From 2011 to 2017, satellites photographed the parking lots of 67,078 stores belonging to 44 large US retailers, and a vendor called RS Metrics turned the images into daily counts of how full each lot was. Funds that bought the feed could see a weak quarter before the retailer reported it. Buying the retailers whose lots filled fastest and shorting those whose lots emptied returned 4.6% over the three days around each earnings release, after stock loan fees. Individual investors, who had no access to the images, were net buyers of the retailers about to report bad news.

Alternative data is information about a company, an industry or an economy that comes from outside the company's own disclosures, its regulatory filings and market prices. Card transactions, satellite images, web traffic, app downloads, job postings, shipping records and social media posts are the common types. Much of it is a by-product of another business, such as a payment processor or a smartphone app, which is why the industry also calls it data exhaust.

How alternative data differs from traditional financial data

Traditional data is what a listed company publishes and what the exchange records: annual and quarterly filings, earnings releases, guidance, conference calls, prices and volumes, and the analyst estimates built on all of them. It arrives on a fixed calendar, everyone receives it at the same moment, and it describes a quarter that has already closed.

Alternative data fills the weeks in between. A panel of card transactions shows a restaurant chain's sales by week and by region while the quarter is still running. App download estimates show whether a new subscription product is catching on months before the company reports its subscriber count.

The price of that head start is precision. A card panel sees only the customers of the banks that supply it, a satellite sees only cars, and an app estimate is a model built on a sample of phones.

Funds use the two together. An analyst forecasting a retailer's revenue starts from the reported numbers and uses the card panel to adjust the current quarter. A quantitative fund turns the same panel into a score across hundreds of stocks and trades the gap between the score and the consensus estimate.

The main types of alternative data

Buyers' catalogues sort datasets into about 20 categories, from consumer transactions to weather. Eight of them, with what each one measures:

Card and transaction data
Anonymized purchase records from card issuers, payment processors, bank account aggregators and e-receipt apps, grouped by merchant. They estimate revenue at consumer companies by week, brand and region. Bloomberg bought one of the best-known providers, Second Measure, in December 2020.
Geolocation and foot traffic
Location signals from smartphone apps, turned into visit counts for stores, malls, factories and offices. The supply depends on app users' consent, which is where privacy regulators have stepped in.
Satellite and aerial imagery
Cars in parking lots, oil in storage tanks, construction progress, crop health and ship positions. Acquiring and processing the images costs enough that select hedge funds were the typical clients of the parking lot vendors, while mutual fund managers and sell-side analysts rarely used them.
Web-scraped data
Prices, product ranges, stock levels, job postings and reviews collected from public websites by software. Web-scraped and transactional datasets are the two largest categories by spending, at 15% and 14% of the total in 2024.
Web traffic and app usage
Website visits, app downloads, daily active users and in-app revenue, estimated from panels of devices and browsers or from data that app developers share with the vendor. Sensor Tower bought its main rival in this category, data.ai, formerly App Annie, in March 2024.
Sentiment and text
News, social media posts, earnings call transcripts and product reviews, scored for tone and topic by language models, usually per company and per day.
Employee and hiring data
Employer ratings, job postings and headcounts taken from public profiles. Hiring that speeds up in one division can show where a company is investing before management says so.
Shipping, flight and trade data
Ship positions from transponders, customs records of imports, and corporate jet movements, used to follow supply chains and, now and then, to spot executives visiting a likely merger partner.

Who buys alternative data

Of 107 private fund managers surveyed in November and December 2025, 90% used alternative data, against 67% a year earlier and 62% in 2023. Hedge fund managers made up 41% of respondents, private equity firms 39% and venture capital firms 20%. More than two-thirds had alternative data budgets above 1 million dollars a year.

Quantitative hedge funds were the first large buyers, because they could test a dataset across hundreds of stocks at once. Discretionary stock pickers followed, using the data to check a thesis on a few names before earnings. Private equity firms use it to screen and value companies that publish no figures at all.

The average dataset was sold to about 20 investment clients in 2025, down from 25 in 2024, as the number of datasets on offer grew from 2,215 to 2,805, so each signal was shared by fewer funds on average.

