Cognitive biases in trading - how to spot a distortion in your own record
After a loss, it is easy to name what happened: confirmation bias, anchoring or overconfidence. The label alone, however, does not explain the decision. You need a pre-order record, observable behaviour and a check for simpler causes such as noise, costs or a change in market conditions. This article examines several cognitive biases relevant to trading, shows situations that should not be mistaken for them and presents procedures for checking them. The object of assessment is how information was used, rather than whether an individual trade ended in a profit.

Write the evidence standard before choosing a conclusion
Finding several supportive headlines is not the same as finding several independent pieces of evidence. They may all repeat one claim. A useful first check is whether the source would pass the same quality test if its direction were inconvenient. The answer should come from a standard recorded before the decision, rather than a new requirement invented for the contrary observation.
Use a wholly fictional report R, measured in index points. An exercise condition for considering a long position requires the latest available primary release for the specified period to be at least 50. The source must identify its issuer, publication time, period and unit. The decision is scheduled for 10:00 UTC, with all records using that clock. This arbitrary threshold is a data-handling exercise, not a claim about a real indicator or a profitable trading strategy.
Trace each statement to its source and available time
| Record | Seen or published UTC | Value | Source status |
|---|---|---|---|
| A | 09:55 | 51.6 | Secondary post; period not identified |
| B | 09:58 | 49.4 | Primary report R; all required fields present |
| C | 09:59 | 51.6 | Copy of A; no new underlying observation |
| D | 10:06 | 50.3 | Primary correction to R; all required fields present |
At 10:00, B is the qualifying record and its value is below the threshold. A fails the prior source requirement, and C merely repeats A. Two copies do not supply two independent observations. D has not yet been published, so it cannot justify the decision at that time. The exercise’s data condition for a long position is therefore unmet.
Now reverse the emotional attraction of the numbers without changing their source fields. The same acceptance test should still apply. If a review finds that the trader accepted A because it supported the preferred conclusion while rejecting B solely because it contradicted it, the record establishes inconsistent use of the stated standard. It does not establish why the inconsistency occurred. A reading mistake, misunderstood timestamp or deliberate rule change still needs checking.
Update after a correction without backdating what was known
At 10:07, the primary correction D is available. Under the exercise’s latest-release rule, its value of 50.3 now passes the threshold of 50. A new assessment can therefore differ from the one at 10:00 without being a bias. Preserve B as the earlier release and D as the correction, together with their publication times. Replacing B in the old decision record would falsely move later information into the earlier information set.
Acceptance of a primary source is not a promise that it will never be corrected. It identifies which record the procedure uses and when. Nor do the index values establish a causal effect on exchange rates. The example tests consistent treatment of evidence. Evaluating the economic hypothesis would require additional arguments and observations beyond this source log.
Use a possible counterexample without making a diagnosis
Wason’s conceptual-task experiment examined how participants tested proposed rules using confirming and potentially eliminating evidence. Its laboratory number task is not a measurement of this trader or the Forex market. The transferable question is practical: which available observation could contradict the proposed explanation, and was it assessed under the same standard?
Keep the review narrow. A repeated headline is a source-dependence issue; an unavailable correction is a timing issue; selective application of a written standard is an observable inconsistency. A cognitive mechanism remains a hypothesis to investigate, not a name supplied automatically by a losing result. Retain rejected sources and reasons as well as accepted ones, so another reader can reproduce the classification. The detailed sections then distinguish confirmation bias, anchoring, hindsight and alternative explanations rather than using one label for every mistake.
Three exercises with worked answers
The exercise data are hypothetical. Choose an answer, then check the calculation.
Continue from the evidence log to specific mechanisms
The detailed sections preserve the research discussion and illustrations. The fictional report above checks sources and timing; it does not diagnose a cognitive bias from an account outcome.
1. A shortcut is not a defect - when a heuristic becomes an error
A heuristic is a rule of thumb: a way of judging a situation that skips part of the information in order to produce an answer faster and more cheaply. In most applications it works well. A driver does not solve equations before overtaking, and a doctor on call does not begin with a full differential diagnosis. A shortcut turns into a systematic error only in a particular environment - one in which feedback arrives rarely, with a delay, and mixed with noise.
