Not broadly true in my opinion. To give an example from the financial services industry:

* Satellite provider or analytics firm sells "car counts" for retail stores

* Financial institution is intrigued; this must correlate with sales, right?

* But..why don't we just buy credit card transaction data and foot traffic data from clearing houses and GPS trace providers?

I often liken satellite imagery to salt. It's great to finish a dish, but should never be consumed alone. If you don't believe the foot traffic data or credit card transaction data you're buying, you can use satellite data to check it or refine the model. But that's a niche within a niche.

The other issue I pointed out in that article is customer savvy--if you're a quant fund sophisticated enough to make use of an arcane data feed, you're very likely sophisticated enough to generate that feed yourself from raw data. And if you do it yourself, it's suddenly part of a "proprietary" solution. So you'd rather just buy images and do the heavy lifting that pay a premium to buy a data feed that doesn't quite solve your problem by itself.

This was our experience with car count data which we sold to most of the sophisticated quant funds a few years back. We sold 'raw counts' and our own 'derived counts' that took into account cloud cover etc and was produced by our data science team. The quant funds completely disregarded our derived counts as useless.

Very interesting, thank you for sharing that!

Having been on the other side of that... quant funds absolutely want raw data wherever possible, but I suspect for the application you're talking about there's some room for derived data; the responses for e.g. credit card data are relatively infrequent (earnings are reported 1/quarter) so using high-dimensional, high frequency data can be more trouble than it's worth. Genscape is a good example of a company with derived data that everyone in the industry subscribes to.

I think this applies to a lot of types of data. Selling "insights" is really just selling a filtered version of the raw data. If the client's needs match up with your filters, then you've saved them a lot of work. However, if the filters don't match up perfectly with the client's needs then they'll need to do filtering and processing themselves. And in that case, they'll likely get better results just working with the raw data, instead of whatever data you think they need.

I suspect it is at least approximately true of all data sets that are both broadly applicable and difficult to acquire (i.e. requires domain expertise to acquire).