Why does a signal die even though the code that computes it has not changed? This note lists the ordinary causes, using a simple momentum rule on the Russell 2000 as the public example, and separates a dead signal from a broken program.
The code can be innocent
A signal dies when the relationship it was measuring no longer shows up in the way the author hoped, or when the costs of acting on it consume the effect, or when the effect was never more than a pattern in one window. The program can still emit yesterday's function of yesterday's inputs. Mathematical obedience is not survival. The first diagnostic question is whether the code still matches its specification. If it does, the death is not an outage. It is information about the rule or about the period.
Crowding is one ordinary cause. A rule that many people can compute from public data can change the prices it is trying to harvest, because the trades that implement it move those prices. Decay is another name for the same family of facts: the opportunity, if it was real, was visible, and visible opportunities attract the activity that shrinks them. A third cause is a change in the mix of companies, the way an index is built, or the cost of trading the names. The signal did not break. The world in which it was computed moved.
A signal dies when the relationship, the costs, or the sample no longer support it, even though the program still computes the same function of its inputs.
Momentum on a public small-cap index
The Russell 2000 is a public index of smaller U.S. companies. A simple momentum rule ranks names by a recent trailing move and prefers the ones that already rose. On some multi-year windows that rule has been discussed to exhaustion in public research, which is exactly why it is the example. A rule everyone can write down is a rule whose death a reader can imagine without a private dataset. The code that ranks last year's winners can keep running forever. The economic content of acting on the rank is a different question, and it can fade while the code stays perfect.
The check is to split causes before inventing a new formula. First, replay the implementation so the ranks match a recorded run. Second, ask whether the index membership used in the rank is point-in-time, because a rule that accidentally uses later membership will look healthier than the rule a person could have traded, and then look sick when the leak is fixed. That sickness is a repair, not a death. Third, look at costs and at how crowded the rebalance is. A momentum rank that looks large before costs and small after them did not die in the ranking step. It died in the spread, the impact, and the crowd at the rebalance.
There is also a statistical death. A rule tried in many variations will have a best variation. That best one can stop looking special as soon as the sample moves forward, because it was special to the sample. The Russell example is public enough that a reader should assume the variation was not the first one anyone tried. A signal that dies out of sample may be telling the truth about the search. Replacing it immediately with a cousin that flatters the new sample repeats the cause instead of reading it.
Death is an allowed outcome
A process that cannot retire a signal will accumulate rules that only worked in a brochure. Retirement means the rule is no longer eligible for the next stage. It does not mean the code is deleted from history. The history is how the next person avoids rerunning the same hope. The Russell momentum sketch is a teaching object. Its death, when it dies, should be classified: implementation, look-ahead, cost, crowding, or a search that fitted noise.
Why a signal dies is therefore not a mystery to paper over with a new parameter. It is a short list of ordinary causes. The code is the first thing to clear and the last thing to blame. A rule that everyone can compute on a public index is the right classroom, because the classroom does not depend on a secret.
This piece does not claim that momentum is dead or alive, that the Russell 2000 is a portfolio, or that a retired signal was once a source of profit. It lists why a rule can stop deserving attention while the code remains correct.
