2. How you deceive yourself when testing an idea
Level 12: A process that survives
Suppose you already have a well-written hypothesis, with its reason, its counterparty and its refutation condition. Now it has to be tested, and here begins the part of the process where the most money is lost, because the adversary stops being the market and starts being you. No bad faith is required. The four mechanisms that follow work perfectly well in honest, intelligent, well-intentioned people, and they work without your noticing, which is precisely what makes them dangerous.
The four deceptions
- Testing many variants and keeping the one that comes out well. If you test forty versions of an idea on the same data, the best of the forty is going to look excellent even if none of them is worth anything. What you have found is not a rule, it is the biggest stroke of luck available in that sample. And the damage does not depend on your intention: it makes no difference whether you tested forty at once or one a week for a year, because what inflates the result is the number of attempts, not the calendar.
- Looking at the data before deciding the rule. It is the most silent deception. You observe that a 20% threshold works better than a 15% one, you adjust, and from that moment on your test no longer tests anything: you have designed the rule with the answer in front of you. The idea and the test stop being independent, and without independence there is no information.
- Counting the hits and forgetting the misses. Memory does not file things away evenly. You remember the trade that went well because you told people about it, and you do not remember the four times you considered something similar and never got round to doing anything. The result is a mental track record with an enormous bias in your favor, built without a single lie.
- Rationalizing after the fact. Once you know the outcome, a coherent explanation always turns up. If you won, it was the analysis. If you lost, it was an exceptional event that could not have been anticipated. Neither sentence can be checked, and both feel true.
Why memory does not work as a record
Here it is worth being precise: the problem is not that one has a bad memory. The problem is that memory rewrites. When you recall a decision, you do not retrieve a file: you reconstruct the scene out of what you know now, including the result. And since the result is already known, the reconstruction tends to make your reasoning at the time look better informed, more nuanced and more prudent than it was. Nobody remembers having been certain about something that went wrong; what they remember is having had doubts.
Annie Duke, who was a professional poker player before devoting herself to decision-making, insists on separating two things that ordinary language mixes together: the quality of a decision and the quality of its result. In any environment with chance in it, a good decision can end badly and a bad decision can end wonderfully. Judging by the result is comfortable, immediate and systematically misleading: it teaches you to repeat what went well, which is not the same as what was correct.
Philip Tetlock found the other side of the same problem while scoring specialists' predictions over a number of years. Average accuracy was mediocre, but what is relevant for this module is how it was corrected: the ones who improved steadily were those who recorded their forecasts with a probability and a date, reviewed them afterwards against what had happened and treated every error as information rather than as bad luck. In the large-scale forecasting tournaments he ran, that habit was enough to separate the best from the rest clearly and stably. It was not some inaccessible talent: it was keeping score.
The only thing that corrects this: writing beforehand
There is no psychological shortcut. You do not get out of these four deceptions by being cleverer, or humbler, or by paying more attention. You get out by creating a record that cannot be modified after the result, because what gets corrupted is not your judgment, it is your memory.
A useful note is made before you know anything and contains, as a minimum:
- The date and the exact decision, written in such a way that a stranger could carry it out identically.
- The reason, in a sentence beginning "this works because".
- The counterparty: who is selling and what constraint they are under, as per the previous module.
- The expected magnitude and the period within which you ought to see it.
- What would make you abandon the idea, in observable terms.
- How many variants you tested before settling on this one. This line is the most uncomfortable one to write, and that is why it is the most informative.
- Your degree of confidence, as a number. Not in order to be right, but so that you can check later whether your confidence was calibrated.
And a rule that looks trivial and is not: you note down everything, including the ideas you discard and the trades you think about and do not make. A record containing only what you executed reproduces the third deception exactly, in notebook format.
An example, declared as such
Imagine two notes about the same trade. The first, written afterwards: "I bought because the sector was cheap and the market was exaggerating; it worked". The second, written beforehand: "I am buying on March 14 because funds with a ratings mandate are being forced to sell; I expect to recover between 8% and 12% in six months; I abandon it if a buyer of size appears with no price constraint; I tested three threshold variants and I am keeping the middle one; confidence, 60%".
Only the second allows you to learn anything. If it goes well and the unconstrained buyer never appeared, you have confirmed the mechanism. If it goes well because one appeared and you made money anyway, you know you were right for a reason other than the one you wrote down, which is a winning trade and a refuted hypothesis at the same time. The first note, by contrast, can be applied to any result, including the opposite one. That is why it is so comforting to write, and that is why it teaches nothing.