Inside a fund, the work splits three ways. A sourcing team finds and negotiates datasets. Engineers clean them and map every record to the right ticker. Analysts or an equity data science team test whether the numbers lead reported results. At multi-manager firms, data is one of the costs many platforms pass through to their investors along with rent, technology and recruiting, as our guide to multi-manager hedge funds explains.

What alternative data costs

Investment managers spent about 2.8 billion dollars on alternative data in 2025, 17% more than in 2024. The estimate multiplies vendor price ranges by client counts across the datasets with enough detail to price, so it is a floor. Counting the 5,600 thinly documented datasets left out of it at the average revenue per dataset would have lifted the 2024 figure from 2.5 billion dollars to 11.4 billion dollars.

As of June 2026, about 31% of datasets cost 25,000 dollars a year or less, including 4% that are free, down from nearly 40% in the previous count. Higher sourcing costs and vendors setting a separate price for each client drove the change.

Quoted prices are a ceiling. Vendors give a range and deals close nearer the bottom: for a dataset quoted at 26,000 to 150,000 dollars a year, the estimated price actually paid is about 63,200 dollars. Revenue per dataset is lopsided. Among datasets with at least one paying investment client, the average earned about 1.6 million dollars across all its clients in 2024. About 415 datasets earned between 100,000 and 500,000 dollars, and 20 earned more than 20 million dollars from investment firms alone.

Size changes the arithmetic. A 300 million dollar fund charging the 2026 average management fee of 1.32% collects 3.96 million dollars a year. Ten datasets at 63,200 dollars each cost 632,000 dollars, 16% of that fee income, before anyone is hired to clean and test them. For a 10 billion dollar fund at the same fee, the same 632,000 dollars is under half a percent of its 132 million dollars of fees.

Prices keep rising. Among the same 107 managers, 81% of managers saw cost increases on data products that had added AI tools, and 89% still planned to raise their alternative data budgets in 2026.

The subscription is only part of the bill. A new dataset also needs a trial period, a legal and compliance review, storage, and engineering time to map millions of records to tickers that change with every merger, spin-off and rebrand.

A worked example: turning a card panel into a revenue estimate

The numbers below are illustrative. A retailer reported 2.0 billion dollars of revenue in the third quarter of last year. A card panel recorded 40 million dollars of spending at its stores in that quarter, so the panel sees about 2% of sales.

  1. This year the panel records 43 million dollars for the same weeks, a 7.5% rise.
  2. Over past quarters, panel growth has run 1.5 points above the growth the company later reported, because the panel's cardholders shop there more than average. The adjusted estimate is 6.0% growth.
  3. Applied to last year's 2.0 billion dollars, that gives 2.12 billion dollars of revenue.
  4. Analysts expect 4% growth, or 2.08 billion dollars. The data points to revenue about 1.9% above consensus.

Each step carries an error. The panel misses cash and online orders paid through other processors, gift cards spent this quarter were bought last year, and a retailer's fiscal quarter may end on a different date from the calendar quarter the panel uses. A fund that trades on a 1.9% gap needs the panel's past forecast errors to be well under 1.9%, measured over enough quarters to trust them.

Does alternative data still predict returns?

Three published studies tested whether a dataset would have made money, each on a different kind of data:

DataSampleResultPublished in
Real-time sales measures, including signals from about 50 million mobile devicesUS retailers3.4% average excess return at the earnings announcementJournal of Financial Economics, 2017
Employer ratings posted by employeesListed US firms, 2008 to 2016Firms with the largest rating gains beat those with the largest falls by 0.74% a month over the next quarterJournal of Financial Economics, 2019
Satellite counts of cars in parking lots44 US retailers, 2011 to 20174.6% long/short return over the three days around earnings, after stock loan feesJournal of Financial and Quantitative Analysis, 2025

The returns cluster around earnings announcements. The data tells its buyers what the quarter looked like before the company does, and the price moves on the day the company confirms it. The mechanism matches the argument Sanford Grossman and Joseph Stiglitz made in 1980: prices cannot fully reflect information that costs money to collect, or nobody would pay to collect it. Our page on the efficient market hypothesis covers its weak, semi-strong and strong forms.

Across 97 return predictors from academic research, long/short returns were 26% lower out of sample and 58% lower after publication, as investors traded on what they had read. The satellite feeds had no measurable effect on the speed of price discovery at first, and their effect on prices built up over the following years as more funds adopted alternative data. Across the market, the arrival of card and satellite data made stock prices more informative, and managers of the covered companies cut back their opportunistic trading in their own shares.

Short-term data also changes what analysts forecast. Across 65 million analyst forecast observations, the informativeness of one-year forecasts has risen about 10 points since 2000, while that of five-year forecasts has fallen about 10. Analysts more exposed to StockTwits, a social network for traders, produced short-term forecasts 0.54 to 0.66 percentage points more informative and long-term forecasts 1.51 to 1.92 points less informative. When near-term information gets cheaper, forecasters move their attention to the next quarter and away from the next five years.

Why alternative data signals stop working

  • Panel bias. A card panel reflects the customers of the banks and apps that supply it. When a supplier leaves, the panel changes and the series breaks, which can look like a change in the company's sales.
  • Short histories. A dataset with five years of history covers 20 quarterly reports per company. A signal tested on so few events, and picked from many signals tried, can look strong by chance. Our article on factor investing covers the same problem in the factor zoo.
  • Vendor churn. In early 2025 about 450 datasets were labelled discontinued in Neudata's catalogue, fewer than 100 of them in the previous year. A model built on a feed that stops has to be rebuilt on another source with a different history.
  • Falling quality. A third of data buyers surveyed in 2026 said the quality of their data sources had worsened over the previous one to two years, and about three in five of them blamed AI-generated content at least in part.
  • Estimates that move. Download and usage figures are model outputs, and vendors revise them. In the App Annie case, the adjustments were made with confidential company data.

When alternative data becomes material nonpublic information

Under US law, trading on material nonpublic information is illegal when the information was obtained or used in breach of a duty of trust or confidence, owed to the company or to whoever supplied the information. Rule 10b5-2 lists when that duty exists in misappropriation cases, starting with an agreement to keep information confidential.

Card spending aggregated across millions of consenting consumers breaches no duty to the company whose sales it measures. The risk sits in how a dataset is sourced:

  1. A company shared its own data under a promise of confidentiality, and the vendor passed it on in a form that reveals that company's numbers.
  2. The data was collected in breach of a website's terms of use or from behind a login.
  3. Personal data was collected without valid consent, a privacy violation even where no securities law is broken.

The App Annie case

App Annie sold estimates of app downloads, usage and revenue to trading firms. Companies shared their confidential app performance data with it on the promise that the data would be aggregated and anonymized before it fed a statistical model. From late 2014 to mid-2018, App Annie used non-aggregated, non-anonymized data to alter its estimates and make them more valuable to sell to trading firms. It told those firms that it had effective controls against the misuse of confidential data, and it suggested ways to use the estimates to trade ahead of earnings announcements.

In September 2021 App Annie agreed to a 10 million dollar penalty. Its co-founder and former chief executive, Bertrand Schmitt, agreed to pay 300,000 dollars and to a three-year bar from serving as an officer or director of a public company. It was the SEC's first enforcement action charging an alternative data provider with securities fraud.

The SEC's 2022 examination findings

Seven months later the SEC's examination staff published a risk alert on how advisers handle MNPI. On alternative data, examiners found advisers with no policies on the risk of receiving MNPI through it, advisers that ran due diligence on some data sources and skipped others, and advisers with no procedure for checking the terms and legal obligations under which a vendor collected its data.

Web scraping after hiQ v. LinkedIn

In April 2022 the Ninth Circuit held that scraping profiles LinkedIn had made public was unlikely to be access "without authorization" under the Computer Fraud and Abuse Act, the federal anti-hacking law. The dispute continued over LinkedIn's user agreement. In December 2022 hiQ accepted a permanent injunction that required it to stop scraping LinkedIn and to delete the data, code and algorithms built from it.

Location data and the FTC

Foot-traffic panels are built from app location signals bought from data brokers. In January 2024 the FTC announced its first ban on a data broker selling sensitive location data. X-Mode Social and its successor Outlogic had sold precise locations that could reveal visits to medical clinics and places of worship. The final order, approved in April 2024, bars the company from sharing or selling sensitive location data.

How funds vet a dataset before buying

Investors ask about this process in hedge fund due diligence questionnaires. A typical review runs in this order:

  1. Trace provenance: who collected the data, from whom, and under what consent or contract.
  2. Ask whether any company supplied its own data, and on what confidentiality terms.
  3. For scraped data, check that only public pages are collected, that no logins are used, and that each site's terms of use have been reviewed.
  4. Confirm that personal information is removed or aggregated before delivery.
  5. Put representations in the contract that the data contains no MNPI and was lawfully obtained, with an indemnity and a right to audit.
  6. Test a trial sample, get compliance sign-off, and repeat the review whenever the vendor changes its sources.

Alternative data in consumer lending

The term has a second meaning in credit. Lenders use rent, utility and phone payments, and the cash flow in a bank account, to score people whose credit files are thin or empty. In June 2025 the CFPB cut its long-quoted count of credit-invisible adults roughly in half and put the share at 2.7%, about 7 million people, in December 2020.

Since September 18, 2021, Fannie Mae's automated underwriting system can count on-time rent payments found in bank statement data that the applicant agrees to share. In a sample of applicants who had not owned a home in the previous three years and had not received a favorable recommendation, 17% could have been approved if their rent history had been counted.

Alternative data an individual investor can use

Paid feeds are priced for institutions, and the satellite study shows what that means at earnings time. Short sellers with the data built positions in the days before bad news, and individual investors bought. The public sees short interest only twice a month, with a delay, too late to follow those trades.

Free sources cover part of the same ground:

  • Search interest by product or brand, from Google Trends.
  • Job postings on a company's own careers page, counted over time.
  • App store rankings for a company's apps.
  • Employer ratings on review sites, the signal behind the 0.74% a month result above.
  • Government statistics on retail sales, freight and energy inventories.

None of these comes with a cleaned history mapped to tickers, so testing one takes the work a vendor would otherwise have done.

This article is general information and does not constitute investment advice.

Frequently asked questions

What is an example of alternative data?

Satellite counts of cars in retailers' parking lots are the best-known example. From 2011 to 2017, daily counts covering 67,078 stores of 44 US retailers let traders spot weak quarters before they were reported, and a long/short strategy built on the counts earned 4.6% over the three days around each earnings release. Card transaction panels, app download estimates, web-scraped prices and job postings are other common examples.

How do hedge funds use alternative data?

Mostly to estimate a company's revenue or a key operating figure before the company reports it, then to compare that estimate with the analysts' consensus. Quantitative funds turn a dataset into a signal across hundreds of stocks. Discretionary managers use it to check a thesis on a few holdings ahead of earnings. Of 107 private fund managers surveyed in late 2025, 90% used alternative data.

Who are the main alternative data providers?

The market is fragmented, with 2,805 datasets counted in the 2025 estimate of spending. Well-known names include Second Measure for card data, owned by Bloomberg since 2020, Sensor Tower for app data, which bought data.ai (formerly App Annie) in 2024, and RS Metrics for satellite parking lot counts. Catalogues such as Neudata and Eagle Alpha help buyers find and compare datasets.

Is alternative data legal?

Yes, when it is lawfully collected and carries no information obtained in breach of a duty of confidence. The SEC's 2021 case against App Annie, settled for more than 10 million dollars, involved app estimates altered with confidential company data that the vendor had promised to aggregate and anonymize. Privacy law applies separately: in 2024 the FTC barred the data broker X-Mode Social and its successor Outlogic from selling sensitive location data.

How much does alternative data cost?

About 31% of datasets cost 25,000 dollars a year or less as of June 2026, and 4% are free. At the top, 20 datasets each earned more than 20 million dollars a year from investment firms in 2024. More than two-thirds of the private fund managers surveyed in late 2025 had alternative data budgets above 1 million dollars a year.