Alternative Data in Investing
Table of Contents
Alternative data in investing uses observations such as card purchases, search interest or store visits to answer questions that company filings and market prices alone cannot resolve. A retailer’s customers may spend more this week, for example, but the card feed might also have added a payment provider. Before treating the increase as a signal, ask what changed: the business, the sample or the way the vendor measured it.
Start with a question you can test. Then find out what the data cover, when you could have used them, and whether they improve a forecast or an investment decision after costs.
What Counts as Alternative Data?
Alternative data come from outside traditional investment inputs such as issuer filings, financial statements and market prices. They can be public or licensed. Their value depends on what they measure and how closely it relates to the decision.[1]
| Source | Possible use | What to check |
|---|---|---|
| Card transactions | Track spending at covered merchants. | Which cards, payment channels and geographies are missing? |
| Store visits or satellite imagery | Track visits or count cars at observed locations. | A device or car count may include staff and repeat visits. It does not tell you who bought anything. |
| Search interest | Track changes in attention to a brand or product. | An indexed sample of searches is not a count of orders. News or complaints can also draw attention.[2] |
| Public product listings | Track advertised prices, promotions or stock status. | An offer or stockout does not reveal how many units sold. Check what the data source allows you to collect and reuse. |
In fundamental analysis, a quarterly filing gives you the company’s reported revenue. A card feed samples purchases and may miss categories that the company includes. Align fiscal periods, divisions and geography before comparing the two. Use the time the relevant information was available: the quarter end and the time EDGAR accepts a filing are different, and the SEC does not guarantee when an accepted filing first appears on its website.[3] An earlier public earnings release can supply the figure. Channel checks help investigate operating activity; a vendor feed can inform that work if its coverage matches the question.
Check Coverage, History and Delivery
Find out who the sample represents
Ask for the data dictionary, a sample file and a history of changes. For a card panel, check which providers and merchants contribute, whether geography refers to a shopper’s home or the store, and how refunds, duplicate transactions and missing observations are handled. Ask what started on each date: collection, coverage of a market, or verified availability to customers. A five-year file may contain only two years of comparable, usable history. See how panel data are structured for why the units you follow matter.
The Opportunity Insights Economic Tracker illustrates the difference. Its public consumer-spending series uses aggregated credit and debit data supplied by Affinity Solutions. It groups spending by where the cardholder lives and uses the median income of the ZIP code for income groups. It does not report individual companies’ sales. The publisher documents a switch from daily observations shown as seven-day averages to weekly observations in June 2022. It marks the latest three weeks provisional because it needs more observations to identify breaks from payment providers entering or leaving.[4]
Use the version available at the decision time
A vintage is a version of the data available at a particular time. The date on a row tells you what period it describes. For a test, separately record when the publisher released each version and when your own process could use it. A correction can change an existing row; a backfill can add old rows that were missing at the time.
To reproduce a revision, open both archived copies of the Tracker’s national Affinity CSV and look up the row whose year, month, day and frequency are 2023,7,16,w. The w indicates a weekly observation despite “Daily” in the filename. Read spend_all, the seasonally adjusted spending-change series for all merchant categories.[4][5]
| Archived file | Repository commit timestamp (UTC) | July 16 value |
|---|---|---|
| July 25 CSV | July 25, 2023, 16:04:57 | 0.190 |
| August 15 CSV | August 15, 2023, 17:53:02 | 0.197 |
After seasonal adjustment, 0.190 and 0.197 mean spending was 19.0% and 19.7% above the January 2020 reference period (January 6 to February 2, 2020), respectively. Neither number measures growth from the previous week. The revision is 0.197 − 0.190 = 0.007, or 0.7 percentage point. The commits prove that the two archived files differ. Their timestamps do not prove when either value first became publicly available, much less when a particular analyst received it. If you cannot establish availability by the decision cutoff, move the test to a later cutoff you can verify or leave that observation out. This is how a point-in-time backtest avoids look-ahead bias: each decision uses only information that would have been available then.[5][6]
Check the License, Privacy and MNPI Separately
- License: Ask who collected the data, how they got it, and whether they can license it for your investment use. Check whether you can store past versions, how long you can keep them, and what you can do with derived scores or shared research. Get the applicable terms and evidence before adopting a feed.
