges that are too narrow; the width-by-step metric catches ranges that never narrow.
  • Correlated assumptions: multiplying factors assumes they move independently, and the worksheet should flag factors that share a cause.
  • Thin public comparables: few data licenses disclose terms, and those that do rarely include the rights, record counts or preparation needed to normalize them, so most ranges rest on private evidence an investor has to ask to see.
  • Disagreeing methods: OpenDataVal found no valuation algorithm best across all tasks, so a range built on one method can look tighter than the evidence allows.
  • Stale evidence: an executed price from a year ago reflects that year's buyers and their training needs.
  • Anchoring: once a midpoint is spoken, both sides drift toward it. A range does not remove the anchor; it only puts it in writing where it can be challenged.

Price questions for any data licensing company, SourceX included

These questions separate a company that measures its price uncertainty from one that narrates it. SourceX's own published terms call an early estimate indicative and fix each deal's exact dataset and terms in a signed agreement before anything is delivered; how its ranges have performed is a question for its investor materials, not its website.

  • Which inputs to your ranges are executed transactions, and which are asking prices or category ranges?
  • For each comparable, which adjustments did you apply, and what low and high factor did each carry?
  • What share of signed prices fell inside the range stated at the first estimate, across how many deals?
  • Where did signed prices land relative to the midpoint, and does that pattern differ by data type or buyer purpose?
  • Which evidence narrowed your ranges fastest: use statements, rights reviews, samples or preparation terms?
  • How many deals ended without a signature after an estimate, and at which step?

Methods note: the worksheet arithmetic and what was not measured

The compounding table, the worksheet, the negotiation ladder and the calibration floors were computed on 2026-10-28 by a short Python script (xs0218_price_range.py, kept with its output in the writer's working folder) with exact fractions, rounding once, half up, at output, and no randomness. All inputs are invented examples in value units (U); they describe no real deal and no SourceX figure.

Rules as the script applies them. Each open adjustment is a multiplicative factor between 0.8 and 1.2, so n open adjustments span 0.8 to the power n through 1.2 to the power n. A comparable's normalized range multiplies its stated value by the product of its adjustments' low factors and, separately, of their high factors. An item counts toward the negotiating range only if it is an executed transaction and its normalized top-to-bottom ratio is at most 2.0; the range is the intersection of the counted items, and an empty intersection is reported as a conflict. The ladder settles one of four assumptions per step by setting its factor to 1.0. Each calibration floor is the largest count of prices inside their ranges whose cumulative Binomial(n, 0.8) probability is below 0.05, for n of 10, 20 and 40.

Public evidence: the Data Banzhaf and OpenDataVal papers on arXiv, OpenDataVal's README at commit 5c2856b, the Bureau of Labor Statistics page on Producer Price Index concepts, the Federal Reserve's June 2026 Summary of Economic Projections, GAO's Cost Estimating and Assessment Guide (GAO-20-195G) and SourceX's how-it-works, FAQ and investor pages, read on 2026-10-28. The external pages other than the GitHub README were verified through search excerpts on their own domains, and a reviewer should open each before release; the SourceX pages were fetched directly. Not measured: any realized data license price, coverage rate or realization ratio, at SourceX or elsewhere; how correlated the four adjustments are in practice; and how often published asking prices differ from signed prices.

Frequently asked questions

Why can't a data license have one fair price?

Because the same records are worth different amounts under different rights, usable scope, preparation and buyer purpose, and each of those has to be estimated. Even the buyer's measured value of the data varies between training runs, as the Data Banzhaf paper reports. A single number hides those estimates; a range shows them and lets each side argue about the right one.

What is a rights-normalized comparable?

It is an executed data license converted to the same terms as the dataset being priced: the same exclusivity, permitted uses and term, the same usable share and the same preparation duties. Each conversion carries a low and high factor. A comparable that needs many conversions produces a wide range and should be treated as context rather than as a price.

How wide should a data licensing price range be?

As wide as the evidence requires and no wider: it should come from stated adjustment factors and, once a company has history, from how often past ranges contained the signed price. If 80 percent ranges contain far fewer than 80 percent of signed prices, they are too narrow; if they contain nearly all of them, they may be too wide to guide a negotiation.

Do asking prices for datasets tell you what data is worth?

Not on their own. An asking price is an opening position, and official price statistics such as the Producer Price Index are built from transaction prices for exactly that reason. Asking prices can show what sellers hope for, so list them, but do not count them as comparables.

What should investors ask a data company about pricing?

Ask which range inputs are executed transactions, which adjustments were applied to each, what share of signed prices fell inside the first stated range, and how many deals were lost after an estimate. A company that can answer with denominators is measuring its uncertainty rather than describing it.

Sources

  • Wang and Jia's Data Banzhaf paper reports that training noise makes different runs of one data valuation algorithm produce inconsistent value rankings, with average Spearman correlations of about 0.038 for Shapley values and 0.856 for Banzhaf values in its stability test. Source
  • The OpenDataVal paper benchmarks eleven data valuation algorithms and finds that no single algorithm performs uniformly best across all tasks. Source
  • OpenDataVal's repository describes a unified data valuation benchmark across image, text and tabular datasets, accepted at the NeurIPS 2023 Datasets and Benchmarks track. Source
  • The Bureau of Labor Statistics Handbook of Methods says Producer Price Index prices are transaction prices that include discounts, premiums, rebates and allowances rather than list prices. Source
  • The Federal Reserve's June 2026 Summary of Economic Projections bases the confidence interval around its median projections on root mean squared errors of forecasts made over the previous 20 years. Source
  • GAO's Cost Estimating and Assessment Guide makes quantifying risk and uncertainty a step whose purpose is to attach a level of confidence to a point estimate. Source
  • SourceX's how-it-works page says an indicative estimate comes first and pricing stays indicative until a Dataset Deal Confirmation is signed. Source
  • SourceX's FAQ says the exact payout and terms are fixed in a Dataset Deal Confirmation that the supplier can sign, question or decline, and that an estimate is indicative. Source
  • SourceX's investor page reserves company and commercial metrics for investors and offers a materials request form. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies