Published industry wording tests derived data in a data licensing agreement with two questions about each output: can the licensed data be rebuilt from it, and can it stand in for the licensed data? This post sets out that wording and what it says about ownership, what copyright and database law add in the US, the EU and the UK, and what a licence should say about scores, aggregates, models and enriched records. An illustrative clause applied to five outputs shows how to sort your own before you sign. Each statement reports what its source says. None of it is legal advice, and the questions it raises are for your counsel.
The uses a licence can grant are set out in what a data licence covers, and passing data to your customers in the entry on the redistribution licence. This post works through one clause from that list: derived data.
The definition of derived data in a data licence agreement
Derived data is information created wholly or partly from licensed data that cannot be readily reverse-engineered to recreate the licensed data and cannot be used to produce a reasonable facsimile of it. That is the definition in the Best Practice Recommendation for Derived Data, version 2.1 of November 2015, from FISD, the Financial Information Services Division of SIIA.
An output has to pass both tests. The rebuild test asks whether a recipient could work back from the output to licensed values; the substitute test asks whether the output could do the job the licensed data does. Whether a blend, a share or a ranking passes often turns on what is published with it: the inputs beside it, or a group so small that it describes one record.
Three more points from the same document:
- Other names. FISD notes that some providers call derived data New Original Works or Original Created Works. Search the draft for both.
- No provider's policy. The working group records differences of opinion and says it should not be assumed to represent the contractual or policy approach of any member provider.
- Published determinations. FISD recommends that providers publish their judgments on which transformations create derived data, with examples whenever possible. Ask your vendor for its examples.
The same split in open data licence templates
Open licence templates draw a similar line in their own words.
| Template | Output outside the licence's sharing conditions | Output that stays under them |
|---|---|---|
| CDLA-Permissive-2.0, last updated 17 June 2021 | Results: any outcome of computational analysis of the data, models and their insights included, free of any restriction or obligation | Shared data, changed or not, which travels with the agreement's text |
| CDLA-Sharing-1.0 | Results, which may hold no more than a de minimis portion of the data, with no obligation on their use or publication | Enhanced Data, your additions and modifications, published under the same agreement |
| ODbL 1.0 | A Produced Work, such as an image or a text made from the contents, with a notice of its source | A Derivative Database, which when used publicly is offered under the ODbL or a compatible licence |
The de minimis limit is a rebuild test written into a definition. ODbL also says that repeated and systematic extraction of insubstantial parts may amount to a substantial part, which is why repeated releases of small aggregates need a check of their own.
Ownership of derived data in a data licensing agreement
FISD's 2015 recommendation is that whoever creates derived data, vendor or subscriber, should be acknowledged as its owner, though the information provider may still require a licence for the underlying data used to create or distribute it. The same recommendation asks the contract to define intellectual property rights over the use, storage and dissemination of derived data.
Ownership and permission are therefore two clauses: title to an output can sit with you while the right to show it or pass it on still depends on your licence to the underlying data.
FISD adds two points:
- Downstream recipients. The 2015 recommendation says a third party receiving derived data should owe nothing under the original provider's contract.
- Models and generative AI. A buy-side paper from the FISD Alternative Data Council, Understanding Investment Firms' Use of Licensed Data with GenAI, version 1.0 of March 2026, applies the same two tests to model outputs: where an output passes both, it says use and ownership should not be restricted. At the end of a term it suggests certifying deletion of the raw data and removing artifacts that could act as functional substitutes, naming high-fidelity embeddings built for retrieval-augmented generation. The paper presents itself as the buy-side view, so read it as one party's position.
Open templates speak of conditions, not title: CDLA-Permissive-2.0 places no restriction on Results but assigns them to no one.
Intellectual property, ownership and derived data beyond the licence agreement
Beyond the agreement, the law protects data differently by country. The US Copyright Office states that copyright in a compilation of data covers its selection, coordination or arrangement, not the data itself. The EU and the UK add a separate database right for a maker who has made a substantial investment in obtaining, verifying or presenting the contents.