- Privacy: Ask what people were told, what permissions apply, and how the vendor limits exposure of individuals, especially in small groups or sensitive locations. Avoid trying to identify people from aggregates. In the InMarket case, the FTC alleged notice and consent deficiencies; the final order restricted that company’s use of location data and required supplier checks.[7]
- Material nonpublic information (MNPI): A fact can be material if there is a substantial likelihood that it would significantly change the total mix of information available to a reasonable investor. The U.S. insider-trading analysis can turn on how the information was obtained and whether duties of trust or confidence were breached. A paid subscription or an “aggregated” label does not settle the question. If you cannot establish the source or suspect the feed contains confidential issuer information, pause use and seek compliance or legal guidance before trading.[8]
Section 204A requires covered U.S. investment advisers to establish, maintain and enforce written policies and procedures reasonably designed to prevent unlawful misuse of MNPI. The SEC’s separate Code of Ethics Rule applies to advisers registered or required to register under section 203. SEC examination staff flagged gaps in documenting vendor diligence and checking whether collection practices changed; their risk alert describes staff observations, not a new rule.[1][9]
Vendor representations deserve special attention. In a settled App Annie order, the SEC found that the company used nonaggregated, nonanonymized confidential app data to alter estimates sold to trading firms, contrary to its representations. The respondents did not admit or deny the findings, subject to the exceptions in the order. Ask for evidence that the source permissions and data handling match what the vendor says.[10]
Worked Example: Turn Card Spending Into a Signal
Suppose you want to know whether in-store card spending at fictional Birch Street Apparel and comparable U.S. retailers improves estimates of quarterly sales growth. You’ll compare a forecast that uses the card data with one based on public information alone. Separately, you’ll test a trading rule. Work with 2022–2023 decisions to develop the method and 2024 decisions to tune it, using only outcomes already known for each fitting or tuning step. Lock the rules by June 30, 2025, then set aside decisions from July through December 2025 as the untouched test, or holdout.[6]
Compare the same markets and adjust for sample size
The fictional license permits internal investment research and storage of the archived versions and derived scores during its retention period. It does not allow publication of raw files. The feed covers completed card purchases net of refunds at physical stores in selected Midwest markets. It misses cash, private-label cards and online checkout. A new provider adds markets this year, so compare only markets and providers covered in both periods. Divide covered spending by the number of active cards in the comparable regional feed each week, including cards that did not shop at Birch Street.
| Observation | Card spending | Active cards | Spend per card |
|---|---|---|---|
| Week ending Sept. 8, 2024: comparable markets | $100,000 | 1,000 | $100 |
| Week ending Sept. 7, 2025: first file, all covered markets | $130,000 | Not used | Not compared |
| Same 2025 week: first file, comparable markets | $111,300 | 1,050 | $106 |
| Same 2025 week: revised file, comparable markets | $107,100 | 1,050 | $102 |
The apparent 30% increase compares different markets. In the first file, $18,700 of the $130,000 came from markets that weren’t covered a year earlier. After matching markets and dividing by active cards, growth is $106 ÷ $100 − 1 = 6%. The revision gives $102 ÷ $100 − 1 = 2%. Per-card spending adjusts for the number of active cards, but the mix of shoppers can still change. And a higher dollar amount can reflect higher prices rather than more units sold.
Make one decision per retailer and reference week
Before the holdout, the trading rule was fixed: evaluate each retailer and reference week once, at the first 9:15 a.m. Eastern cutoff on a trading day after the initial file becomes usable. Use that first usable version for the comparison. If it shows comparable spending per card up at least 5% from the matched week a year earlier, buy once at that session’s regular open. Otherwise, skip that observation permanently. Rereading the file or receiving a revision cannot trigger another entry for that retailer and week. Exit any trade at the open five trading sessions later. For this observation, the first file is released after the Wednesday close:
- Sunday, September 7: The reference week ends.
- Wednesday, September 10, 4:15 p.m. Eastern: The first file is released, three calendar dates after week end. Validation finishes at 4:30 p.m.
- Thursday, September 11, 9:15 a.m. Eastern: The decision uses the available 6% figure. It clears the 5% threshold, so the simulation enters at the 9:30 a.m. open.
- Friday, September 19, 10:00 a.m. Eastern: Late refunds revise the figure to 2%. That would miss the threshold, but it cannot change what was known on September 11.
Keep the released file ID, release and validation times, and the merchant mapping used for each decision. A later version of any of them can quietly change which trades appear in a backtest.
Validate Forecasts and Trades Separately
Forecast test: For the public-only baseline, forecast the same sales level the retailer reported for the comparable quarter a year earlier. That predicts 0% year-over-year sales growth.[6] Compare it with a forecast that adjusts that baseline using the card signal. Predict the same quarterly sales growth for each retailer using the same information cutoff and the same holdout observations. If you test additional public inputs, such as a retail survey, use the same available releases in both forecasts. The Census Bureau distinguishes the month its survey measures from the release date, and ALFRED preserves dated versions of supported economic series.[11] Once the company releases results, compare the two forecasts’ absolute errors in percentage points of year-over-year sales growth. Report how often the panel was usable and check whether improvements extend beyond one region or retailer.
