Hedge funds and AI · AI in investing
Alternative data: satellite images, card spending and the law
What alternative data is, how funds use card panels, satellite images, web and app data to estimate results early, and the insider trading and privacy limits on it.
On this page
Alternative data is information about companies and the economy that does not come from the usual sources: financial statements, filings, press releases and market prices. Satellite photos of store parking lots, totals from millions of card transactions, website visits, app downloads, shipping records and the tone of social media posts are all alternative data. Investment firms buy it to learn, before a company reports, roughly what the report will say.
It is a real business and a real legal risk. Data that seems anonymous can be traced to people, and data that a vendor obtained in breach of a promise can be material non-public information, which is illegal to trade on. This lesson explains the main kinds, how funds turn them into estimates, why a bigger dataset is not always a better one, what it costs, and the insider trading and privacy limits on all of it, with the US cases that drew the lines.
What counts as alternative data
The SEC's examination staff gave a working definition in an April 2022 risk alert: information increasingly used in financial analysis beyond traditional financial statements, company filings and press releases. Its examples were satellite and drone images of crop fields and retailers' parking lots, aggregate credit card transactions, social media and internet search data, geolocation data from consumers' phones, and email data from apps. The staff added that alternative data does not necessarily contain material non-public information.
| Kind | What it measures | What it can estimate | Main weakness |
|---|---|---|---|
| Card transactions | Spending at named merchants by a panel of cardholders | A retailer's or restaurant chain's sales | The panel is not the whole customer base |
| Satellite and drone images | Cars in lots, crops, oil tank levels, ships | Store traffic, harvests, inventories, trade | Weather, cloud cover, and what a photo cannot show |
| Web traffic and search | Visits, searches, prices and stock on websites | Demand, pricing power, product launches | Bots, changes in how sites are built, scraping limits |
| App data | Downloads, active users, time in app | User growth for app-based businesses | How the vendor collected it; consent and privacy |
| Location data | Phones visiting stores, factories, offices | Foot traffic, factory activity | Privacy law and re-identification |
| Text and sentiment | Tone of news, social posts, job ads, reviews | Shifts in demand, hiring, reputation | Noise, manipulation and fake accounts |
Turning raw images, transactions and text into numbers is where machine learning and language processing come in (AI in investing: what it does at funds and brokers, and what it can’t), and judging whether a signal from the numbers is real is the problem described in Machine learning in quant funds: features, overfitting and regimes.
How funds use it
The most common use is nowcasting: estimating a number for a period that is over, or nearly over, but not yet reported. A company's sales for the quarter are fixed the day the quarter ends, but investors do not see them until the company reports weeks later. Alternative data narrows that gap.
The estimate is worth something only if it differs from what other investors already expect, and only if it is accurate enough to trust. A fund compares its nowcast with analysts' consensus forecasts. If the card data points to sales well above consensus, and the data has tracked reported sales well in past quarters, the fund may buy before the report. If the data agrees with consensus, it has learned little that is not already in the price.
Funds also use alternative data to check a story ("is the new product really selling?"), to watch risks (falling traffic at a key customer), and for economic and commodity views, such as estimating shipments, crop sizes or oil in storage before official figures appear. The company's own reports remain the final word: Form 8-K: what triggers it, the item numbers and how to read one and the quarterly 10-Q are what the estimate is ultimately scored against.
Why a bigger dataset is not always a better one
A card panel sees only its own members. Two kinds of error follow, and they behave differently.
- Random error comes from the luck of who is in the sample. It shrinks with the square root of the panel's size: a panel 100 times larger has one tenth of the random error.
- Bias comes from who tends to be in the panel. If the panel's members are younger, richer or more likely to shop online than the store's customers as a whole, their spending can grow at a different rate. A bigger panel of the same kind of people has the same bias.
Vendors and funds fight bias by comparing past panel estimates with the sales companies later reported and adjusting for the gap. That works until the panel's make-up changes, for example when a card issuer that supplies the data loses or gains a group of customers. Precision is easy to buy; accuracy has to be checked every quarter.
Costs and decay
Alternative data is expensive in ways that go beyond the price of a subscription. Raw data has to be cleaned, matched to the right companies and tickers, checked against reported results and watched for changes in how it is collected. That takes data engineers, researchers and lawyers. The legal review alone, discussed below, is a standing cost.
