A Signal Registry: Cataloguing What Exists

A research team of fifteen people, five years in, has built hundreds of signals — momentum variants, value ratios, sentiment scores — scattered across old notebooks, half-remembered project folders, and a few people's heads. A new hire spends their first month rebuilding a signal that already exists two folders over, because there was no way to find out it was already there. A signal registry solves exactly this: a single catalogue, one row per signal, that anyone on the team can search before starting new work.

What belongs in a row

A useful registry entry is more than a name. It needs the signal's definition precise enough that someone else could reproduce it, which data it depends on, who built it and when, its rough performance characteristics (does it work, in what universe, over what period), and its current status — live in production, tested and shelved, or actively being researched. Missing any of these turns the registry into a list of names nobody can actually act on: knowing a "value momentum" signal exists is useless if nobody can tell whether it worked, what data it needs, or whether someone already tried it and found it didn't survive transaction costs.

The registry's value depends entirely on people actually updating it, which is the part that quietly fails in practice. A registry that's accurate for the first six months and then drifts out of date as researchers move on to new things without updating their old entries becomes worse than no registry at all — it actively misleads a new hire who trusts a stale "status: shelved" label on a signal that was actually killed for a data bug that's since been fixed.

Worked example: what a search saves

A researcher wants to test whether earnings-day price drift predicts next-week returns. Without a registry, they build the signal from scratch — two weeks of work — only to discover in a team meeting that someone tested almost the identical idea eighteen months ago and found it disappeared after accounting for bid-ask spread on earnings days. With a registry, a five-minute search surfaces that prior entry, its data source, its exact result, and the reason it was shelved, letting the researcher either skip the idea entirely or start from the known failure point and try a genuinely different variant — turning two wasted weeks into five minutes plus a more informed starting point.

What this means in practice

A signal registry pays for itself the moment a team is large enough, or old enough, that no single person can remember everything that's already been tried — which for most desks is well within the first year or two. Keeping it useful is a process problem more than a technical one: the registry needs to be the natural byproduct of the research workflow (updated when a project wraps, not as a separate chore) rather than a side task everyone quietly skips.