To run a trial of an alternative dataset well, treat it as two things at once: a short licence and an experiment whose pass marks are written before the data arrives. This guide takes a fund's data or research lead through what to fix before delivery, what the trial agreement should cover, and a test plan for the trial period, reading the industry's published template term by term. It is a working guide, and no substitute for your counsel's reading of the contract.
How to run a trial of an alternative dataset, and what the agreement should cover
Run a trial of an alternative dataset in three stages. Before delivery, write down the decision, the companies you track and the pass marks, and clear compliance. Sign a trial agreement that covers the evaluation period, permitted use, confidentiality, derived results and deletion. Then run the tests in a fixed order and record each result against its pass mark.
A published standard exists for the agreement. The Alternative Data Council, a set of working groups inside FISD, a division of SIIA, publishes a Trial Data License and Mutual Non-Disclosure Agreement as a template. The council states that adoption is entirely voluntary. Among its publications we found no test plan for the trial itself, so the plan below is our own.
Before the data arrives: five things to fix in writing
A result read without a pass mark can be argued either way, so fix five things in writing before delivery.
- The decision and the pass marks. Name the decision the dataset would feed and how often it is made, and write the figure that counts as a pass for each test below. A pass mark chosen after the results only describes them.
- The list of companies. The companies you track, as security identifiers or website domains, frozen on the day the plan is dated. Coverage is measured on this list, never on the vendor's total.
- The compliance review. Send the questionnaire ahead of the data. FISD's trial due diligence questionnaire, which the council's page calls a shorter checklist for trial data, asks how the data is collected, whether it contains or may be derived from material non-public information, and whether personal information is collected. It also requests a data dictionary or schema, a sample of fewer than 100 rows from more than three months ago, and the relevant terms of any contracts under which the vendor buys data.
- The vendor's own description. FISD's Data Vendor Tear Sheet records answers the trial then verifies on rows: the start date of the history and any breaks in it, the update frequency and the delay before delivery, the security identifiers mapped, whether the history is point-in-time, and which timestamps the data carries.
- Where the data will sit, and the exit. One schema or bucket, open to the named evaluators only, so that deletion is one step and can be shown. Name who signs off the result and what follows each outcome: a licence, an extension agreed in writing, or deletion.
How the review is applied has drawn examiners' attention too. In a risk alert of 26 April 2022, the examinations staff of the US Securities and Exchange Commission (SEC) reported advisers that did not apply their due diligence process to all sources of alternative data, and advisers unable to demonstrate, for instance with documents, that their procedures were applied consistently. The alert is a staff statement and does not mention trials. A dataset on trial is a source like any other once it sits in your systems, so the review and its record belong before delivery. The guide to alternative data due diligence questionnaires covers the full version.
What a trial agreement with an alternative data vendor covers: the licensing terms of a pilot evaluation
A trial agreement is a short data licence joined to a confidentiality agreement. The industry's published template sets the evaluation period, limits use of the dataset to evaluation, names who may handle it, says who owns the results you derive and what is returned or destroyed afterwards, and carries the vendor's representations about how the data was gathered.
A paper called a trial, a pilot or an evaluation is read for the same points. FISD's template is the reference text used here. Its introduction says that negotiation between counsel often delayed the start of an evaluation by a few weeks to months, and gives as a goal common terms often accepted by both vendors and buyers. The table reads it term by term.
| Term | What the FISD template provides | What you still settle |
|---|---|---|
| Evaluation period | Set in the order form. It starts on receipt of the complete data, as confirmed by the customer, runs for a number of days the parties fill in, and can be extended by prior written agreement. | The days, from the hours in your test plan |
| Permitted use | A worldwide, non-exclusive licence for the customer and its affiliates to access, use, display, modify, aggregate, store and analyse the data, solely for evaluation. | Whether a live decision may rest on trial data |
| Who may handle it | Representatives who need it, and nominated consultants: service providers that support the evaluation, for instance by transmitting, converting or storing the data. The customer answers for their acts. | The providers, by name |
| Confidentiality | Mutual, and it survives termination. The data is confidential information, used only for the evaluation. | Anyone who must be told |
| Derived results | Tools, insights and analysis the customer generates from the data are owned by the customer without restriction, provided they could not serve as a substitute for the vendor's product. | How a fitted model is treated |
| Return or destruction | On written request of the disclosing party, with exceptions that include backups where permanent deletion is infeasible, one copy held for regulatory or retention purposes, and derived results. | Whether deletion follows the last day unasked, and how it is evidenced |
| Vendor representations | Lawful collection, including under privacy law. Commercially reasonable steps so that the data holds no material non-public information and no personal information. At least 30 days' notice of a planned change to sources or methodology. | Whether the same statements carry into the production licence |
| Warranty | Beyond those representations the data is provided as is, with accuracy and fitness for a purpose disclaimed. | No blank to fill: this is why the trial tests the data |
| Order form | The description of the data, its schema, the history, the file format and naming, the delivery method, frequency and timing, and any fees. | Every line, before signature |
Evaluation use and trading. The licence is for evaluation only, and a separate clause states that nothing in the agreement is intended to limit the customer or its affiliates from trading any financial product. How the two read together when a position is informed by trial data is for your counsel to settle in the order form before the data arrives.
