Why is a failing model removed rather than kept around as a warning or a maybe? This note explains the removal, using a mean-reversion rule on SPY that cannot survive a different regime as the public example.
A failed model still occupies a path
A model that has failed its tests and remains wired into a notebook, a configuration, or a person’s habits is still a model in use. People will reach for it when it flatters the day. The failure will be remembered as nuanced. Removal is the practice of making the failure operational: the model is marked ineligible, the path that would call it does not call it, and a later revival has to start a new review rather than resume an old enthusiasm. Keeping it close, in case it was early, is how ineligible work stays eligible.
Removal is not humiliation and not a deletion of history. The history is required so the same rule is not reintroduced under a new name next quarter. What is removed is the authority to proceed. A museum of failed ideas, clearly labeled, is compatible with that. A folder of maybe-later ideas, unlabeled, is not. The label is the culture. Without it, every author is a lobby for a revival.
Removing a failing model means marking it ineligible and taking it off every path that could still call it. The history is kept. The authority to proceed is not.
A rule that only liked the calm
Take a simple mean-reversion rule on SPY that looked tidy on a calm window and then met a window in which the fund trended because the index was repricing news. Suppose the stated tests included that second window, and the rule only remained attractive if those days were dropped. The rule has failed. The cultural question is what happens on the afternoon of the failure. If the response is to shorten the window until the picture returns, the tests were decorative. If the response is to mark the rule ineligible and write down why, the tests were real.
The SPY example is public and intentionally plain. Exotic models are not the only ones that deserve removal, and they are not the ones most likely to be kept out of affection. A simple rule that several people have discussed is harder to remove, because removal feels like a comment on the discussion. The culture has to prefer that awkwardness to a pile of rules that only work in the brochure window. Regime tests exist so this moment arrives on a schedule rather than in a surprise. When it arrives, removal is the default, not the escalation.
There is a practical detail. Removal includes the copies. A notebook that still imports the failed rule will be the copy someone runs. A reviewed module that still exposes it will be the copy someone wires. The person who wrote the rule is often the worst person to find every copy, which is why the written review names the locations. A failing model is removed when those locations no longer offer it as a living choice. A paragraph in a postmortem is not enough if the import still succeeds.
Revival is a new object
Nothing prevents a later person from proposing a different rule that learns from the failure. That proposal is new. It does not inherit the old rule’s anecdotes, and it does not skip the tests by citing the earlier work as almost. The earlier work is evidence of a dead end, which is useful, and evidence of nothing else. Treating removal as temporary by default empties the word.
Why a failing model is removed is so that ineligible work cannot be used by habit. The mean-reversion sketch on SPY is the example of a failure that should be boring to execute. The firm’s culture is measured by whether the boring thing happens.
This piece does not identify a live model, claim that mean reversion has failed in general, or describe a removal that produced a profit. It explains why a rule that misses its own tests, illustrated with a SPY sketch, is taken out of use. Leaving the import in place and calling the rule retired is not removal.
