AI in investment · Editorial

Port queues, parking lots, power draw — how do you read alternative data?

Alternative data promises an unfiltered view of the economy. The filter is the reader — this is the reader's field manual.

Alternative data entered investing through the side door: satellite images of parking lots, vessel-position feeds, job postings scraped at dawn, the hum of electricity meters. The pitch writes itself — see the economy before the statistics do. Artificial intelligence then multiplies the promise, because models are famously willing to learn from whatever they are fed. That generosity is exactly why the reading matters. A model does not ask whether the parking lot was photographed on a holiday.

This essay is the desk’s field manual for reading alternative data — not which feeds to buy, but how each kind of feed can quietly lie.

Port queues: from ships to headlines

A harbor is a camera-friendly machine. Vessels broadcast their positions; queue lengths outside a berth can be counted from space. From such counts, vendors build proxies for trade volumes, and from those proxies, others build investment signals. The chain is plausible, and each link has a soft joint: a queue can grow because trade is booming, because cranes are broken, because a typhoon parked the fleet for a week, or because a shipping line changed its calling schedule for reasons of bookkeeping. The count is real. The interpretation is a crowd of hypotheses wearing one number.

Counting is not understanding

Every alternative feed has a coverage — the slice of the world it actually sees — and coverage is where the bias lives. Satellite imagery forgets the night and the monsoon. Card-transaction panels cover only the people who carry that card. Web-scraped job postings cover only employers who write postings. None of these biases is disqualifying; all of them are load-bearing. The honest question is not is this accurate but accurate about whom — and Coverage consists of entities, not pixels: a port can rename its terminals, a company can merge and change its identifier, a product can be re-binned into a different category mid-year. If the vendor’s entity mapping cannot survive a merger, an appearance of continuity may be pure bookkeeping.

Who checks the counter?

The industry term is ground truth: an independent measurement against which the proxy can be compared. For a trade proxy, customs statistics exist after a lag; for footprint data, filings exist after a lag; for consumption panels, official retail series exist after a lag. The lag does not weaken the check — the proxy’s whole value is arriving earlier, so the check answers the different question of whether the early number was pointing anywhere near the late one. Vendors who never reconcile their feeds against any slower truth are asking readers to hold a map that has never met a road.

The one-number temptation

The most dangerous product of alternative data is a single clean number — “the congestion index” — because a number with a definite article acquires an authority its provenance never earned. The desk’s rule of thumb: the moment a proxy is described without its mechanism, someone is either skipping the epistemology or hiding the maintenance work. Both are worth knowing about.

Questions this desk brings to any feed

  • What exactly is being counted, with what sensor or source, and what misses the count?
  • When the counting method changed, was the history restated or spliced?
  • How are entities identified across time, and what happens through a merger?
  • Which slower, independent measurement would this feed have to agree with — and does it?
  • Who sees the raw stream under the license, and does the answer change what “independent” means here?

None of these questions is hostile. Good vendors enjoy them; the others were never selling data — only the posture of data.

An endnote on patience with proxies

Nothing here argues against alternative data; cheaper, faster, wider senses of the economy are a genuine gift to research. The argument is against reading proxies as verdicts. A feed is a witness, and this desk’s job — the only job, since nothing is sold or advised on this site — is to interview the witness before the machines quote it. If you know a feed that deserves an interview, the desk takes suggestions.