Every few months a new “alternative data” source gets pitched as the next edge — satellite imagery, credit-card panels, app-install data, geolocation. Most of it is correlated noise dressed up as alpha. A small fraction is durable signal.

The question worth asking, before any vendor demo: what does this data tell me that I can’t get from the public version, and how long does the gap last?

A simple test

Before paying for a new dataset, I run it through three filters:

  1. Lead time. Does this dataset move before the data the rest of the market is looking at? If yes, by how many days?
  2. Decay. Once enough people are using it, is the edge gone? Most alternative data is a decaying-half-life trade.
  3. Cost of being wrong. What’s the false-positive rate look like in regimes that aren’t the back-test window?

If the answer to (1) is “a few hours” and (2) is “as soon as the second-largest hedge fund subscribes,” you’re looking at a feature, not a moat.

Where alternative data actually pays

The pattern that consistently works isn’t a single source — it’s the combination of two or three weak signals into something with a higher information ratio than any of them alone. That’s harder for a competitor to replicate, because the value isn’t in the dataset; it’s in the integration.

The boring conclusion: most alternative data is overpriced, the integration work is undervalued, and the people quietly building proprietary feature sets out of public data are eating the lunch of teams who think a vendor subscription is a strategy.