Alternative data in trading research means information outside conventional price feeds, financial statements and company announcements. It can help analysts test whether customer demand, production or business activity supports an investment thesis. Its value depends on what it measures, when it becomes available and whether it adds anything the market has not already priced in.
The practical starting point is a question, not a subscription. “Are sales weakening before the next earnings release?” gives a research team something to test. “What can we do with this enormous dataset?” mostly gives it storage bills.
What Counts as Alternative Data?
Alternative data includes transaction records, web activity, satellite imagery, shipping information and other observations of economic activity. It is not synonymous with unstructured data: a transaction dataset can arrive in tidy rows, while an image needs processing before an analyst can use it. This distinction appears in the CFA Institute report on AI and alternative data in investment management.
| Data category | Research question | What to check |
|---|---|---|
| Aggregated payment transactions | Is consumer spending changing? | Panel coverage, refunds and merchant mapping |
| Website and app activity | Is customer interest translating into usage? | Measurement changes and paid promotions |
| Satellite imagery | Has activity changed at a facility or location? | Cloud cover, observation frequency and interpretation |
| Shipping records | Are goods moving faster or slower? | Reporting delays and incomplete routes |
| Job advertisements | Where might a company be expanding? | Duplicate listings and vacancies that remain unfilled |
Treat these questions as hypotheses, not automatic trading rules. More website visits do not establish higher profits. More job advertisements do not establish higher employment. Each observation needs a defensible connection to the business outcome being forecast.
Connect the Observation to an Investment Thesis
Consider a hypothetical retailer. An analyst wants to estimate quarterly sales before the company reports. A payment panel might offer evidence about spending, but the analyst should first separate growth at existing stores from growth caused by new openings. Refunds, currency movements and changes in the panel also belong in the analysis.
Suppose observed spending rises 8%. That does not justify an 8% revenue forecast without further work. The panel may exclude cash purchases, cover only one country or contain more customers than it did last year. Build the comparison around a consistent sample and reconcile its coverage with the company’s reported business.
Then compare the forecast with market expectations. A business can grow and still disappoint investors. Alternative data should help estimate the gap between likely results and expectations, rather than simply label activity as good or bad. The next step is combining observations with news, prices and company filings, so the signal has financial context.
Check Coverage Before Model Accuracy
Ask the provider what population the dataset represents, how records enter or leave the sample, and how collection methods have changed. Request coverage histories, revision policies and examples of missing observations. A smooth chart is not evidence of a stable sample.
For physical observations, inspect the measurement conditions too. NASA’s nighttime light products include corrections for atmospheric effects, moonlight and other distortions. Its Black Marble product documentation shows why a satellite measurement requires more care than comparing two bright patches on a map.
Apply the same discipline elsewhere. Distinguish zero activity from missing data. Check whether a brand maps to the correct listed parent. Investigate abrupt jumps before interpreting them as business changes. A collection outage should not become a recession forecast.
Respect When the Data Became Available
Maintain separate timestamps for the event, the provider’s publication and your system’s receipt. For a historical simulation, use the version that was available at the decision time, not a corrected file downloaded months later.
Consider transactions recorded on Monday but delivered on Thursday. A Monday trading decision cannot use them. If the provider later revises merchant classifications, preserve the earlier mapping so researchers can reconstruct what the strategy actually knew.
Test the dataset against a simple baseline using price and conventional financial information. Keep later periods outside model development, account for execution costs and examine whether results depend on one company or one unusually favourable period. The related guide to data leakage and overfitting in AI backtests covers those testing problems in greater detail.
Review Data Rights and Information Risk
For US investment advisers, alternative data raises questions about material nonpublic information, or MNPI. The SEC’s April 26, 2022 risk alert on investment adviser MNPI compliance identified inconsistent provider diligence, weak documentation and failures to reassess providers after collection practices changed. The alert records staff observations; it is not a new rule.
Before approving a feed, ask who collected the records, how they obtained them, what permissions support their use and whether confidentiality restrictions apply. Refer unresolved issues to legal or compliance staff rather than treating a vendor assurance as the end of the review.
Keep evidence of approval and establish review triggers. A new upstream supplier, a change in collection method or unexpected customer detail should prompt another assessment. Procurement approval should not last forever simply because the invoice does.
Decide Whether the Feed Earns Its Place
Start with a bounded pilot: one investment question, one dataset and a written success criterion. Measure forecast improvement, research time saved and performance after realistic costs. Include licensing, engineering, storage and review work in the budget.
Retain a feed only when its contribution survives scrutiny. Monitor delivery delays, coverage changes and forecast errors after deployment, and set conditions for suspending its use. Alternative data earns its place when it improves a decision through a traceable process—not when it makes a dashboard look busy.