Skip to content

Deal economics

How to present uncertainty in a data valuation: confidence ranges and a memo template

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Present a data valuation as a conditional range, not one number: a low, base and high case, each defined by named assumptions about buyer utility, completeness, rights, costs, timing and alternatives, with the evidence for each, confidence words you have defined, and the single fact that would move the range most. Say plainly that the range is an internal estimate, not a quote or promised payment.

Value drivers

Dataset valuation confidence

This page covers how to present uncertainty in a data valuation: confidence ranges and a memo template.Illustrative

Key takeaways

  • A dataset valuation should be a range tied to named conditions, not a single figure.
  • Build low, base and high cases as coherent stories across six drivers: buyer utility, completeness, rights, costs, timing and alternatives.
  • Define confidence words once in the memo and use only those; borrow the structure from calibrated scales such as the IPCC's.
  • The rights position usually moves the result most; when rights are unknown, the honest memo shows no range yet.
  • Every memo should say it is an internal estimate, not an offer, quote or promise of payment.

How should a company present uncertainty in a data valuation?

Present it as a range tied to named conditions, not as one figure. State a low, base and high case, the assumption that defines each, the evidence behind each assumption, how confident you are in words you have defined, and which single fact would move the range most. Then say what the range is not: a quote, an offer or a promise of payment.

That structure lets an owner, a CFO and a board discuss a possible data license honestly. It keeps the conversation on the conditions that drive value, such as rights, buyer fit and preparation cost, instead of on a number someone heard secondhand and anchored to.

The audience for the memo is internal: the owner, finance, counsel and anyone on the board who will decide whether the company explores a license. It is written before any buyer conversation and updated as evidence arrives. It is not a document for a buyer, and it does not replace a negotiated price.

This page gives a range-based valuation memo template you can fill in as is, the drivers that set dataset valuation confidence ranges, wording for confidence, an illustrative worked example and the cases where a range is the wrong tool.

Why does a single number mislead in a dataset valuation?

A single number hides the assumptions that produced it, and in a dataset valuation those assumptions are large. Whether the records can be licensed at all, whether a buyer needs them, how much preparation they need and whether the license is exclusive can each change the outcome by more than any refinement of the arithmetic.

Financial reporting already has a model for this. IFRS 13 separates fair value measurements by the observability of their inputs, and Level 3 measurements rely on unobservable inputs. For recurring Level 3 measurements, the standard asks for a narrative description of how sensitive the value is to changes in unobservable inputs that could produce a significantly higher or lower result. A data valuation for a first license is almost entirely unobservable inputs, so it deserves at least that much disclosure.

Tax rules also accept ranges. The U.S. transfer pricing regulation describes an arm's length range built from comparable transactions and, where comparables cannot be fully adjusted, an interquartile range from the 25th to the 75th percentile of results. You will rarely have comparables for a dataset, but the principle carries over: a defensible range is built from stated evidence and stated method, and its width reflects what you do not know.

What should a range-based valuation memo include?

The memo should let a reader who was not in the room understand the range, the reasons for it and what would change it, on one or two pages. Use these ten blocks in this order, and date and version every copy.

  • Fill in blocks 2 to 5 before anyone writes a number. If the evidence inventory is thin, the memo should say so in block 1 and stop at block 9.
  • Keep the assumptions register as a separate tab in the spreadsheet that holds the arithmetic, so every scenario reads its settings from one place.
  • Write block 7 last. The confidence statement should summarize blocks 4 to 6, not anticipate them.
  • Give the memo a single owner. Ranges that several people edit without versions drift toward whoever spoke last.
What should a range-based valuation memo include?
BlockWhat to writeEvidence to attachCommon mistake
1. Purpose and statusWhy the memo exists and the sentence: this is an internal estimate, not an offer, quote or forecast of paymentRequest or board minute that asked for itLetting the memo circulate without the status line
2. ScopeSystems, record types, date range and volume in scope; what is excludedData profile one-pager or export countsValuing everything the company holds instead of what can be offered
3. Rights positionWho owns the records, contractual limits, privacy constraints, open questionsContract review notes and counsel's statusAssuming de-identification creates a right to license
4. Evidence inventoryEach fact used, its source and its dateExports, counts, quotes, buyer requests if anyMixing verified counts with guesses in one column
5. Assumptions registerEvery unobservable input, its low, base and high setting and whyReasoning for each settingHiding the assumption that matters most in a footnote
6. ScenariosLow, base and high cases, each defined by a coherent set of assumptionsAssumptions register rows used by each caseBuilding cases by adding and subtracting a fixed share
7. Range and confidenceThe range and a confidence statement in defined wordsDefinitions table for confidence termsUsing likely or conservative without defining them
8. SensitivityWhich two or three inputs move the result most and by how muchData table or tornado chart from the modelTesting inputs that do not matter and skipping ones that do
9. What would change the rangeFacts that would narrow, widen or move it, and how to get themOwner and date for each open itemPresenting the range as final
10. Sign-offPreparer, reviewer, date, version and next review dateReviewer commentsNo version history, so stale memos keep circulating

How do you build the low, base and high cases?

Build each case as a coherent story, not as a percentage above and below a guess. The low case is the outcome if the weaker plausible answer holds for each driver that is genuinely uncertain; the high case is the outcome if the stronger plausible answer holds; the base case is your best judgment. Every case should be something you could defend in a board meeting if it turned out to be the true one.

Six drivers do most of the work in a dataset valuation. Write the evidence you have for each, then set the low, base and high settings from that evidence.

  • Keep each case internally consistent: a high buyer-utility case usually also assumes higher preparation effort, because the valuable records are the detailed ones.
  • Never let the high case assume a resolved rights question that is still open; put open rights questions into what would change the range instead.
  • Express preparation cost and timing in the same units as the value, so net figures can be compared across cases.
How do you build the low, base and high cases?
DriverQuestion it answersLow case ifHigh case if
Buyer utilityWould a model trained or evaluated on these records do something useful better?The task is common and public data covers itRecords capture rare, verified outcomes buyers cannot get elsewhere
CompletenessHow much of the history is linked, labeled and usable?Key outcomes or links are missing for many yearsRecords link request, action and outcome across most of the history
RightsCan the company license these records for the intended use?Customer contracts or privacy terms restrict reuseOwnership is clear and contracts permit the license after review
CostsWhat will export, de-identification, review and delivery take?Data is spread across systems with free text full of personal detailsOne system exports cleanly with structured fields
TimingWhen would any money arrive and on what conditions?Payment depends on later acceptance or milestonesTerms are simple and payment follows delivery
AlternativesWhat else could a buyer use instead?Similar data is available from many sources or synthetic data sufficesFew organizations hold comparable records

How do you word confidence without overstating it?

Define your confidence words once, in the memo, and use only those. The IPCC guidance for its assessment reports is a well-tested model: it pairs a calibrated likelihood scale, in which likely means a 66 to 100 percent probability and very likely means 90 to 100 percent, with a separate qualitative confidence level that reflects the amount of evidence and agreement behind a finding.

Metrology uses the same idea in numbers. The international guide to uncertainty in measurement expresses a result with a coverage interval and an expanded uncertainty, a standard uncertainty multiplied by a stated coverage factor, and asks that the factor be stated when it could be ambiguous. The lesson for a valuation memo is to state what the range is meant to cover, not just its endpoints.

A company memo does not need probabilities, but it does need consistent words. The table below is a starting vocabulary to adapt; the definitions are editorial suggestions, not a standard.

How do you word confidence without overstating it?
Term in the memoUse it whenEvidence typically behind itAvoid
SupportedThe input comes from a count, export or signed documentVerified record counts, contract textUsing it for a reasoned estimate
EstimatedThe input is reasoned from partial evidenceSample review, staff time logs, vendor quotesPresenting it as a fact
AssumedNo evidence yet; the setting is a judgment to testStated reasoning onlyLeaving it out of the sensitivity test
Within the range is expectedMost supported and estimated inputs agree on the base caseSeveral drivers with evidenceImplying a probability you did not compute
Range is indicative onlyKey drivers are still assumedFew verified inputsCirculating it outside the working group

Illustrative worked example: a field service history

This example is illustrative. The company, figures and settings are invented to show the method, and each case is expressed as a percentage of the base case, not as money. They are not market data and say nothing about what any dataset would earn.

A home services company with roughly a decade of jobs in one field service system wants to know whether a license is worth exploring. Its memo scopes five years of linked jobs, estimates and invoices, excludes call recordings, and records that counsel has not yet reviewed two older customer contract templates.

The low case, at an illustrative 40 percent of the base case, assumes the older contracts restrict reuse, so only three years are in scope, and that technician notes need heavy redaction. The base case assumes five years are in scope with moderate redaction. The high case, at an illustrative 150 percent of the base case, assumes the same five years plus strong buyer interest in the callback and warranty outcomes, which are rare in public data. Preparation effort rises from low to high because the high case includes more free text.

The memo's confidence line reads: rights are assumed pending counsel's review, counts are supported, and preparation effort is estimated from a two-week sample. Its sensitivity block shows the rights question moves the result more than every other driver combined. The next step it records is counsel's review of the two contract templates, with an owner and a date, not a negotiation.