Its value also decays. The first funds to use a dataset can learn something the market does not know. As more buyers use the same data, its information reaches prices earlier, and the surprise left on the reporting day shrinks. A dataset that once gave a lasting edge may end up only confirming what prices already show. Why trading strategies stop working describes the same fading for trading rules. For this reason the economics of alternative data favour large firms that can spread the costs over many positions.
The legal limits
Material non-public information
US insider trading law forbids trading on material non-public information (MNPI) in breach of a duty. Information is material, as the SEC put it in its 2000 release adopting Regulation FD, when there is a substantial likelihood that a reasonable shareholder would consider it important in making an investment decision. Under the misappropriation theory, it is also illegal to trade on information taken in breach of a duty of trust or confidence owed to its source; SEC Rule 10b5-2 lists situations that create such a duty, including an explicit agreement to keep information confidential.
That is why alternative data is checked so carefully. Aggregated, lawfully collected data that anyone could buy is generally not MNPI. Data that a vendor obtained by breaking a confidentiality promise, from an employee leaking a company's figures, or under terms that forbid its use for trading, can be. The same Regulation FD release describes the mosaic: an analyst may assemble many pieces of non-material information into a conclusion that is valuable, and that is legitimate research. The question is always where each piece came from.
Insider trading defines the term, and Insider trading filings: Forms 3, 4 and 5 and how to read them covers the separate, legal trading by corporate insiders that they must report on Forms 3, 4 and 5.
What SEC examiners found
The SEC's April 2022 risk alert reported that examiners had seen advisers use alternative data without written policies reasonably designed to address the risk of receiving MNPI through it. Some advisers ran ad hoc, undocumented diligence on data vendors. Some had no policies for assessing the terms, conditions or legal obligations attached to how the data was collected, including when red flags appeared. Some had an onboarding check for vendors but no process for repeating it as time passed or collection practices changed. In its examination priorities for fiscal 2027, published October 23, 2026, the Division of Examinations still lists the risks of alternative data sources among the areas it reviews.
The App Annie case
On September 14, 2021 the SEC charged App Annie, a seller of mobile app performance data, and its co-founder and former chief executive, Bertrand Schmitt, with securities fraud. The SEC called it its first enforcement action charging an alternative data provider with securities fraud. According to the SEC, App Annie told the companies that shared confidential app data with it that the data would be aggregated and anonymized before being used in its estimates. From late 2014 to mid-2018, it instead used raw, non-aggregated data to adjust the estimates it sold to trading firms, and it misrepresented to those firms how the estimates were made and how strong its compliance controls were. Without admitting or denying the findings, App Annie agreed to pay a $10 million penalty, and Schmitt a $300,000 penalty and a three-year bar from serving as an officer or director of a public company.
The case shows the risk from both sides: the companies that shared data were misled about its use, and the funds that bought the estimates were misled about where they came from.
Privacy
Data about people's spending, browsing and movements is personal data, even when the names are removed. In June 2024 the Federal Trade Commission finalized an order against Avast, a security software maker, which it said had sold detailed, re-identifiable browsing data collected through its software to more than 100 third parties through a subsidiary, Jumpshot, without adequate notice or consent. The order required a $16.5 million payment, banned Avast from selling or licensing browsing data for advertising, and required it to delete the data and any products or algorithms derived from it. The FTC's case was about consumer deception, not trading, but the lesson carries over: a buyer of "anonymized" data relies on the vendor's consent and collection practices, and an order can reach the models built from it.
Where alternative data fails
- Bias. A panel, an app's users or a satellite's view is a sample, not the whole picture.
- Change. Vendors change their collection, sources join and leave, and a website redesign can break a scraper overnight.
- Short histories. Many datasets cover only a few years, too few quarters to tell a real relationship from a lucky one.
- Too many datasets. Testing hundreds of feeds against prices is a search, with the overfitting that comes with it.
- Crowding and decay. Once many funds hold the same data, the edge moves into prices earlier.
- Legal and reputational risk. A data source that breaks a confidentiality promise or privacy law can turn research into an enforcement case.
How Finabase does it
Finabase does not sell alternative datasets, and its pages do not show card, satellite, app or location data. A symbol page (The Finabase symbol page: quote, chart, financials and news) brings together the public, traditional sources that alternative data tries to anticipate: the live quote and chart, company financials and earnings, headlines on the News tab, and on the Holders tab the top institutions and funds from quarterly 13F filings (13F filings: what institutional holdings reports show and miss), insiders and recent Form 4 trades. The insider trading page summarizes recent insider purchases and sales across the market (Earnings calendar, ex-dividend dates and insider trades on Finabase). All of it is public, and none of it is a recommendation.
Frequently asked questions
What is alternative data in investing?
Information about companies or the economy from outside the usual financial statements, filings and press releases, such as card spending, satellite images, web traffic, app usage and social media. Funds use it to estimate results before companies report them.
Is using alternative data legal?
Generally yes, if it was lawfully collected, contains no material non-public information obtained in breach of a duty, and respects privacy law and the terms under which it was gathered. The SEC expects advisers that use it to have written policies to check those points.
Do hedge funds really use satellite images?
Some do, along with many other datasets. The SEC's own examples include satellite and drone images of crop fields and retailers' parking lots. How much any dataset adds after its cost is hard to see from outside.
Can individual investors use alternative data?
Large commercial datasets are priced for institutions. Individuals can still read the primary public sources well: company filings, insider trading reports, institutional holdings and news.
Summary
Alternative data is non-traditional information, from card spending and satellite photos to app and web activity, that funds use to estimate company results before they are reported. It can work, but a larger dataset only cuts random error, not bias, and every edge fades as more buyers arrive. The limits are legal as well as statistical: data obtained in breach of a duty of confidentiality can be material non-public information, the SEC has charged a data vendor with fraud, and privacy regulators have ordered the deletion of data and the models built from it.
Check your understanding
4 questions on this lesson. Each answer comes with the reason, and nothing is scored.
A fund grows its card-spending panel from 10,000 to 1,000,000 cardholders. The panel’s members are younger and shop online more than the store’s customers overall. What improves?
- The random error shrinks, but the bias from who is in the panel stays the same
- Both the random error and the bias disappear
- Nothing improves, because sample size never matters
Show the answer
A. The random error shrinks, but the bias from who is in the panel stays the same Random error falls with the square root of the panel size: in the worked example, from ±1.57 to ±0.16 percentage points. Bias comes from the kind of people in the panel, and a bigger panel of the same people keeps it. The estimate becomes precise but can still be 2 points wrong.
Why do funds care whether an alternative data vendor collected its data in line with its promises?
- Data obtained in breach of a duty of confidentiality can be material non-public information, and trading on it can be insider trading
- Because only data from public filings may legally be used in research
- Because the SEC approves each dataset before funds may buy it
Show the answer
A. Data obtained in breach of a duty of confidentiality can be material non-public information, and trading on it can be insider trading Under the misappropriation theory, trading on material information taken in breach of a duty of trust or confidence is illegal. In the App Annie case, the SEC said the vendor used confidential data it had promised to aggregate and anonymize. Lawfully collected, aggregated data is generally not MNPI, which is why funds check the source.
What is the most common way funds use alternative data such as card transactions?
- To estimate a company’s sales for a period before the company reports, then compare that with what analysts expect
- To replace the company’s reported results, which are no longer needed
- To learn exactly what a stock’s price will be on a given day
Show the answer
A. To estimate a company’s sales for a period before the company reports, then compare that with what analysts expect This is nowcasting. A quarter’s sales are fixed when it ends but not reported for weeks. An estimate is valuable only if it differs from the consensus and has tracked reported results well before; the company’s filing remains the final word.
Why does an edge from a new alternative dataset tend to fade?
- As more funds buy and trade on the same data, its information reaches prices earlier and less surprise is left
- Satellites and card networks stop working after a few years
- The SEC bans each dataset after two years of use
Show the answer
A. As more funds buy and trade on the same data, its information reaches prices earlier and less surprise is left The first users can learn something the market does not know. Once many buyers hold the same data, prices adjust before the company reports, so the dataset ends up confirming what prices already show. Costs of buying, cleaning and checking the data stay high, which favours large firms.
Sources and further reading
- SEC Division of Examinations — Risk Alert: Investment Adviser MNPI Compliance Issues (April 26, 2022)
- SEC — Press release 2021-176, SEC charges App Annie and its founder with securities fraud (September 14, 2021)
- SEC — Final Rule: Selective Disclosure and Insider Trading (Regulation FD), Release 33-7881 (August 15, 2000), on materiality and the "mosaic"
- Cornell LII — 17 CFR 240.10b5-2, duties of trust or confidence in misappropriation insider trading cases
- FTC — Press release: FTC finalizes order with Avast banning it from selling or licensing web browsing data for advertising and requiring it to pay $16.5 million (June 27, 2024)
- SEC Division of Examinations — Fiscal Year 2027 Examination Priorities (October 23, 2026)
- Finabase calculations: sampling error and bias for an invented card panel and retailer