The market is exactly that kind of environment, and that is the whole difference between a cognitive trap and an ordinary mistake. If, instead of seeing the consequences of your decisions immediately and unambiguously, you get them after hours or weeks, in a version blurred by a thousand things you had no influence over, then learning from your own experience stops working in the obvious way. The shortcut is never corrected, because nothing corrects it.
One thing follows from this and is worth stating at the outset: this is not an article about the human mind being defective. A tool is not defective because it was used outside the range in which it performs.
A label is not an explanation
The fact that a behaviour can be described as confirmation bias does not prove that confirmation bias caused it. The description fits a great many runs - including those in which there was no distortion at all. Adding a name after the outcome is known is cheap and always succeeds, so it carries no information.
For the name to mean anything, two conditions must be met. First: a record made before the decision exists, against which what was actually done can be compared. Second: the simpler explanation has been checked and rejected. There are several simpler explanations and they are almost always closer at hand: chance, transaction cost, the way the order was executed, a change in market conditions, plain lack of knowledge. An ordinary informational error - the trader did not know something that could have been checked - is not a cognitive bias but a gap in preparation.
This whole text assumes, moreover, that the rule the trader is executing has any justification after costs. Whether a set of rules has a positive expected value is settled by section 8. Psychology does not create an edge where none exists, and diligently executing a baseless rule produces nothing but diligently executed losses.
Seven distortions and the limits of this article
This article covers seven mechanisms. For each, we examine the research evidence, the behaviour that can be observed and the alternative explanations to consider. The examples and review rules help you analyse decisions, rather than assign yourself another label. Learning more names cannot replace a record that allows a particular case to be assessed.
Three families of mechanisms belong to other articles in this section and I do not develop them here. The reaction to a loss, the reference point, the disposition effect and sunk costs are the psychology of losses. Judging a decision by its outcome, and the full application of hindsight bias to settling trades, is outcome versus process. The escalation sequence after a loss is revenge trading and tilt. The arithmetic of losing streaks is handled by the article on uncertainty, and the construction of the record by the article on the journal.
2. Confirmation bias: an asymmetry in what evidence has to prove
Confirmation bias is not a matter of a person looking for evidence for their own thesis. Looking for evidence for a thesis is a normal way of working and spoils nothing by itself. The distortion lies in the asymmetry: data supporting the hypothesis is accepted at a lower quality threshold than data contradicting it. The same argument, made by the same analyst, is judged sound when it is on our side and speculative when it is on the opposite one. Raymond Nickerson's 1998 literature review organises this as a family of phenomena rather than a single effect: selective search, selective interpretation and selective recall operate independently of one another[4].
A classic demonstration comes from Peter Wason’s 1960 task[3]. Twenty-nine psychology students received the sequence 2, 4, 6 and had to discover the rule by proposing sequences and receiving feedback on whether they fitted. The rule allowed any three increasing numbers. The sequence 8, 10, 12 would confirm both that rule and the narrower hypothesis of “consecutive even numbers”; the increasing sequence 1, 2, 3 would distinguish them. Participants varied: six stated the correct rule on their first announcement; the others made errors or did not reach an answer. These frequencies cannot simply be transferred to traders. The practical lesson concerns testing: look for information that can distinguish your hypothesis from a competing one.
How to check for selective use of evidence
The moment at which the asymmetry switches on is usually a single one: opening the position. Before the click the trader has a hypothesis; after the click they have a position - and from that moment contradicting information costs more, because it entails a concrete action. There are three observable consequences. The number and choice of sources checked changes. The way the same signal is described changes: a break of a level was a "signal" before entry and is a "test" after it. And the exit condition changes - not on paper, but in practice, through reasons added along the way for why it does not apply just now.
What this is not. It is not confirmation bias when a trader holds a position in accordance with a condition written down before entry - even if the position is losing and even if it ends worse. Nor is it confirmation bias to change your mind after new information, if that information genuinely appeared later. The criterion is single and it is formal: was the threshold for accepting evidence set before that evidence was seen. Holding a position as a mechanism belongs, in any case, to the article on the psychology of losses; here I am interested only in how the selection of information changes.