Trading test: Better estimates of sales need not produce better trades. Investors may already expect the improvement, or weak margins may offset stronger sales. Research that studies operating estimates, earnings surprises and stock returns treats these as different outcomes.[12] Compare returns against a relevant benchmark over the same period, with turnover and realistic execution costs. In a momentum strategy, as in a card-based strategy, excluding companies that later delist can distort the historical test.
For Birch Street, suppose $10,000 of opening-price exposure gains 0.40% before costs, with an exit at the open five trading sessions later, on September 18. The gain is $10,000 × 0.004 = $40. Assume $15 per leg for spread, fees and market impact combined, or 15 basis points of starting notional on each leg. The trade keeps $40 − $15 − $15 = $10, or 0.10% of the starting notional, before the data subscription and taxes. It needs a 0.30% gross return just to cover those execution costs. These costs are assumptions, and the example says nothing about typical trading costs or strategy performance.[13]
Check how fast the signal fades. Before opening the holdout, specify delayed entry tests. For example, compare entry on the first eligible open with entry one session later for the same qualifying observations, using the same holding period and separately estimating costs for each schedule. Test other holding periods separately so changes in entry lag and holding period don’t get mixed up. If the benefit disappears with the delay you expect in practice, that matters more than the best result at an earlier cutoff. Changing rules after looking at the holdout calls for a new test period.
Common Mistakes
- Counting the same evidence twice. Two vendor products can rely on the same upstream feed. Check whether apparent corroboration adds independent observations.
- Confusing a missing value with zero. Missing, delayed or suppressed data do not show that customers stopped buying. Record why the value is missing, and apply a rule set before the test.
- Keeping only companies or providers still covered today. Build the eligible universe from historical records, including losses of coverage and delistings. A benchmark survivorship example shows why omitted histories can change the result.
- Turning one good estimate into a purchase decision. Check out-of-sample accuracy and net returns before deciding whether the data license, maintenance and monitoring are worth it.
When a Signal Stops Helping
A result can stop working because the feed changed or because the market changed. Monitor delivery lags, covered merchants, missing data and source changes separately from forecast error and net trading results. A provider switch may show up first as a change in the number of cards before it changes the measured growth rate. A signal can also lose its trading value while still tracking the business. Keep original results and compare the new picture against a simple baseline; model risk management explains how to set up that monitoring.
Frequently Asked Questions
Sources and Notes
Sources checked November 18, 2026. Birch Street Apparel, the panel and the trade are hypothetical. The separately attributed Opportunity Insights example uses published archived files.
- SEC Division of Examinations: Investment Adviser MNPI Compliance Issues (April 26, 2022), Sections I–II and footnote 9.
- Google Trends: FAQ about Trends data and export, embed and cite Trends data.
- SEC: EDGAR APIs and Webmaster FAQ on filings and timestamps.
- Opportunity Insights: Economic Tracker, data documentation and data dictionary. Consumer-spending data supplied by Affinity Solutions. See Chetty, Friedman and Stepner, The Economic Impacts of COVID-19 (2024), Section II.
- Opportunity Insights: Consumer Spending revisions and file commit history; see the archived consumer-spending documentation for the July 25 vintage and the August 15 vintage. Exact archived CSV links are in the table.
- Hyndman and Athanasopoulos: Forecasting: Principles and Practice, Sections 5.2, seasonal naïve method, and 5.10, time-series cross-validation.
- Federal Trade Commission: InMarket Media Decision and Order, Parts II–III and VI, and finalization notice (May 2024).
- U.S. Supreme Court: Basic Inc. v. Levinson, 485 U.S. 224, 231–32; 17 CFR § 240.10b5-1.
- Investment Advisers Act Section 204A, 15 U.S.C. § 80b-4a, and 17 CFR § 275.204A-1, Investment adviser codes of ethics.
- SEC: In the Matter of App Annie Inc. and Bertrand Schmitt (September 14, 2021).
- U.S. Census Bureau: Advance Monthly Retail Trade Survey methodology, release schedule and release archive. Federal Reserve Bank of St. Louis: FRED versus ALFRED and release-date notes.
- Froot, Kang, Ozik and Sadka: What do measures of real-time corporate sales say about earnings surprises and post-announcement returns? (2017).
- Frazzini, Israel and Moskowitz: Trading Costs of Asset Pricing Anomalies (2012); NYSE 2025 trading calendar.
