Hindsight bias
A hire doesn't work out, and the interview you were pleased with at the time reorganises itself into a list of warnings you apparently ignored. A company fails and the reasons are all over the coverage, none of which anyone wrote down while it was still trading. The feeling in both cases is the same, and it's specific enough to have a name: once you know how something turned out, the outcome stops looking like one of several things that could have happened and starts looking like the only one that was ever going to.
Three things, not one
Hindsight bias is usually described as knowing it all along, which is the third of three effects that behave differently. The psychologists Neal Roese and Kathleen Vohs, reviewing five decades of the research in 2012, separate them:
Memory distortion. You misremember your own earlier estimate, and the misremembered one sits closer to what happened. You said 30% and recall having said 60%.
Inevitability. The event comes to seem like it had to unfold that way, whatever anyone's estimate had been.
Foreseeability. You believe you personally saw it coming, which is the layer that does the damage to learning.
They stack, and they have different causes, which is why an intervention that dents one of them can leave the others intact.
Why it happens
None of this is dishonesty, and it isn't a weak memory either.
Your earlier judgment isn't filed away somewhere to be retrieved on request; it gets reconstructed each time you go looking, out of what you currently believe about the situation. Once the outcome is part of what you believe, it's part of the raw material the reconstruction is built from, and the result arrives feeling like recall rather than inference. Baruch Fischhoff, the psychologist who first isolated the effect, gave subjects a historical account of the 1814 war between the British and the Gurkhas and told different groups different outcomes. Each group rated its assigned outcome as the more probable one, and each denied that being told had influenced them.
The denial is what makes it hard to work around: the bias produces no sensation of its own, and the revised memory feels exactly like the original one.
What it costs
Learning from your own record stops working. If you evaluate past calls from memory, you're grading a version of your reasoning that has already been edited toward the result, so the calls you got wrong look like near misses and the ones you got right look like insight. A track record assembled this way can only tell you that you're a reasonable judge of things, which is what it would have told you regardless.
Judging other people gets unfair. Kim Kamin and Jeffrey Rachlinski, studying how people assign liability in hindsight, asked one group whether a town should take an expensive flood precaution and another group, told the flood had happened, whether failing to take it had been negligent. Far more of the second group condemned the decision than the first group had thought the precaution worth taking on the same facts. Performance reviews, incident post-mortems and malpractice suits all run this design by default.
Institutions learn the wrong lesson after a failure. The warning signs are always everywhere afterwards, because a failure tells you which of the thousand ambient signals to go and collect. Post-mortems that start from the outcome reliably produce a story in which somebody should have known.
Warning people doesn't fix it
Telling people about hindsight bias, including immediately before testing them on it, does very little. Neither does instructing them to be careful, or offering them money to be accurate.
The one debiasing move with reasonable support is to consider the opposite: write out, concretely, how the other outcome could have happened, and what would have caused it. Roese and Vohs report this as the intervention that holds up, and the mechanism is straightforward, in that generating the alternative reintroduces the possibilities the outcome deleted. It's real and it's partial. It's also work you have to remember to do at a moment when nothing feels uncertain.
What actually works
The fix is a record made before the outcome existed, and it's clerical rather than psychological, which is why it works where the mental disciplines only half do. A sentence written at decision time, with a date on it, can't be revised by what happens next, so comparing what you thought against what occurred stops depending on a witness who has an interest in the verdict. When Michael Mauboussin asked Daniel Kahneman what one thing would most improve an investor's performance, the answer was a cheap notebook: write down what you expect and why, so that hindsight can't later persuade you that you knew.
The version I keep adds a number to it, because "I thought it would probably work" survives contact with almost any outcome, while "70%" doesn't. Twenty dated entries with numbers on them will tell you things about your own judgement that reviewing the same twenty decisions from memory cannot. The template has the fields, the method essay explains them, and the evidence for the practice is its own page.
Full disclosure: I make Reckon, an iOS decision journal — the prediction and the confidence number get locked to the date you wrote them, and the app asks again a month or two after the decision resolves, which is the point at which your memory of the call has drifted furthest and the written version is worth the most. One-time, no subscription. A dated note in any app you already have does the same job, and the app is for when the follow-up is the part you'd let slide.