Alternative data can turn activity outside conventional financial feeds into research features, provided timing, vintages, coverage and rights survive scrutiny.
Alternative data
Alternative data is information used in investment analysis from sources outside conventional financial statements, company disclosures and market price feeds. Shop-visit estimates, aggregated payment transactions, satellite images and web activity can reveal an aspect of economic activity before it appears in a company's reported results. For a quant, the opportunity is to construct a useful feature; the challenge is establishing what it measures, when it was available and whether the apparent relationship survives a credible test.
“Alternative” describes the source and use, rather than a quality grade. Publicly accessible information can qualify; an expensive proprietary feed can still be noisy or redundant. A vendor's sentiment score is already an analytical transformation of underlying text, so its measurement method matters alongside the source.
A visit is not a sale
Suppose a fictional provider reports 100 visits to a shop in an earlier period and 120 in a current, equal-length period. With the same observation rules and comparable coverage, the recorded count changes by
100120−100=0.20=20%.
That is a change in the provider's count. It is not a measurement of revenue growth. Some visitors buy nothing, basket sizes change, purchases can move online, and the observed devices may cover a changing fraction of customers. A location feed also needs rules for distinguishing a customer from an employee, a passer-by or a repeated device observation.
Hypothetical recorded shop visits rise from 100 to 120, or 20%. The activity period precedes vendor release and receipt plus validation; only then is this version usable. Visits do not measure revenue or establish alpha.
Original hypothetical counts, with equal-length periods and fixed observation rules. The sequence shows information eligibility, without assuming a particular release delay or trading fill.
Even a good forecast of sales would not establish alpha. The information might already be reflected in prices, and a strategy must still justify its forecast horizon, trading rule and results after transaction costs. The economic story connects the measurement to a target; evidence must establish whether that connection is useful.
Keep both the activity clock and the release history
A record dated to the activity period may arrive much later. In the shop example, the count of 120 is unavailable to the strategy until the vendor actually releases it and the strategy has received and completed the required processing. Attaching the count to the earlier activity date in a backtest creates look-ahead bias.
Preserve the period measured, publication and receipt timestamps, usable-time rule, timezone and version identifier. If a provider later expands its panel or changes its method, a newly delivered history can differ from the history clients originally received. Replacing old inputs with that reconstructed history changes the historical experiment. A fixed lag cannot recover an original value that has been overwritten.
A vintage is a version of the information available at a specified historical time. The St. Louis Fed's real-time-period documentation illustrates the distinction between today's view of past observations and information known in a past period. A private provider needs its own documented availability contract. An economic-data vintage date does not establish a private feed's intraday receipt time.
The Point-in-Time Data lesson develops version selection, missing-data rules and snapshot checks. A public preview is available; the full lesson requires membership.
Ask who enters the sample
Coverage describes which shops, people, devices, regions or payment channels the feed observes. Selection bias arises when the process determining inclusion systematically distorts an inference about the intended population. More records do not automatically repair that mismatch.
For example, a card panel omits purchases made through other payment methods; a mobile-app panel may favour particular types of users. In its evaluation of transactional data, the U.S. Bureau of Labor Statistics explains that merchant-based card records differ from goods-and-services spending measures and may not represent all consumer transactions. The lesson for a forecast is to define the population it actually addresses.
Track panel membership and methodology changes alongside the feature. A rising visit count can reflect wider collection rather than rising activity. Comparing a stable panel can help separate these explanations, but restricting the panel to entities that survive until the end introduces survivorship bias. The restriction itself needs a historical eligibility rule.
Dataset selection creates another problem: trying many providers and keeping only the one with the best test result makes that result part of the research search. Preserve the trial history and assess a frozen hypothesis on untouched later evidence to limit overfitting.
Permission to buy is not the whole provenance chain
A usable feed needs separate answers about collection rights and consent, the licence granted to the purchaser, privacy and reidentification risk, and potential material nonpublic information (MNPI). A purchase agreement does not, by itself, establish that every upstream collection or downstream use is permitted. Removing names does not by itself establish that location traces cannot identify a person.
These questions depend on the data, use and jurisdiction. In a U.S. investment-adviser context, the SEC examination staff's 2022 MNPI risk alert identified weaknesses in alternative-data diligence, including collection terms and follow-up when collection practices changed. The alert also explicitly distinguishes alternative data from data that necessarily contains MNPI; it is a staff risk alert, rather than a new rule.
Consent and downstream safeguards are concrete issues: the FTC's 2024 final X-Mode/Outlogic settlement announcement describes restrictions on sensitive location data and supplier checks for informed consent. That U.S. case illustrates why the provider's commercial availability is not a complete rights assessment; it does not set a universal rule for every dataset.
For research, retain the source lineage, permitted uses, methodology version and historical delivery record alongside the measurements. Those records let a reviewer distinguish a change in economic activity from a change in collection, processing or entitlement.
Sources
Marcos López de Prado, Advances in Financial Machine Learning (Wiley, 2018), Chapter 2, §§2.2.1–2.2.4: revisions, analytics and alternative sources. The count example and diagram above are original.