Deletion runs on request. Return or destruction follows a written request from the disclosing party, not the calendar. If your policy is to hold no data without a licence, schedule the deletion yourself and say what deleted means. NIST's Guidelines for Media Sanitization define sanitisation by its result: reading the target data off the media is no longer feasible at a given level of effort.
What you learned stays yours. Derived results belong to the customer, up to the point of substitution. Where the vendor's data can be discerned in a result or separated from it, the confidentiality terms go on governing that result. On a plain reading, a score or a chart is derived and a table that restates the vendor's rows is the data.
Representations have drawn enforcement. On 14 September 2021 the SEC announced that an alternative data provider for the mobile app industry and its co-founder had agreed, without admitting or denying the findings, to settle securities fraud charges over misrepresentations about how its alternative data was derived, and described the matter as its first enforcement action charging an alternative data provider with securities fraud. The template writes a vendor's account of its collection into the contract from the first day.
The template also provides that click-through or other online terms presented with the data have no force, so the signed text governs. For the production licence that follows, what a data licence covers sets internal use, embedding and redistribution side by side.
A test plan for the trial period
Run the tests in this order, because each depends on the one before. Write the pass mark for each before delivery.
| Test | What you compute | What the result tells you |
|---|---|---|
| 1. Delivery | Files, schema, date range, duplicates and missing values, against the order form | Whether the product is the one described |
| 2. Mapping | The share of your list matched on a security identifier or a domain, and among matches the share with each field you need | Coverage of the companies you track |
| 3. Timing | For each record, the gap between the date of the event and the time it was observed | Which date your research may use |
| 4. Stability | One closed period pulled twice, as far apart as the trial allows, compared row by row | Whether history is rewritten |
| 5. Depth of history | Closed decision periods of your own strategy in the delivered history, against the minimum your research process sets | Whether your own test can run at your standard |
| 6. Your own test | The relation between the data and your outcome, on a design fixed before delivery | Whether the data adds to what you hold |
| 7. Operations | Delivery times against the stated frequency and delay, and the time taken to answer a question | What production would be like |
Timing: the date you could have known
FISD's learning point on point-in-time data gives the reason for the third test: a strategy validated by back-test needs a history that shows, at each point in time, what was known and when it was known. Without it the result can carry look-ahead bias and survivorship bias. The tear sheet asks whether a point-in-time history is true, meaning collected on the event date, or backfilled, and which timestamps are added. The trial checks those answers: subtract the event date from the observation time on a sample of records, read the whole distribution, and key your research on the observation time.
How much history is enough for your strategy
No single figure answers this, and the template's order form leaves the history as a line for the parties to fill in. The answer is a count you make yourself. Count the closed, non-overlapping decision periods of your strategy in the delivered history that pass the timing and stability tests, leaving out any span marked as backfilled. Compare the count with the minimum sample your research process requires before it accepts a result. If the count falls short, the dataset can still pass the other tests, and your own test waits until the record meets your standard.
Your own test, and the hours it takes
Whether a dataset adds information to your process is a question only your own outcomes answer, on a design fixed before delivery: the measure, the comparison and the periods held back. The Guide to Alternative Data that the council publishes says that evaluating a new data source is complicated and that integrating one can take time, so count the hours before you agree the number of days. A result on past data is a finding about that sample, recorded with its design.
A worked example: a 30-day trial with Acme Robotics on the list
This example is illustrative: the research team, the dataset, every figure and Acme Robotics are invented.
Assumptions:
- The team tracks 400 listed companies, each held as an ISIN. Acme Robotics is one of them.
- The dataset is daily hiring data: a row per company and day, and each posting with the date the employer gives and the date it was first observed.
- The order form sets an evaluation period of 30 days from confirmed receipt.
- Pass marks, dated before delivery: at least 85% of the list matched; at least 70% of the list with one or more postings; at least 90% of postings first observed within 7 days of the employer's date; no row changed between two pulls of one closed week.