Record the evidence and the condition for changing the hypothesis
You can see this in the record without introspection. The field "what would refute this hypothesis" is empty, or filled with a sentence that cannot be checked ("if the market goes against me"). Justifications for holding the position appear after opening, not before. The number of sources checked is clearly larger after entry than before it.
The rule that bypasses this is a single one and fits in one sentence: you write the invalidation condition down before submitting the order, in observable form - a level, a time or a specific event about which it can be stated that it "occurred" or "did not occur" without interpretation. Trigger: the moment before the click. Action: recording the condition plus one piece of information that would refute the hypothesis if it appeared. Completion criterion: the condition occurred - position closed; it did not occur - the position stays. Audit: both fields are in the record and can be reviewed across a series of decisions.
How to recognise that the rule is not working. If the invalidation condition never occurred in twenty consecutive entries, check whether it is observable and consistent with the plan. Its absence alone does not prove a defective definition: the condition may be rare, or a loss may have followed a separate exit rule. A stronger warning sign is a condition that cannot be checked unambiguously. Another symptom is a condition that occurred and was recorded, but prompted action only later, with an added justification. In that case the rule exists but was not executed — a problem discussed in the article on procedures.

3. Overconfidence comes in three forms
Overconfidence is, in everyday use, one word for three different things. Don Moore and Paul Healy separated them in 2008 in a review published in Psychological Review[5]. Overestimation concerns your own performance on a task: I think I did better than I did. Overplacement concerns the comparison with others: I think I am better than most. Overprecision concerns the width of an interval: I state a narrower range of possible values than my data justifies.
The most important observation from that work is that these three forms do not move together, and sometimes move in opposite directions: on hard tasks people overestimate their own performance but underrate their standing relative to others, and on easy tasks the reverse. So there is no single "overconfidence" as a trait - there are three separate discrepancies between what is stated and what is the case.
In trading each of them looks different. Overestimation of your own performance is a remembered hit rate higher than the counted one. Overplacement is the belief that statistics about retail clients as a whole concern somebody else. Overprecision is a forecast in the form of a point rather than an interval - and it has the most direct consequence, because whoever states a single value has no scenario for the rest.
What account data shows
The empirical material most often invoked here is the work of Brad Barber and Terrance Odean. In their 2000 analysis covering 66,465 households using an American discount brokerage between 1991 and 1996, the highest-turnover households achieved about 11.4% a year net, the average household about 16.4%, the least active about 18.5% and the market about 17.9% - a difference between the extreme turnover groups of about 7.1 percentage points, largely a matter of costs[6].
It has to be said clearly, though, what that study does not show. Overconfidence was not measured in it; it is the authors' interpretation, a proposed explanation of the observed relationship between trading frequency and net result. The relationship itself is correlational and has competing explanations. [EXTRAPOLATION] The study concerns common stocks on the American market in the 1990s, under a completely different cost structure from the retail currency market. I carry over only the direction: trading more often raises total costs, and costs are subtracted from the result regardless of what the trader thinks of themselves.
What this is not. High conviction is not in itself a distortion. A trader who states 70% and is right in 70% of cases is simply well calibrated, and lowering their confidence would make their decisions worse. The problem is not the level of the statement but the systematic discrepancy between the statement and the frequency - and this is one aspect of overconfidence a trader can assess using their own data.
4. Calibration - assessing stated confidence
Calibration is the agreement between a stated probability and the frequency of hits. If “about 65%” was recorded before entry and the target was reached in roughly 65% of those cases, the results are consistent with the statement; if it was reached in half, the discrepancy needs investigation with allowance for sampling uncertainty. This is one aspect of overconfidence that can be assessed using your own data. The pre-outcome record, common confidence bands, hit frequency, uncertainty and sample size are covered in article 10.1; we use the same bands here to avoid incompatible records. The comparison also needs checks for different entry setups, instruments and session times, which may affect hit frequencies. Good calibration alone is not an edge: a stated and observed 55% hit rate can coexist with losses at an unfavourable profit-to-loss ratio.
5. Anchoring and reference values
Anchoring is the influence of a value learned earlier on a later estimate, including when that value carries no information at all. Amos Tversky and Daniel Kahneman showed this in 1974 with the wheel-of-fortune experiment: subjects watched a number being drawn and then estimated what percentage of United Nations member states are African states[1]. The group that saw the number 10 gave a median of 25%; the group that saw 65 gave a median of 45%. The number came from a wheel, and everybody saw it happen.
The objection that this is an artificial laboratory task is fair, so it is worth citing a study in a real domain. In 1987 Gregory Northcraft and Margaret Neale invited two groups - professional estate agents and students - to value the same real house, which participants could inspect, together with a full information pack[7]. The only element that differed was the listing price given. The valuations of both groups moved with it. More interesting than the effect itself is the agents' reaction: asked about the factors they had taken into account, most of them did not mention the listing price. [EXTRAPOLATION] The property market is not the currency market, and a one-off valuation is not a trading decision. I carry over only two things: experience in a field does not protect against anchoring, and an anchored person usually does not name the anchor among their premises.
The anchors a trader has
The most common one is the entry price. The market does not know it and has no reason to respect it, yet every later assessment is measured from it: not "the price is such and such and the structure looks like this", but "I am such and such an amount below my entry". The second is a round level. The third is your own earlier forecast - a number written into the record a week ago comes back as a reference point, although it was produced with a different set of information. The fourth is the account high or the day's result, but that already belongs to the article on the psychology of losses, which treats the reference point as a separate mechanism.
Not every reference value is an anchor. A level at which orders genuinely cluster carries information about market structure, and using it is not an error. The entry price can matter too - for instance when calculating the cost of holding the position. The criterion is simple to apply and requires no looking inside your own head: can you justify this decision without stating your own entry price? If the decision rests on current information independent of the entry price, there is less reason to suspect anchoring. If the justification falls apart without that price, check whether it still has a genuine role, for example in costs or a rule recorded in the plan. Only the absence of such relevance supports an anchoring interpretation.
With round levels the caveat from the first section applies. The claim that "liquidity is sitting" at a given place rests on a picture that a retail participant in the spot market does not have - they see a slice of their broker's quotes, not an aggregated order book (section 11). The level may be significant, but the sentence "I know there are orders there" is usually stronger than the data the trader has.
What shows up in the justifications for modifications
The signal is textual and easy to catch during a review: in the justifications for closing or modifying a position the same number keeps appearing, and that number is the entry price. The second signal is protective orders placed at equal distances from the entry instead of at places that follow from structure.
The rule: you write the justification for every modification of an open position without using the entry price. Trigger: any change to a protective order, a target or a size. Action: one sentence describing what in market structure justifies the change. Completion criterion: the sentence is written, the change made or abandoned. Audit: at review you can see how many justifications could be formulated and how many could not.
How you know it is not working. If the sentences can always be written and always sound equally general ("structure still favourable"), the rule has turned into a formula. The test is the number of modifications that, after it was introduced, did not happen at all - if that number is zero, the rule stops nothing.
6. Recency and availability - two different mechanisms
The availability heuristic consists in judging the frequency of an event by how easily an example of it comes to mind. Tversky and Kahneman described it in 1973[2]: subjects systematically overstated the frequency of events that were easy to recall, if only because they were reported in the media or had a vivid form. The recency effect is a special case of it - the most recent observations come to mind most easily, so they weigh the most, regardless of how many of them there are relative to the whole history.
I keep them separate because in trading they lead to opposite behaviours. Availability makes you avoid an instrument or a situation that once ended spectacularly badly, and at the same time disregard the quiet, repetitive, boring costs - the spread, slippage, the cost of financing a position, whose mechanics are described in section 11 - which over a year add up to a far larger amount. Recency makes you rebuild a rule on the basis of the last dozen or so decisions, despite the existence of a longer record that says something different.
The arithmetic that settles the second situation is handled by the article on uncertainty. I will quote one number from it as a reference point: at a 55% hit rate and 250 decisions a year, a run of at least six losses in a row appears in 68.6% of runs and at least eight in 20.3%. A streak that looks like a breakdown is, in this order of magnitude, an arithmetical norm - and this is not an argument for always keeping the rule, only for the fact that the length of a streak on its own settles nothing.
