Hindsight Bias: Why the Past Looks Predictable

Hindsight bias is the revision of what a person remembers having expected, once the actual outcome is known. Its target is not the verdict on the past decision but the memory of one's own state before it. Rewriting the memory is what makes hindsight bias the most destructive error in this set: it removes the evidence needed to detect any of the others.

Most of the errors in this subject area attack a decision at the moment it is taken. Hindsight bias waits until afterwards and attacks the only surviving witness. Baruch Fischhoff reported the effect in the Journal of Experimental Psychology in 1975, showing that people told how an episode had turned out then reported having assigned a higher probability to that ending beforehand than people who had never been told. The people reporting were not lying and they were not careless. Each of them was asked to remember, each remembered, and what came back had already been rewritten. Hindsight bias rests on the fact that a memory is rebuilt at the moment it is asked for rather than lifted out of storage, and the material it is rebuilt from includes the answer.

What does hindsight bias actually distort?

Two things that sound like one thing need separating. The first is the assessment of a past decision: was it sensible, given what could be seen at the time. The second is the person's memory of their own prior expectationWhat somebody actually believed before the outcome was known.: what the person honestly thought was going to happen before anything happened. Hindsight bias leaves the first alone and works on the second.

Here is the everyday version. A household member is unwell on a Tuesday and the household decides to wait a day before going to a doctor. On Thursday it turns out to have been something that needed attention on Tuesday. Ask anybody in that household on Friday what they thought on Tuesday, and a very high proportion will say some version of "I did feel it was more serious than the others thought". Some of them genuinely did. Most of them did not, and there is now no way to tell which is which. Nobody in that household has told a lie. The distortion is not in the story anybody tells about the decision, it is in the raw material that story is being told from.

Corrupted raw material is why the effect is so often underrated when it is first met. Hindsight bias sounds like a mild vanity, a tendency to claim credit for having seen things coming. The effect is worse than a vanity. A vanity can be argued with. A revised memory arrives with exactly the same feeling of certainty and detail as an accurate one, and it arrives first, before any argument can start.

Why is a memory rebuilt rather than retrieved?

The word remember suggests a filing cabinet. A document is requested, a drawer opens, the document comes out unchanged. A mind does not answer a question about its own past expectation that way. Almost nobody files an expectation. The mind stores fragments: some of what was known, some of how it felt, some of what was said. When the question arrives, those fragments are assembled into an answer that makes sense right now, and the assembly is called reconstructionRebuilding a memory from current knowledge at the moment it is asked for, rather than retrieving a stored one..

Reconstruction is not a defect. Reconstruction is what lets anybody answer questions about their own past at all, given that no mind has the storage to file everything. The trouble is only in what goes into the rebuild. On 12 October the ingredients available were what was known on 12 October. On 31 March the ingredients available are what was known on 12 October plus everything that has happened since, and that second pile includes the result. The rebuild uses whatever is on the shelf, and after the fact the outcome is on the shelf, so it goes into the expectation the same way every other ingredient does.

Why does nobody catch themselves doing it?

Because the rebuild is seamlessProducing no sense at all that anything has been revised.. Seamlessness is the property that turns an ordinary quirk into the most awkward error in the set. An ordinary change of mind can be felt happening. There is a before and an after, and whatever moved the person can usually be named. Nothing like that occurs here. The revised expectation does not arrive labelled as a revision. The revision arrives as the memory, in the first person, with texture and detail, and often with a supporting recollection of having said something at the time.

Set it beside a mistake that can be caught. A column of figures added wrongly produces a total that will disagree with something eventually, and the disagreement is a signal that can be acted on. Hindsight bias produces no disagreement. The error revises the record and the recollection of the record at the same moment, so the two still match perfectly. An internal consistency check is exactly the wrong instrument for an error that maintains internal consistency.

Hindsight Bias vs Outcome Bias: what is the actual difference?

Hindsight bias and outcome bias get treated as one thing constantly, including by people who know both names, and separating them is the single most useful move anybody can make on either. Outcome biasJudging the quality of a decision by how it turned out rather than by what could be known when it was taken. was set out by Jonathan Baron and John Hershey in the Journal of Personality and Social Psychology in 1988, and is defined in full in its own treatment. The comparison needs only what outcome bias acts on, set against what hindsight bias acts on instead.

Outcome bias produces a verdict. Somebody looks at a decision, sees how it ended, and grades the decision by the ending. The sentence it generates is about the decision: a poor call, and look what happened next. Hindsight bias produces a claim about a person's own past mental state. The sentence it generates is about the expectation: I thought at the time that this would happen. One is an assessment. The other is testimony.

The difference matters because the two have completely different repairs. Outcome bias has a well known correction, and it is a question: what could actually be known when the decision was taken? Ask it honestly and the verdict changes. The result is deliberately set aside, and only the information available beforehand is allowed to count. The correction works on outcome bias and fails completely on hindsight bias. The answer to what could be known at the time is supplied by a memory that has already absorbed the result. The repair for the first error runs through the machinery the second error has already got to.

What happens when the log is read back at the two dates?

Take the case that runs through this set of errors. Meera Sundaram, an invented investor whose decisions are recorded in the invented Palash decision log kept by Palash Advisory Services Private Limited, took one decision on 12 October. She sold Suvarna Chemicals Limited whole at Rs 4,60,000/-, and she kept Kesari Logistics Limited, then standing at Rs 1,95,000/-.

The log is read back on 31 March. Suvarna Chemicals Limited had risen a further 8.0 per cent after the sale, so the same holding, left alone, would have stood at Rs 4,96,800/-. She did not have that Rs 36,800/-. Kesari Logistics Limited had fallen a further 20.0 per cent, from Rs 1,95,000/- to Rs 1,56,000/-, a fall of Rs 39,000/-. Put the two together and the pair costs Rs 75,800/-. One case is never evidence that any rule works.

Now the step that looks harmless. Meera is asked on 31 March what she expected on 12 October. Suppose that on 12 October the probability she would honestly have put on Suvarna Chemicals Limited rising from there was 20.0 per cent, low enough that she was content to sell. On the illustrative scale below, somebody asked with no knowledge of the outcome reports that same 20.0 per cent. Somebody asked once the whole outcome is known reports 48.0 per cent, or 2.4 times the figure they actually held, and reports it with no sense of having changed their mind about anything.

At that point the 12 October decision has become unreviewableA decision whose reasoning can no longer be assessed, because the evidence needed to assess it has been altered., and note carefully what that does and does not mean. Unreviewable does not mean the decision was wrong. Unreviewable does not mean it was right either. Unreviewable means the question can no longer be answered by anybody, including by Meera. The only surviving account of what she thought beforehand has been overwritten by what happened afterwards. A decision that has stopped being answerable is a stranger and more serious situation than a wrong one.

Why does this one destroy learning from a record?

Learning from past decisions has a shape, and the shape has one weak joint. A decision is taken, something happens, and then what happened is compared with what was expected. The comparison is the whole engine. The comparison is what distinguishes reasoning well and getting an unlucky result from reasoning badly and getting a lucky one, or neither. Without the comparison, nothing is left but a list of things that occurred.

Hindsight bias removes exactly one of the two terms in that comparison, and it removes it quietly. The outcome is preserved perfectly, written into the values and the dates. The expectation is not preserved at all. Nobody wrote it down, and the mind asked to supply it now supplies something built from the very outcome it is meant to be compared against. The comparison then runs between the outcome and a version of the outcome. Two versions of one outcome always agree, so the review reports that everything was foreseen and finds nothing to change.

The missing term has a consequence worth naming plainly. A person who reviews decisions without a written prior expectation will tend to conclude, honestly and repeatedly, that they saw most things coming. Such a conclusion is not a character flaw and it is not arrogance. The conclusion is the arithmetic of a review run on a corrupted input, and it will happen to a careful person and a careless one alike.

Does knowing about hindsight bias help?

Very little, and this is the finding that surprises people who have just learned the name. With most errors, being told about them buys something. Once a vivid recent story is recognised as pulling on the judgement, a pause becomes possible in which to ask whether the story is representative, and the pause does real work. Awareness supplies a moment in which to intervene.

Hindsight bias does not offer that moment. There is no point at which a revision is noticed taking place. The revision does not present itself as a revision. Somebody who has understood hindsight bias perfectly will still, three months from now, remember having expected whatever happened, and will feel the ordinary confidence of remembering. Awareness has nothing to catch hold of. Knowing about hindsight bias changes what a person concludes about other people's recollections far more than it changes their own. The defence therefore has to be built out of paper rather than out of attention.

What is the only reliable defence?

A written recordAn expectation committed to paper before the outcome is known, which memory has no way of revising afterwards. made before the outcome is known. The written record is the whole answer, and the list has no second entry. Every other candidate defence works by asking somebody to remember accurately, and remembering accurately is the exact capability the error removes.

Think about what a written line actually is. A written line is a physical object, or a file with a date on it, that exists outside the mind doing the remembering. On 31 March, memory can revise the recollection of the reasoning, the recollection of the mood, the recollection of what was said out loud, and the recollection of how confident everybody sounded. Memory can revise every single item in the process except the sentence that was already written down before anything happened. The written sentence is not stored in the place the revision operates on.

Here is the household version again, and it makes the point faster than any argument. Two people disagree about what was said at a meeting six months ago, and both are completely sincere. If one of them wrote a line in a notebook that evening, the disagreement ends in four seconds, and it ends without anybody having to be accused of anything. The notebook is not superior memory. The notebook is a different kind of object.

What must a written expectation contain to work?

Not everything written down beforehand does the job, and this is where most first attempts fail. The test is simple and slightly uncomfortable: could the sentence turn out to be false? If no possible future makes it wrong, it has recorded nothing, however carefully it was composed and however sincere the person composing it was.

A note saying the holding may rise or may fall passes every test of honesty and fails this one completely. Such a note is true whatever happens next, so on 31 March it can be read as consistent with any outcome at all, and it will be. The same goes for the reasoning is sound, or the situation is being watched closely. A written expectation earns its place only by naming something specific enough that a future can contradict it. A number and a condition beat a paragraph of prose every time. The note in the figure above works because it says one in five, and because it says what a rise on the same information within two quarters would mean.

When is a sense of inevitability legitimate?

Hindsight bias does not imply that nothing was ever predictable, or that anybody who says they saw something coming is fooling themselves. Plenty of things are predictable, and plenty of people do predict them. The feeling of obviousness is not the problem. The feeling of obviousness in the complete absence of any contemporaneousMade or written at the same time as the event, rather than reconstructed later. evidence is the problem. In that situation the feeling is equally consistent with having seen it coming and with never having thought about it at all.

So the question to put to a claim of foresight, including one's own, is not whether it feels true. The question is whether anything exists that was made before the outcome. A written note from beforehand converts a feeling of inevitability into an ordinary fact that can be checked, and without one the feeling carries no evidential weight in either direction. Notice that this cuts both ways: it is just as wrong to dismiss somebody who genuinely did call it as to believe somebody who did not, and the record is what distinguishes them.

How does somebody reviewing decisions actually use this?

Two people are served by the same point. Devika Rao, the adviser at Palash Advisory Services Private Limited, sits down with a client to look back at a decision. She asks the natural question: what was the client expecting at the time? Whatever comes back is testimony given after the result, and Devika Rao has no way to weigh it. If she treats it as evidence, her review is measuring the outcome twice and calling the second measurement an expectation. If she discards it, she has no expectation at all to compare against. Either way the conversation cannot do the job it was arranged to do, and the reason has nothing to do with the client being unreliable.

The person deciding alone is in a harder position, not an easier one. Nobody is in the room whose account differs. A committee at least generates disagreements, and a disagreement is a signal that something is being reconstructed. One person reviewing their own decisions gets a single, coherent, entirely consistent account every time. The practical rule is negative and specific: any recollection of a prior expectation, including one's own and including one given in complete good faith, carries no evidential weight unless something written before the outcome supports it.

The scale of the problem is visible in the log itself. Of the 240 decisions it records, a written reason was set down on 84 of them, or 35.0 per cent. The remaining 156, being 65.0 per cent, reach any later review with no prior expectation attached to them at all.

The same gap is why a conduct file matters to a professional for reasons beyond compliance. A file that shows what was expected and why, written at the time, is the only thing that lets anybody afterwards distinguish a sound process with a poor result from a poor process with a good one. Any duty an intermediary has to keep such records is set by the Securities and Exchange Board of India.

What does hindsight bias not explain?

Hindsight bias does not explain why the 12 October decision was taken. Whatever drove that sale, it was in place before anybody knew anything about 31 March, and it belongs to the errors that act at the moment of deciding rather than afterwards. Hindsight bias arrives only when somebody looks back.

The error says nothing about what happened next either. The 8.0 per cent rise and the 20.0 per cent fall are entries in one record, and hindsight bias has no view on why they occurred or on what any past movement implies about any future one. Hindsight bias is a claim about memory and about nothing else, and stretching it into a claim about what will happen leaves the subject altogether.

Meera Sundaram, Devika Rao, Palash Advisory Services Private Limited, the Palash decision log, Suvarna Chemicals Limited and Kesari Logistics Limited are invented.
Educational material. Not advice on any investment, tax, budget or market position.