Confirmation bias: why we find only the arguments in favour

We find only the arguments in favour because the bias sits in the query, not in the conclusion. The decision to trade matures before the gathering of data begins, and from then on the search is worded in such a way that a negative answer physically cannot come back: a person who has already bought the security types "growth prospects" into the search box, not "why the idea will not work". The arguments against are not rejected — they simply never make it into the sample. Hence the main consequence: confirmation bias has to be fought at the input, by changing the wording of the question, and not at the output, by trying to weigh honestly what has already been collected.

The query is biased, not the conclusion

The definition of the term is covered briefly in the site's glossary — confirmation bias. But a definition does not explain why the bias is so persistent in people who regard themselves as analysts. The persistence rests on the fact that the procedure looks flawless: hypothesis, data gathering, conclusion. Only one link is broken — the data are sought according to whether they are compatible with the hypothesis. Formally a test took place; in fact what took place was an illustration.

The bias works in layers, and each of them, taken separately, seems reasonable. Search: you open the pieces whose headlines promise a familiar thought. Interpretation: an ambiguous fact is read in favour of the position ("the report is weak, but those are non-recurring write-offs"). Memory: a month later you remember the forecast that came true and do not remember that next to it stood the opposite forecast, which you considered no less convincing at the time. The sum is a feeling of accumulated rightness that is backed by nothing except selection.

The multiple as a mirror

The most common entry point for the bias is an indicator that can be read in either direction. You look at 3,78 and see what you came to see: a low value becomes confirmation of cheapness, a high value becomes confirmation of quality. Both readings are available at the same time, so the multiple never refutes the position. That makes it ideal fodder for the bias — and a poor sole argument. The discussion of why a low P/E does not mean "cheap" is useful here not as a calculation method but as a list of the mechanisms that can make the denominator unreal.

The check that really cuts is to compare accrued profit with cash. Profit is the result of accounting decisions, and it is the easiest place to read in what you wish to see; the cash flow statement leaves less room for a charitable interpretation. If cash flow diverges from profit and you explain the divergence by "the specifics of the industry", that is exactly the moment when the bias is at work. The statements themselves can be opened in the issuer reports section, and the {{instrument:SBER}} card gives you the data before a narrative has stuck to them.

The regulator says what you want to hear

Central bank communication is built in such a way that it almost always contains a fragment to suit any position: the signal comes with conditions, caveats and a list of risks. A holder of long OFZ reads into the same text a hint of easing, while a seller reads a hint of caution, and each of them can produce a verbatim quotation. That is why the breakdown of the signal's wording is worth reading before you have a position in long bonds: afterwards you will look for confirmation in the text, and you will find it. The same trap lies in calculating the return on bonds: when there are several acceptable ways to calculate yield, the bias will choose the most flattering, and there is more on this in the piece on why yields differ between brokers.

Confirmation comes from what you decided to measure

The most hidden form of the bias is the choice of metric. While holding a losing position, a person watches the depth of the drawdown, because depth fluctuates and regularly offers grounds for hope. The duration of the drawdown grows monotonically and offers no hope, so nobody looks at it. Yet it is duration, not depth, that determines the cost of holding on. There is a similar blindness with liquidity: as long as you are not exiting, liquidity goes unnoticed, and its absence does not refute the thesis — it simply is not on the list of what is observed until the moment it becomes decisive.

A dividend record confirms the thesis especially readily: the payment was made, therefore the idea is right. It is more useful to look not at the fact of the payment but at the calendar of upcoming payments — {{dividend_calendar|limit=5}} — and at what the payment is funded from.

What really works: a refutation condition written down in advance

Step 1 — before the trade, state which observation will make you exit: not "if it falls", but a specific event in the financial statements, in liquidity, in the regulator's policy. Step 2 — write it down somewhere it cannot be rewritten later: in hindsight the bias edits the conditions instead of admitting the mistake. Step 3 — set a review date that is not tied to the price: otherwise the review will happen only after a rise, that is, at the moment when it will confirm you.

And a separate technique: frame the search from the opposite side. Instead of "why will this security rise" — "what would have to be true for it to fall, and how can that be checked against the data". A query with a negative answer built into it returns a different sample. This does not make you objective, but it deprives the bias of its main tool — the right to word the question.

What this piece does not claim

The site has no data on how often confirmation bias affects the decisions of a particular investor, and such measurements are difficult in principle: the effect is visible only in comparison with a decision that was not taken. So what is described here is the mechanism and the ways to intercept it, not an estimate of its scale. The terminology can be checked in the glossary and in the term card.

{{callout:warning}}If, after reading this, you are looking in your own position for confirmation that the bias does not apply to you, you are already inside the mechanism described. Test the opposite: which observation would convince you that it does.{{/callout}}