Alternative data is information used to study economic activity but created outside familiar financial sources such as market prices, official company reports filed with regulators and press releases. Satellite images, website visits, aggregated card transactions and online reviews are common examples. An unusual dataset is not automatically useful or reliable.
From raw observation to a reading
The starting point is an observation of the real world: vehicles in a car park, ships in motion or use of an app. Turning it into research requires traceable provenance, correct dates and a comparison with the activity it is supposed to measure. Alternative data complements official company reports and prices; it does not replace them by definition.
“Alternative” describes the source relative to traditional financial data. “Unstructured” describes the format. A satellite photograph can be both; an aggregated spending table may be alternative while already being structured.
Checks before use
The first question is what the sample truly represents. Customers of one payment provider, for example, may not resemble all customers of a company. Geographic coverage, observed devices and changes in collection methods can create movements that look economic but are only measurement changes.
Next, reconstruct the data chain: who collected it, which transformations it underwent, which rights permit its use and when it arrived. Privacy, licences and contractual terms are not minor technical details. The applicable duties depend on the jurisdiction and on the organisation using the data.
Finally, compare cost with marginal usefulness. Purchase, cleaning, company mapping and maintenance may cost more than the information gained. A signal that worked in the past can also weaken when the sample, observed behaviour or number of market participants using it changes.
Advanced: time, leakage and decay
A professional test preserves point-in-time versions: on each date it uses only what was available then. If a backtest includes corrections published later, future information enters the past and creates data leakage. Retrospectively removing failed companies or vanished users can also introduce survivorship bias.
Out-of-sample validation helps limit overfitting, but it is not enough. Delays, revisions, coverage and entity mapping—matching observations to the correct companies—still need monitoring. A change in any of them can change the economic meaning of the signal.
Alternative data does not necessarily contain material non-public information. The staff of the SEC's Division of Examinations has nevertheless reported deficiencies at some US advisers involving vendor diligence, collection terms and potential MNPI risk. A defensible process therefore documents provenance, red flags, periodic reviews and use decisions. The dataset is a research input, not a legal shortcut or a promised return.
Sources
- SEC Division of Examinations, Investment Adviser MNPI Compliance Issues — US definition, examples and observations on alternative-data diligence.
- CFA Institute, Data science and AI: A guide for investment managers — the data lifecycle, non-traditional sources, quality and governance principles.
- CFA Institute Research and Policy Center, Unstructured Data and AI — the distinction between alternative and unstructured data and responsible research use.
- CFA Institute Research Foundation, Investment Model Validation: A Guide for Practitioners — survivorship and look-ahead bias, out-of-sample tests and time-ordered validation.