| Where | Source and date | What it says | Ask counsel |
|---|---|---|---|
| United States | Circular 14, US Copyright Office, reviewed July 2020 | Copyright in a compilation of data covers its selection, coordination or arrangement, not the data. A purely mechanical collection, such as a white-pages directory, gets no compilation copyright | Which restrictions on your outputs rest on the contract alone? |
| European Union | Directive 96/9/EC of 11 March 1996 | Article 7: a maker whose investment in obtaining, verifying or presenting the contents is substantial, judged by quality, quantity or both, may prevent extraction or re-utilisation of the whole or a substantial part of the contents, judged the same way. Article 8: the maker of a database made available to the public may not prevent a lawful user from extracting and re-utilising insubstantial parts for any purpose, within the part the user is authorised to use, and the user may not conflict with normal exploitation of the database or unreasonably prejudice the maker. Article 15: a contract term contrary to Article 8 is null and void | Does the clause on derived data reach insubstantial parts? |
| European Union | CV-Online Latvia v Melons, Case C-762/19, Court of Justice, 3 June 2021 | A specialised search engine that copies and indexes the whole or a substantial part of a freely accessible database, and lets its users search it, is extracting and re-utilising it, which the maker may prohibit where this puts at risk recovering its investment through normal operation of the database, for the national court to verify | Could an output put at risk the investment in the database it came from? |
| United Kingdom | Sui generis database rights, Intellectual Property Office guidance on GOV.UK, published 30 January 2020 | Database right protects contents without any test of originality, on substantial investment. For databases made on or after 1 January 2021, only UK citizens, residents and businesses qualify in the UK, and they cannot hold the right in the EEA | Which country's database right, if any, applies? |
The same guidance suggests that UK owners of such databases consider licensing agreements or copyright for protection in the EEA. CC BY 4.0 draws on the same line: a database in which you hold database rights and into which you put all or a substantial portion of a licensed database's contents is Adapted Material, though its individual contents are not.
The substantial part test asks how much of the contents an output takes, by quantity or quality; the 2021 judgment asks whether a use puts the investment at risk, which is close to the substitute test. Where a country's law protects little of a dataset, ask counsel how much of the restriction on your outputs comes from the agreement alone.
What to write for scores, aggregates, models and enriched records
Each kind of output can fail the tests in its own way.
| Output | Where a test can fail | What to settle in writing |
|---|---|---|
| Score | Inputs or evidence shown beside it let a reader solve for a licensed value | Whether the score alone is derived, which inputs and evidence may be shown, and its version |
| Aggregate | A group of one company, or one where every member or none qualifies; two releases one company apart | The smallest published group, a rule for repeated releases, and who may see the figures |
| Model | Weights that store rows, or an index built to return them | Whether training is allowed, which artifacts survive the term, and whether a retrieval index counts as data |
| Enriched record | A copied field is the licensed value, and a set of copies can replace the feed | That appended fields stay licensed data, which computed fields are derived, and what happens to both at the end of the term |
The use cases on account scores in a sales platform, enriching company records in a data product and training models on dated company signals work through each build.
An illustrative example: five outputs under one clause
This example is illustrative: Acme Robotics is an invented company, and the clause was written for this page. It is not a recommended form, its wording takes positions that counsel would weigh, and the sorting below follows from this clause's words alone, not from any other contract.
Assumptions:
- Acme Robotics, a maker of warehouse robots, licenses Intent Scores and Technology Changes for internal use and for display in a reseller portal.
- The clause below is its only wording on derived data.
- Acme's account score is 0.6 times the intent score for one topic plus 0.4 times its own engagement score, both on a scale of 0 to 100.
- For one account, the intent score is 80 and the engagement score is 50.
- In a quarterly benchmark segment of 140 companies, 21 added software in one technology category.
- Acme's classifier returns one probability per account, and its weights hold no licensed rows.
The clause:
- Licensed Data means the data delivered under this Agreement, in any format, and any copy, extract or reformatting of it.
- Derived Data means data the Licensee creates by processing Licensed Data, alone or with other data, that (a) cannot readily be reverse-engineered to recreate Licensed Data or any part of it, and (b) cannot be used as a substitute for Licensed Data or any part of it. Data that reproduces values of Licensed Data, whether renamed, reformatted or combined with other data, is Licensed Data.
- As between the parties, the Licensee owns Derived Data and may keep and use it after this Agreement ends. Disclosure of Derived Data to third parties is governed by clause 9.
Clause 2 follows the FISD definition and adds the words "or any part of it", so that one recoverable value fails, and a sentence that settles copied values. Clause 3 follows FISD's ownership recommendation.
Output 1: the account score alone, in Acme's sales dashboard. 0.6 × 80 + 0.4 × 50 = 48 + 20 = 68. One number with two unknowns has many solutions: an intent score of 60 with an engagement score of 80 also gives 36 + 32 = 68. Nothing licensed can be recovered, and one blended number cannot do the work of weekly scores by topic. Derived Data.
Output 2: the same score in the reseller portal, beside the engagement score and the weights. Take 0.4 × 50 = 20 from 68 to get 48, divide by 0.6, and the licensed 80 reappears. Test (a) fails: not Derived Data, so clause 3 does not apply and the agreement's other terms govern its display.
Output 3: the quarterly benchmark. 21 ÷ 140 = 0.15, so 15% of the segment added software in the category. No company's record can be read from the share, and a share cannot replace a dated feed per company. Derived Data. Two variants fail test (a): a segment of one company, which reads 100% or 0%, and a later re-cut of 22 out of 141, which shows that the one company added made the change.