The exception: had the review found that customers own the job records, the memo would not show a lower range. It would show no range at all, because there would be nothing the company could license, and it would say so.

Which assumption moves a dataset valuation range most?

The rights position usually has the largest effect, because it decides whether records are in scope at all. After rights, buyer utility and preparation cost tend to dominate, and their interaction matters: the records buyers value most are often the most expensive to prepare.

Find out rather than guess. Build the memo's arithmetic in a spreadsheet with one cell per driver, then vary one or two drivers at a time. Microsoft's guidance describes a one-variable or two-variable data table that recalculates a formula across a grid of input values; for more than two variables, Excel's scenarios tool switches between saved sets of values.

Report sensitivity as swing, not as a list. For each driver, show the result with that driver at its low setting and at its high setting while every other driver stays at base. Sort drivers by the size of the swing. The top two or three usually explain most of the range, and they are where the next piece of evidence should come from.

Watch for drivers that move together. If the high buyer-utility setting only holds when the detailed free-text notes are included, and those notes are what drive redaction cost, testing the two drivers separately overstates the high case. Test them as a pair in a two-variable table.

Record-level methods can sharpen the buyer-utility input when a sample is available. Ghorbani and Zou's data Shapley measures each training record's contribution to a model's performance; even a rough version on a sample can show whether value is concentrated in a few record types, which narrows the range more than any amount of debate.

When is a range the wrong thing to circulate?

Stop and do not circulate a range when the rights position is unknown, when the scope includes regulated or sensitive records nobody has reviewed, or when the evidence column is mostly assumed. In those cases the honest memo says what must be learned first.

Narrow the scope when one system or record type carries most of the value and the rest adds cost and risk. A narrower memo with supported inputs is more useful than a broad one built on assumptions.

Request specialist review when the memo will be used for a decision with consequences: counsel for rights and privacy, a finance reviewer for tax and accounting treatment of any proceeds, and a technical reviewer for buyer-utility claims. This page is general information; it is not valuation, legal, tax or accounting advice.

Bringing the memo to a first licensing conversation

A range-based memo is the right preparation for a first conversation about licensing, because it shows what the company knows, what it does not and what would change the answer. It also protects the company from anchoring on a figure before rights and scope are settled.

With SourceX, the company that owns the records approves the scope and terms and signs before a single record is delivered, so the memo's scope and rights blocks line up with approvals the company itself gives. The rest of the path, from qualification and rights review to contracting, preparation, delivery and payment, is run by SourceX.

A company does not need a finished memo to start: SourceX's first step is a short assessment that requires no data uploads. Keep the range itself for internal use.

Frequently asked questions

What are Level 3 valuations?

Under IFRS 13, Level 3 fair value measurements use unobservable inputs, meaning assumptions the entity develops because market data is not available. The standard asks for added disclosure of how sensitive Level 3 values are to those inputs. Most first-time dataset valuations rely almost entirely on judgment-based inputs, which is why a range and a sensitivity block fit them.

How do you calculate the value of data for a license?

Start from what a buyer could do with the records, then adjust for completeness, rights, preparation cost, timing and alternatives. Each of those is uncertain, so the output should be a low, base and high case tied to named assumptions rather than a single number. Formula-only approaches that multiply record counts by a rate skip the drivers that matter most.

Should I share the valuation memo with a buyer?

Usually not. The memo is an internal decision document that records assumptions and open questions. Share the scope and a data profile instead, and keep the range for the people deciding whether to proceed.

How wide should the range be?

As wide as the evidence requires and no wider. If the range is so wide it supports no decision, the memo has done its job by showing which fact to establish next, usually the rights position or the buyer-utility question.

How often should the memo be updated?

Update it whenever a driver's evidence changes, such as counsel finishing the rights review, a sample being profiled or terms being proposed. Date and version each update so older copies are not mistaken for the current view.

Sources

  • IFRS 13 requires, for recurring Level 3 fair value measurements, a narrative description of sensitivity to unobservable inputs that could produce a significantly higher or lower value. Source
  • The U.S. transfer pricing regulation describes an arm's length range and, where comparables cannot be fully adjusted, an interquartile range from the 25th to the 75th percentile of results. Source
  • The IPCC AR5 uncertainty guidance note defines calibrated likelihood terms, including likely as 66 to 100 percent and very likely as 90 to 100 percent, alongside confidence levels. Source
  • JCGM guidance on measurement uncertainty defines expanded uncertainty as standard uncertainty multiplied by a coverage factor and asks that the factor be stated when ambiguity is possible. Source
  • Microsoft documents one-variable and two-variable data tables in Excel for recalculating a formula across a grid of input values. Source
  • Ghorbani and Zou proposed data Shapley to quantify the value of each training datum to predictor performance. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies