3. When to abandon something that stops working

Level 12: A process that survives

This is the hardest problem in the course and it is best dealt with plainly: a bad run and an edge that has disappeared look exactly the same from the inside. There is no indicator that separates them in real time, nor any loss threshold that tells one from the other, and anyone who says otherwise is selling something. All you can do is build criteria that do not depend on waiting for the result to speak.

Why the noise takes so long to clear

Put some numbers on the table. Imagine a genuinely good strategy, with a ratio between return and variability of 0.5: it makes, on average, half of what it swings. That is an excellent result, of the kind very few managers sustain for decades. With that figure, the probability that any given year ends in a loss is around 30%. The probability that one specific three-year period ends entirely below zero is far lower, roughly one in thirty-four. But that is not the useful question, because you do not live one specific three-year period: you live a succession of them. Over a decade there are eight three-year periods chained together, and the probability of running into a spell of three bad years at some point in those ten years rises to around one in six. That is to say: what from the inside looks like an anomaly demanding an explanation is, seen from outside, what happens to one in six. It is not a rare case: it is what is to be expected if you wait long enough.

Now the question that matters. How many years of results are needed to distinguish, with reasonable confidence, that strategy from another that contributes absolutely nothing? With a ratio of 0.5, on the order of sixteen years. If the real edge were 0.3, more modest but still very valuable, around forty-five years would be needed. And if it were 0.25, more than sixty.

Read that again. An entire professional career may not be enough to demonstrate that a moderate edge existed. Nobody has that patience, and not out of weakness of character: clients leave sooner, mandates are reviewed sooner and the investor himself has to eat. This means that the decision to abandon is never going to be taken with sufficient evidence. The question is not how to obtain certainty, but how to decide well without it.

The two errors, and why they are not symmetrical

There are two ways of getting it wrong. Abandoning something that was still working, right after the worst run and right before the recovery, which is the most repeated pattern in the history of investing. And holding on to something that is already dead, paying costs for years to finance somebody else's discipline.

Both hurt, but not equally: the first is a visible opportunity cost and the second a slow haemorrhage that also consumes attention and credibility. Neither is resolved by looking at the results curve, because the curve is precisely the piece of data that does not discriminate.

Criteria that do not depend on the recent result

If the result does not inform you, you have to judge the mechanism. Three questions, in this order:

• Has the reason it worked changed? Go back to the written hypothesis from the first module and look at the counterparty line. If your edge existed because a type of participant was obliged to sell, check whether that obligation still exists. A change in regulation, in mandates or in accounting can eliminate your edge without any return figure announcing it.
• Has it become crowded? This is the phenomenon from level 4. The signs are observable without waiting for the result: the idea appears in the general press, there are products that package it up, the money devoted to it has grown several times over, and the spreads you were exploiting have narrowed. If the edge is still there but only a tenth of the money fits that used to fit, your problem is not one of refutation but of capacity.
• Has the structure of the market changed? Level 5 left the clearest example: who buys, on what criterion and in what proportion. Changes in how trading is done, in the composition of the index, in the trading hours or in the relative weight of the different participants can alter the ground under your feet. The signal may still be true and the price at which you manage to execute it may no longer leave a margin.

Michael Mauboussin offers the tool that puts all this in order: separating the role of skill from that of chance according to the environment. Where chance weighs heavily, the short-term result contains almost no information and the process has to be judged instead. Howard Marks adds the practical complement: you cannot know where the market is going, but you can know what point you are at, and what point the idea you were exploiting is at.

The rule that has to be written beforehand

Since the decision is going to be taken with insufficient data, how it is taken has to be fixed in advance, and not when it hurts:

• A minimum holding period, consistent with the horizon declared in the hypothesis. If you said six months, six months.
• A loss limit which is not a criterion of being right but of survival: it is not there to tell you that you were wrong, but to make sure you are still in the game.
• A periodic review of the mechanism, on a fixed date, in which you answer the three questions above in writing and without looking at the recent return.
• An exit by refutation, the condition from the first module, which is the only one that allows you to abandon without having waited sixteen years.

Real case: value's years in the wilderness

Between 2007 and 2020, the strategy of buying the cheapest shares by multiples and avoiding the most expensive went through one of the worst periods on record since data has existed. It was approximately thirteen years of results worse than the market as a whole. Thirteen years is longer than almost any professional can bear without explanations, and quite a few well-known managers closed down or changed approach during that stretch.

Over those years two narratives coexisted, both reasonable and both unprovable at the time. One said that it was a very long but statistically normal run. The other said that the reason had died: accounting measures companies whose value lies in intangibles worse and worse, so the very numerator of the comparison had become misleading. The second was not an excuse, it was exactly the right question: has the reason it worked changed.

From the end of 2020 onwards the spread narrowed sharply over several quarters. Whoever held on out of conviction and whoever held on because they had reviewed the mechanism obtained the same result, and that coincidence is a trap: one of the two had a process and the other got lucky, and not even the result allows you to tell them apart.