# From alternative data to alpha signals
A parking lot counted from space told a quiet story before the earnings release. In the mid-2010s, hedge funds began buying satellite imagery of retailer parking lots, counting cars week over week, and inferring foot traffic ahead of quarterly sales. The desks that did this well got a read on revenue before the market did. That is the promise of alternative data: information that is not in the standard financial filings, turned into a tradable edge.
This lesson walks through how a long/short equity desk (a fund that buys stocks it expects to rise and sells short stocks it expects to fall) converts messy raw data into a signal: a numeric score that ranks stocks by expected return. The hard part is not the data. It is avoiding the traps that make a signal look brilliant in backtest and worthless live.
Alternative data is any dataset outside traditional sources (financial statements, analyst estimates, price and volume). Common categories:
The FactSet and Eagle Alpha ecosystems catalog thousands of such vendors. For a solid primer on the category, see the CFA Institute overview of alternative data.
A feature is a single measurable input, for example "year over year growth in card spend for company X this week." Three quick examples of how each raw source becomes a feature.
You never see individual transactions. The vendor delivers aggregated spend by merchant. You mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → merchants to public tickers (a nontrivial step: "SBUX" covers thousands of store brands and franchisees), then compute weekly year over year growth. The feature: card-implied revenue growth versus the same quarter last year.
A computer vision model counts cars in a retailer's lots across hundreds of locations. You normalize by location and season (December looks nothing like February), then aggregate to a company-level index. The feature: change in traffic index versus trailing average.
You pull the transcript, split management remarks from the analyst Q&A, and score tone. A simple approach uses a finance-specific sentiment dictionary. A modern approach uses a language model to score uncertainty, hedging, and topic shifts.
# Simplified: sentiment delta on an earnings-call transcript
from transformers import pipeline
clf = pipeline("sentiment-analysis", model="ProsusAI/finbert")
remarks = [
"Demand accelerated across all regions this quarter.",
"We are seeing some softness we cannot yet fully quantify.",
]
scores = clf(remarks)
# Feature = mean(positive) - mean(negative), then compare vs the
# same company's prior-call baseline, NOT vs an absolute threshold.The comparison to the company's own prior baseline matters. Some management teams are chronically upbeat. A signal built on absolute tone just ranks personalities.
This is where most alt-data projects quietly fail. A backtest (simulating a strategy on historical data) can look spectacular and still be an illusion.
Lookahead bias means using information in your backtest that you would not have had at that moment in real life. It is the most common and most fatal error.
Concrete version: a data vendor "stamps" a card-spend record with the transaction date. But that record was not delivered to subscribers until two weeks later. If your backtest acts on the transaction date, you are trading on data from the future.
The fix: use point-in-time data, which records what was known as of each date, including the delivery lag. Always ask a vendor: "What is the delivery latency, and can you provide point-in-time snapshots?" If they cannot, assume the worst case lag.
Same trap with fundamentals: a company may restate earnings months later. Backtesting on the restated number leaks the future.
Data-snooping happens when you test so many variations that one looks great by pure luck. Test 200 signal formulas and a few will show strong historical returns even if all are random noise.
The fix:
If your card panel only covers companies that still exist and skips those that went bankrupt, your backtest inherits a rosy sample. Same if the satellite vendor added coverage of a hot retailer only after it became popular. Always check when each name entered the dataset.
🎬 [VIDEO: "Backtesting and the Dangers of Overfitting" — youtube.com — a clear walkthrough of how multiple testing inflates backtest results and how to guard against it]
Once features are clean and point-in-time, the workflow is disciplined.
Step 1: Combine features into a score. Rank stocks each week by, say, card-implied growth plus satellite traffic change plus call-tone delta. Rank-based scores are more robust than raw values because they resist outliers.
Step 2: Neutralize the obvious. A raw signal often just bets on sectors or size. If all your top-ranked names are small-cap retailers, you have a sector bet, not an alt-data edge. Desks neutralize exposures by ranking within sector and adjusting for market cap, so the remaining signal is stock-specific.
Step 3: Simulate realistically. Include transaction costs, borrow costs for shorts, and turnover. A signal that must trade daily can be eaten alive by costs. Model the market impact of your own trades.
Step 4: Measure decay. Alt-data edges erode as more funds buy the same data. Test whether the signal weakened in recent years. The parking lot trade is far more crowded now than a decade ago.
Step 5: Out-of-sample and paper trading. Run the finalized signal on held-out data, then paper trade it live before committing capital.
Knowledge check
1. According to the lesson, what is the fundamental challenge in converting alternative data into a profitable alpha signal?
2. What best captures the definition of a 'signal' as used in this lesson?
3. A satellite counting cars in retailer parking lots to infer foot traffic ahead of earnings is valuable primarily because it provides information that is:
4. Select ALL correct answers. Which of the following would be classified as alternative data according to the lesson?
Select all the correct answers.
5. Select ALL correct answers about the relationship between raw data, features, and signals in this lesson's framework.
Select all the correct answers.
Alt-data is not just a quant problem. It carries legal and reputational risk.
Material nonpublic information (MNPI). Regulators prohibit trading on MNPI (important information not available to the public). If a card vendor's panel is so granular it effectively reveals one company's internal sales figures, using it may cross a line. Reputable vendors aggregate and anonymize to stay clear of this. Every dataset should pass legal review.
Personal data. Card and location panels can implicate privacy rules. Data must be aggregated so individuals cannot be identified. Provenance matters: how was consent obtained, and was scraping done within a site's terms.
Cost versus capacity. Alt-data is expensive, and a single dataset can cost six figures annually. A signal only justifies that cost if it works at the fund's asset size. A brilliant signal that only works on a few million dollars of positions is irrelevant to a large fund.
The SEC's guidance and enforcement history on alternative data and expert networks is worth tracking; the SEC newsroom publishes actions that shape what desks consider acceptable.
Most individual alt-data signals are weak. A single signal might have a low information coefficient (a correlation between predicted and actual returns often in the low single digits as a percentage). The value comes from combining many weak, uncorrelated signals into a portfolio, and from disciplined execution. There is no single dataset that prints money, and any desk promising that is selling a story.