# What is alternative data?


Alternative data is information about a company, collected from non-traditional sources. Things like job postings, product prices, app rankings, store foot traffic, receipts, and more.

Portfolio managers and quantitative analysts at hedge funds and investment banks use it for an earlier, finer read on revenue, market share, and company activity. HR teams at large employers use it to benchmark pay, hiring, and attrition.

But this is where most definitions stop. A buyer needs one more answer: will the series hold up under a backtest?

## The standard definition

Traditional data is what a company publishes about itself, or what an exchange, regulator, or statistical agency publishes about it. At the core are fundamentals such as filings and financial statements, plus official payroll statistics and government surveys. The company-reported numbers are standardized, audited, and late: quarterly at best, often months after the quarter they describe.

Alternative data comes from ordinary activity. Operations, transactions, and observation generate it, outside the company's reporting process. So it arrives earlier than the report, at the level of a single posting, product, or store.

It also arrives unstandardized. Nothing at the source enforces a schema, so any normalization is added afterward, by the vendor or by you.

Purchased datasets tend to fall into six families:

| What it measures | Typical sources | Who uses it |
|---|---|---|
| Hiring and workforce | Job postings, employer reviews, advertised salaries, headcount by location | Hedge funds screening for expansion or distress; HR teams benchmarking pay and attrition |
| Products and prices | Listed prices, discount depth, assortment, stock status | Funds tracking demand and pricing power; investment banks mapping a sector |
| Attention and adoption | App rankings, downloads, ratings, web traffic, social followers and engagement | Funds monitoring consumer traction before it reaches a filing |
| Spend and transactions | Card transactions, point-of-sale records, emailed receipts | Funds nowcasting revenue and share of wallet |
| Physical activity | Satellite imagery, foot traffic, geolocation, store locations | Funds counting cars, construction, or store visits; investment banks sizing a market; corporates looking to expand |
| Legal and regulatory | Court dockets, patent and trademark filings, assignments, permits, regulatory approvals | Funds pricing litigation and IP risk; banks mapping patent portfolios; corporates watching rivals' filings |

## What it is not

- **Not a synonym for big data.** Size describes the pipeline, not the source. A small, well-maintained series from one regulator is traditional data. A billion rows of the same series is still traditional data.
- **Not necessarily scraped.** Card transactions, satellite imagery, and emailed receipts are alternative data, and they have nothing to do with a web crawler. Scraping is one collection method among several.
- **Not a signal by itself.** A public job posting is alternative relative to a payroll report. It is not secret, and it is not a signal on its own.

## Redefining alternative data

The six families all measure activity. Activity is only useful when it is tied to a company, and the value is what the source says.

For a valid backtest, the definition needs two additions. Identifiers must be mapped and versioned, and extraction must be true to source.

**Entity mappings are a living graph.** Each row must resolve to a company through a moving graph of registrations, tickers, ISINs, and subsidiary trees. A rename or acquisition can break continuity. The mapping that found a company yesterday should find the row today.

- A job posting is only useful once it resolves to a company you can trade.
- An empty parking lot is just a photo until you link it to a specific store.
- A social media handle is only useful once you know who owns it.

**Reproducible, deterministic extraction.** Each row's value is what the source says, not what an AI model inferred or interpreted. The same source returns the same value on an identical run. In the case of derived values such as for sentiment analysis, the model or algorithm used is versioned and the value can be reproduced at a later date.

- A full parking lot is not attributed to one store while the neighboring stores are filling it.
- A product record doesn't show a $49.99 list price when the page listed 49.999,00 €.
- A job posting for a Locomotive Engineer is not grouped under Engineering.

Without these two additions, failures arrive quietly in the backtest. Signal is obscured, and the value of the data is lost.

## Five questions to ask a vendor

1. **What is the collection date of the historical data?** Not the earliest date it claims to cover. The dates the rows were actually observed. The older rows were either collected then or rebuilt later.

    *Many types of alternative data are a snapshot of the world at a specific time. You cannot factually rebuild an observation from its artifacts. So a vendor filling historical gaps from later snapshots loses accuracy. That is look-ahead bias, and it arrives through the vendor's archive rather than your code: every row carries information from after the decision you are testing.*

2. **Do you version the entity mappings?** When a company's ticker or ownership changes, the vendor should be able to show the mapping as it stood on a past date.

    *Imagine a row's link to a company changes. Without a dated mapping, you cannot read the row as of its record date. It orphans, or it resolves through today's structure. You are backtesting a history that did not exist when the rows were collected.*

3. **What happens to an acquired or delisted company?** Its history either stays with the old entity, follows it into the acquirer, or vanishes when the ticker stops trading. If it vanishes, the extract carries survivorship bias.

    *Drop the delisted names and the extract becomes a universe of survivors. The backtest never owns a loser, so every strategy looks better than the one you would have run. Keep an acquisition's rows under the acquirer and the acquirer looks stronger than it would have.*

4. **Which fields are extracted or inferred using AI?** A large language model (LLM) in the process can fill a field with something plausible and wrong. Every value of importance should trace back to the source through code instead of a black box.

    *If a vendor interprets or extracts the data with an LLM, it gets what the field probably says, not what it actually says. Blanks get filled instead of left empty, and the same source can yield a different value on an identical run. In our experience, involving AI (even frontier models) in the extraction process lowers the quality of the data significantly in cases where other methods were available.*

5. **Which rows match an identifier in my universe?** A percentage rate answers a different question. A partial sample file hides which rows failed and how.

    *A join rate hides two different failures. Some rows find no company, and some find the wrong one. A vendor can raise the first number by getting loose with the second, so ask to see the matches, not the percentage.*

Ask for the answers in writing, before the trial.

A vendor that does not know or cannot answer all five correctly has already told you how the backtest will end. Skip it and move to the next vendor.

## How to use the definition

Use the standard definition to narrow the vendor list. It tells you what a vendor measures.

For a backtest, use the narrower definition. Alternative data means a series you can join to an entity, rebuild as of a decision date, and reproduce from the same source.

That won't tell you whether the data carries signal. But it will keep you from wasting time and money on data that never could.

Our company, [Octofact](https://octofact.com/), collects alternative data on that narrower definition. All of our series are continuous, point-in-time, and reproducible. See our datasets at [https://octofact.com/datasets](https://octofact.com/datasets).