Output 4: enriched records in Acme's CRM. Each account gets the names and observation times of its latest technology adoptions. Every appended value is licensed, failing test (a), and together the copies can replace the feed for those accounts, failing test (b). Licensed Data, whatever the fields are called.
Output 5: the classifier. On the stated assumption, no licensed value can be recovered from its weights, and a probability per account cannot do the work of a dated feed. Derived Data. A retrieval index over the same rows needs sorting of its own: the 2026 FISD paper names high-fidelity embeddings built for retrieval as artifacts that could serve as functional substitutes.
| Output | (a) Rebuild | (b) Substitute | Under the clause |
|---|---|---|---|
| 1. Score shown alone | Passes: 68 has many solutions | Passes | Derived Data |
| 2. Score with its inputs and weights | Fails: 80 is recovered | Not reached | Not Derived Data |
| 3. Benchmark, 21 of 140 | Passes while no cell states one company | Passes | Derived Data |
| 4. Enriched CRM records | Fails: values copied | Fails | Licensed Data |
| 5. Classifier | Passes on the stated assumption | Passes | Derived Data |
Run the rebuild test on sample data
Before you sign, run the rebuild test on any aggregate you plan to publish, on sample data for the companies you track. The query joins Technology Changes to a table of your own that places each company in a segment, and lists the cells that state company values: groups below your minimum size, and groups where no member or every member made the change. The category, the period and the minimum size are yours to set.
with cells as (
select s.segment,
count(distinct s.company_id) as companies,
count(distinct e.company_id) as adopters
from my_segments s
left join company_tech_events e
on e.company_id = s.company_id
and e.category = 'technology'
and e.change = 'added'
and e.technology_category = :category
and e.observed_at >= :period_start
and e.observed_at < :period_end
group by s.segment
)
select segment, companies, adopters
from cells
where companies < :min_companies
or adopters = 0
or adopters = companies
order by companies;An empty result means no cell states a company's record. Compare each release's segment sizes with the next, so that no segment changes by a single company.
Where Fokals data fits a derived data clause
We license Fokals data by written agreement, for internal use, embedding in a product or redistribution. Each score in Intent Scores comes with the dated evidence behind it, so a product that shows the evidence shows dated licensed records, and the agreement can name the right that covers that display. Every observation is dated, daily and weekly datasets are written once and never revised, and labels and scores carry a named version on every row. A bulk export delivered as a package of files carries a manifest naming its period, versions and licence, so a derived output can record which delivery and version produced it. Sample data for the companies you track is sent on request with the data dictionary and the methodology.
Frequently asked questions
What is derived data in a data licence?
Derived data is information you create from licensed data that cannot readily be worked back to the licensed values and cannot stand in for them. FISD's Best Practice Recommendation for Derived Data, version 2.1 of November 2015, sets out this two-part definition. The definition in your own agreement is the one that governs, and some providers use other names for the same idea, such as New Original Works.
Who owns derived data under a data licensing agreement?
The agreement decides. FISD's 2015 recommendation is that the vendor or subscriber who creates derived data should be acknowledged as its owner, while the provider may still require a licence to the underlying data used to create or distribute it. Read the ownership clause together with the clauses on use, disclosure and the end of the term, because owning an output is a separate matter from the right to publish it.
Is an aggregate of licensed data always derived data?
Not under the FISD tests. An aggregate passes the rebuild test only when no published figure describes an individual record. A group of one company, a group where every member or none qualifies, or two releases that differ by a single company all let a reader recover company values. Set a smallest group size and a rule for repeated releases, and test both on sample data before you publish.
Is a model trained on licensed data derived data?
Under the FISD tests it depends on what the model can return. The FISD Alternative Data Council, in a buy-side paper of March 2026, applies the same two tests to model outputs and names high-fidelity embeddings built for retrieval as artifacts that could serve as functional substitutes. Settle in the agreement whether training is allowed and which model artifacts may be kept after the term ends.
Does EU database law limit what a data licence can restrict?
Under Article 8 of Directive 96/9/EC, the maker of a database made available to the public may not prevent a lawful user from extracting and re-utilising insubstantial parts of its contents for any purpose, within the part the user is authorised to use and without conflicting with normal exploitation of the database, and Article 15 makes a contract term contrary to Article 8 null and void. Whether the vendor's database falls under the directive, and what counts as insubstantial for your use, are questions for your counsel.
Are facts in a dataset protected by copyright in the US?
The US Copyright Office's Circular 14 states that copyright in a compilation of data extends to its selection, coordination or arrangement and not to the data itself, and that a purely mechanical collection, such as a white-pages telephone directory, gets no compilation copyright. Ask counsel which restrictions in your agreement rest on copyright and which rest on the contract alone.
The queries and code on this page are examples to adapt. Test them in your own environment before you rely on them.
What this page says about the products it names was checked against their public documentation on 7 October 2026. Product and company names are trademarks of their owners. Fokals is not affiliated with them or endorsed by them.