Building an Internal Alpha Library

Every research desk eventually rediscovers the same idea twice. Someone tests a reversal signal in small caps, it doesn't clear the hurdle, they move on, and eighteen months later a different researcher tests the same idea under a different name, spends three weeks on it, and gets the same answer. The waste is not the second test — negative results are cheap to rerun. The waste is that nobody could check, in five minutes, whether it had already been tried.

What actually goes in it

An alpha library is not a folder of old backtest PDFs. It is a structured record, one row per signal, that answers four questions without anyone having to ask the original author:

• What is it, exactly? The formula or construction rule, in a form someone else could re-implement without guessing — not "value-ish score," but "book-to-price, sector-neutralised, winsorised at the 1st and 99th percentile."
• How was it tested, and what did it score? Universe, period, IC, decile spread, turnover, capacity, all from the desk's standard A Benchmarking Harness for New Signals so the numbers are comparable to every other row.
• What does it correlate with? Both other signals in the library and any factors already in the live book. A signal that scores well but correlates 0.7 with something already trading is worth a fraction of what the raw number suggests.
• What is its status? Live, tested-and-shelved, or tested-and-killed, with a one-line reason. "Killed: doesn't survive point-in-time restatement of estimates" saves the next person weeks.

Why the correlation column matters most

The single most valuable thing a library does is stop a desk from mistaking a new coat of paint for a new idea. A researcher builds a "supplier concentration" signal, runs it through the harness, and gets a respectable IC of 0.03. Without a library, that's the end of the story — ship it. With a library, a query shows it correlates 0.8 with an existing quality signal already live in the book. The 0.03 IC isn't new information; it's an old signal wearing a new name. That changes the decision from "add this" to "this doesn't earn its own capital allocation" — see Incremental Alpha of a New Signal.

The same lookup works in the other direction and saves time, not just capital: a researcher about to spend a week building a momentum variant can search the library first and find three prior attempts, two killed for the same reason (doesn't survive after costs in the relevant universe) and one shelved because it worked but had no clean economic story. That's a week saved before a line of code is written.

What kills a library in practice

Libraries fail for a boring reason: entry is expensive and optional, so people stop doing it under deadline pressure. The fix is not a policy memo, it's making entry a byproduct of running the benchmarking harness — if testing a signal automatically logs it, nobody has to remember to write it up afterward. A library that requires a separate, manual write-up step will have gaps exactly where they matter most: the messy, marginal, ambiguous results that don't feel worth documenting but are precisely the ones the next researcher needs to see.

The second failure mode is letting the library rot. A signal marked "live" that was quietly turned off eight months ago, or a correlation number computed against a book that has since changed shape, actively misleads. A library needs an owner and a refresh cadence — quarterly is typical — or it becomes a source of confident wrong answers rather than no answer.

Done well, the library compounds: the fiftieth signal tested on a desk is cheaper to evaluate than the fifth, because there's a growing map of what's already been tried, what killed it, and what it looks like next to everything else already trading.