The working:
- Mapping. 352 of the 400 ISINs match: 352 / 400 = 88%. Pass.
- Field coverage. 290 of the 352 matched companies have at least one posting: 290 / 400 = 72.5% of the list. Pass. The other 110 companies, 48 unmatched and 62 matched with no posting, are listed by country and size.
- Timing, shown on one company. Acme Robotics has 40 postings. The gap from the employer's date to first observation is 0 to 1 day on 22 of them, 2 to 3 days on 12, 4 to 7 days on 4 and longer on 2. Within 7 days: 22 + 12 + 4 = 38, and 38 / 40 = 95%, above the mark. Yet 12 + 4 + 2 = 18 postings, or 45%, were observed two or more days after the employer's date. A back-test keyed on that date would treat them as known before the dataset held them, so the team keys on first observation.
- Stability. One closed week is pulled on day 3 and again on day 24. The 290 companies give 290 times 7 = 2,030 daily rows in each pull, and no row differs. Pass.
The same timing count then runs over every posting. If it also clears 90%, all four pass marks are met, and the note records each figure beside its mark. The depth-of-history count and the team's own test are reported separately. On day 30 the team signs a licence or drops the trial schema and logs the deletion.
Running the plan on Fokals data
Fokals data is licensed by written agreement, for internal use, embedding in a product or redistribution. Sample data for the companies you name is sent on request with the data dictionary and the methodology, and the tests above run on it as follows.
| Test | What the data carries |
|---|---|
| Compliance review | Collected from first-party company sources and public records, and processed in-house. Company-level data throughout. The sourcing statement is public. |
| Mapping | One stable company ID. A listed company carries its ticker, MIC, ISIN, LEI and share-class FIGI, and its SEC CIK where it has one. |
| Timing | Every record carries the time it was observed. In Job Postings the posting date a company's board gives and the first-seen date are separate fields. |
| Stability | Daily and weekly datasets are written once, after the period closes, and are never revised. A first observation sets a baseline and is never counted as a change. |
The worked example maps onto the Hiring datasets, Job Postings and Hiring Activity. The stability test is the one set out in why we write each table once.
Frequently asked questions
What should a trial agreement for an alternative dataset cover?
A trial agreement should cover eight things: the evaluation period and when it starts, a licence limited to evaluation, who may handle the data, mutual confidentiality, ownership of derived results, return or destruction afterwards, the vendor's representations about how the data was gathered, and an order form describing the data, its history, its format and its delivery. FISD's Alternative Data Council publishes a template with all eight, which the parties adapt.
How long should a trial of an alternative dataset last?
Set the length from the work, since the published FISD template leaves the number of days blank. Add up the hours of your test plan, and allow for two pulls of the same closed period far enough apart for a revision to show. The template starts the period on your receipt of the complete data, as you confirm it, and lets the parties extend it by prior written agreement.
How much history do I need to evaluate an alternative dataset?
There is no general figure. Count the closed, non-overlapping decision periods of your own strategy in the delivered history, after removing any span that is backfilled or fails your timing test, and compare the count with the minimum sample your research process requires. The answer differs between a strategy that decides daily and one that decides quarterly, which is why the test is yours to run.
Do I have to delete trial data when the evaluation ends?
That depends on the agreement you signed. Under the FISD template, return or destruction follows a written request from the disclosing party. Its exceptions cover backups where permanent deletion is infeasible, one copy held for regulatory or retention purposes, derived results and insubstantial amounts. If your own policy is stricter, schedule the deletion yourself and keep a record of it.
Can I keep my analysis after a data trial ends?
Under the FISD template, yes: tools, insights and analysis generated from the data are owned by the customer without restriction, provided they could not be used as a substitute for the vendor's product. Where the vendor's data can be discerned in a result or separated from it, the confidentiality terms still apply to that result. Other agreements differ, so read the clause on derived data before you start.
Is there a standard trial agreement for alternative data?
Yes. FISD's Alternative Data Council publishes a Trial Data License and Mutual Non-Disclosure Agreement template, developed with buy-side, vendor and independent participants, and a shorter due diligence questionnaire for trial data. Adoption is voluntary, and the council notes that its documents may not suit every situation. Among its publications we found no standard test plan for the trial itself.
Should compliance review a dataset before a trial starts?
Yes, before the data arrives. A dataset on trial sits in your systems like any other source. FISD publishes a shorter questionnaire for trial data, which asks how the data is collected, whether it contains material non-public information and whether personal information is collected. In a 2022 risk alert, a staff statement that does not mention trials, SEC examinations staff reported advisers that did not apply their due diligence process to all sources of alternative data.
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.