A fact of research life, uncomfortable to every buyer of signals and inconvenient to every seller: predictive strength fades. Not occasionally — as a rule. An investment signal that worked inside a backtest and then, carried politely into live use, slowly loses its grip on the future is not broken; it is maturing. Decay is so ordinary that the interesting question stops being does it happen and becomes which mechanism is happening now.

This essay walks the four wear patterns this desk looks for when covering an AI-derived investment signal, and the questions that surface erosion early. It is written for readers of research, in the belief that noticing decay is a craft, not an alarm.

Crowding: the audience becomes the mechanism

A signal works partly because the pattern it reads is under-read. Publish the pattern — or worse, let enough capital follow it — and the pattern starts to reflect its own audience. Everyone arrives at the same conclusion slightly earlier each quarter, the easy portion of the profit is consumed at the door, and what remains is mostly coordination risk. The tell is in the fundamentals of the signal’s own popularity: capacity figures, turnover growth, how much of the pattern is now common knowledge. A signal does not need to vanish to fail; it needs only to be fully priced.

Vendor drift: the pipe changes behind you

Machine-fed signals depend on data pipelines, and pipelines age like plumbing — invisibly. A vendor improves a cleaning step and splices the improvement into history, or edits coverage because a license changed, or quietly retires a source that no longer ships. Each such event changes what the signal is actually reading, without changing its name. The model keeps scoring; the world under the scores has shifted. Because drift lives in the data supply chain, the countermeasures are contractual and archival: revision policies in the license, restatement notices, and a habit of re-running the signal on genuinely old copies of the feed. A vendor unwilling to discuss its own pipe’s history is renting you yesterday’s map at today’s price.

Regime shifts: the ground moves, the model doesn’t notice

Many patterns are really conditions — ways markets behaved while interest rates, credit spreads or volatility stayed in a certain band. When the band breaks, the pattern does not exit loudly; it simply stops meaning anything. AI systems make this harder to see, not easier, because a well-trained model produces confident outputs regardless of regime, and because training data tends to be dominated by whatever period was most abundantly recorded. The honest disclosures here are ones this desk asks every source for: the regimes the training window covered, how the signal behaved at the edges of that window, and whether anyone defined what a “different world” would look like before the world provided one.

Revisions: yesterday’s number, today

Macro and corporate data arrive alive and are corrected for months afterward. A signal anchored to the corrected series performs against history that no live process ever saw — a flattering shadow of the fog in which real decisions were made. Decay enters when the shadow is what the model learned from: as corrections accumulate and re-releases stack, the live feed and the learned pattern drift apart grain by grain. The defense is a small, stubborn question — which vintage of each number was used — and the surprising discovery that many confident quant narratives were built on vintages nobody bothered to record.

The wear-inspection questions

Condensed, the questions this desk pins beside any aging signal:

How much capital visibly follows this pattern, and has that been re-checked?
What does the license say the vendor must disclose about restatements?
Which regimes does the training window actually cover — and which does it merely neighbor?
Are first-release values available for the series the signal leans on?
Has anyone re-run the signal on truly dated copies of its inputs, lately?

The questions do not stop decay. They replace surprise with a maintenance calendar — which is the difference between a signal that is aging knowingly and one that is failing privately.

An endnote on entropy, told plainly

Everything above is description, not counsel: this desk holds no view on what anyone should trade, sells no signals, and answers only to readers — inquiries about decaying datasets, sources worth covering, or disputes with this essay are all equally welcome at the inquiry desk in Taipei. Entropy is not a scandal; it is the ordinary weather of quantitative research, and reading the weather is what the articles here are for.

Why do AI investment signals decay?