Strategy Decay & When to Switch Off

Decay is a persistent, structural fall in what a strategy produces, relative to what its own validated history said to expect. The two halves of that sentence both matter.

'Persistent and structural' rules out a run of losses. Losing runs are not evidence of anything on their own — they are the ordinary texture of a probabilistic process, and a system that has never had one has almost certainly not been run long enough.

'Relative to its own validated history' is the part people skip. Without a stated expectation, every drawdown feels like decay, because a drawdown always feels worse from inside it than it looked in a table. The expectation has to exist on paper before the drawdown arrives, or you will be judging a system with the one instrument that is guaranteed to be biased: your mood while it is losing.

Four mechanisms, and they are unrelated to each other. Applying the response that suits one of them to a situation caused by another is the most common expensive mistake in this part of the work.

Once you know which of the four you are facing, what to do is nearly obvious. Getting to that knowledge is the whole difficulty, because the evidence available in the moment — a sequence of disappointing outcomes — is consistent with all four.

So the diagnosis cannot be made from the recent result. It has to be made against something fixed. That something is a band: the range of behaviour your validation work already said to expect, extracted and written down before a rupee was committed. Inside the band, a bad stretch is information you already had. Outside it, and with enough decisions behind it to mean anything, you have something new.

The band is not something you can construct after the drawdown starts, because by then you know the answer you want. It is an output of the walk-forward and out-of-sample work covered earlier in this module, recorded at the time.

Set a minimum sample, in advance, below which no verdict may be reached at all — not 'watch closely', not 'reduce a little', nothing. This is the single most useful sentence in the whole procedure, because it removes the first six weeks of live trading from the argument entirely.

There is no universal number. What you can do is make the number a consequence of your own study rather than a preference: if your validation windows disagreed a lot, you need more decisions before live results separate from noise; if they agreed closely, fewer.

Read the last row first. Most of the arguments people have with themselves about decay are settled by noticing they have not collected enough decisions to be having the argument.

Crowding is the one cause that gives you warning, and the warning arrives in the execution log rather than in the equity. That is worth knowing, because the execution log is the part nobody reads.

The signature is a widening gap between the price your study assumed and the price you actually got. Fills land further from the signal price. Partial fills become routine on sizes that used to complete in one go. The queue position you used to get at the open stops being available. On less liquid NSE names, the depth simply is not there at the moment your rule wants it, and your own order becomes a visible part of the move.

None of this shows up as a losing trade at first. It shows up as a slightly worse average, distributed across every decision, which is exactly the shape that hides inside ordinary variation until it has been accumulating for months. If you record the assumed price and the achieved price on every order — one extra column — the drift becomes visible far earlier than any drawdown threshold would catch it.

The instinct when live disappoints is to adjust — widen the stop, lengthen the lookback, add a filter that would have skipped the worst trades. It feels like maintenance. It is refitting.

What you are doing is selecting parameters using data that includes the recent losses. That is exactly the procedure that produces an overfitted study, except now it is being run on a smaller sample, under emotional pressure, without a held-out set, and without any of the discipline you applied during research. Every one of the protections built earlier in this module has been suspended precisely when the temptation to suspend them is strongest.

There is a further problem: once you retune, you no longer know what you are running. The live record before the change and after it are records of two different strategies, so your sample resets to zero at the worst possible moment. A system that gets retuned every time it underperforms has never been allowed to be wrong for long enough to be evaluated at all.

No, and the distinction is worth holding clearly because the two get merged constantly.

The AI module's treatment of model drift is about a dependency you rent: a hosted model is changed by its provider, so the same prompt meets a different system and the output quietly shifts. You did not change anything; the tool moved underneath you.

This topic is the other case entirely. Your model has not moved at all. The parameters are exactly the ones you fitted, the code is byte-identical, and the market it was fitted to is the thing that changed. Nothing in your stack is faulty. The diagnostic question is different too: for a hosted-model change you re-run a fixed regression prompt against a frozen input and compare; for strategy decay you compare live behaviour against a band your own validation produced. Same symptom of 'it got worse without erroring', two unrelated causes, two unrelated tests.

Write the ladder before you need it. Each rung pairs an observed condition with a response you have committed to, so that the decision under pressure is a lookup rather than a judgement.

Monitoring fails for boring reasons. It gets proposed as a daily discipline, survives a fortnight, and then quietly stops — which is the same failure mode as the thing it was meant to catch.

So separate the two tiers. The scheduled tier runs whether or not anything seems wrong, because nothing will seem wrong until it is late. The triggered tier runs on events known to precede a break. Neither replaces the other: scheduled checks catch slow erosion, triggered checks catch discrete breaks.

Anchor the scheduled review to something that already happens rather than to a date you have to remember — the first weekend of the month, or the session after expiry week ends. And keep a log entry for the reviews where nothing changed, because a run of unremarkable entries is precisely what makes an eventual change legible, and lets you bracket when it began.

Both failure modes are emotional, and they are opposite.

Holding on too long is sunk cost dressed as conviction. You built it, you validated it, you told someone about it. Switching it off feels like admitting the months were wasted, so you find reasons the current period is unrepresentative — and every drawdown supplies those reasons for free.

Switching off too early is capitulation timed by discomfort. The pain of a drawdown peaks near its worst point, which is exactly where the pre-committed ladder is hardest to follow and exactly where abandoning is most expensive. If the decision is being made because the loss feels intolerable, the real problem is position size, not the strategy — and that problem does not get solved by switching anything off.

The defence against both is the same, and it is structural rather than emotional: the threshold, the minimum sample and the response were all fixed while nothing was at stake. Under pressure you are not deciding. You are executing a decision your calmer self already made.

This procedure has limits, and they should be stated as plainly as the procedure itself.