What this is not. Here the counter-example is particularly important, because it is easy to turn it into a ban on thinking. Giving more weight to recent data is sometimes correct. If the conditions under which the rule was created have changed - a different level of volatility, a different monetary policy regime, a different cost structure at the broker - then the old observations describe a world that no longer exists, and the recent ones are more informative. A change of market regime is not a cognitive bias; it is an event. Likewise, a reaction to an event that genuinely changed the conditions of trading a given instrument is not an availability effect, even if the event was spectacular.
The same criterion as before settles it: is there a condition, written down before the streak, at which the trader decided in advance that the rule would be out of date. If such a condition is met, the change has a basis, even if a losing streak happens to be under way. If there is none, the only premise left is what the streak looks like - and the look of a streak does not distinguish a working rule from a broken one.
Compare the timing of the loss and the rule change
The signal is dated, so it requires no judgement. It is enough to set the date of each change to the rule against the date of the last loss before it. If changes systematically fall within a day or two of a loss, that is a reason to check the prior premises. The timing association alone does not prove that the loss caused the change.
The rule: changing a rule requires pointing to a condition written down earlier, and a fixed waiting period passes between the event and the change being introduced. It is triggered by the intention to change, not by the change itself: from that moment the intention is an entry with a date, a stated condition and a justification, and only after the waiting period do you come back to it to confirm or abandon it. The waiting period is not a pause for cooling down - it is an interval after which you can see whether the premise was a premise or a reaction to the last loss. After a review what remains is a dated list of changes, and that list is the only material on which anything can be checked.
How you know it is not working. If in the last three changes the stated condition was different every time and every time formulated only at the moment of the change, the waiting period merely shifts the moment at which the trader adds the justification. The second symptom: changes have stopped happening altogether, while the record shows that the invalidation condition occurred and was ignored - then the rule is not protecting the process but freezing it.
7. The gambler's fallacy, the hot hand and the lesson of 2018
Two beliefs, one source. The gambler's fallacy is the expectation of a reversal after a streak: since there have been six losses, the seventh decision "must" turn out better. The hot hand is the expectation of continuation: since there have been five hits, the sixth "should" hit too. Both assume that the process remembers its own past.
The arithmetic for independent trials is short. If a single decision ends in a loss with probability 45%, then the chance of seven losses in a seven-trial block specified in advance is 0.457 = 0.37%. But the probability that the seventh decision ends in a loss, given that the six before it already did, is still 45%. These are two different quantities, and confusing them is the essence of the gambler's fallacy: the figure of 0.37% concerns the whole sequence seen before it began, not the decision the trader is currently facing.
That this is not a theoretical problem can be seen in data from real bets. In 2005 Rachel Croson and James Sundali analysed recordings from a casino in Reno: after a run of results of one colour, the share of bets placed on the opposite colour rose with the length of the run[14]. [EXTRAPOLATION] Roulette has a known distribution and independent trials; the market has neither. I carry over only this: people change their behaviour in response to the length of a streak alone, including when the stake is real.
Why this is the best available example of epistemic caution
The history of the "hot hand" matters more in this article than the effect itself. In 1985 Thomas Gilovich, Robert Vallone and Amos Tversky analysed basketball shots and ruled that belief in a hot streak has no support in the data[9]. For three decades this was a textbook claim and the flagship example of intuition about randomness failing.
In 2018 Joshua Miller and Adam Sanjurjo showed that the method used in that analysis contains a downward bias[10]. The mechanism can be shown with three coin tosses and worked out by hand. Write out all eight sequences of three tosses. In every sequence in which a head occurred that was not the last toss, count the share of heads among the tosses immediately following a head. Average those shares across sequences. You get 5/12, that is about 41.7%, not 50%. The coin is fair; the bias is introduced by the very procedure of selecting the tosses that come "after a head" from a finite sequence. Once that correction is applied, the data from the controlled experiment that formed part of the 1985 work speaks rather for the existence of the effect than against it.
What this is not. Increasing exposure after a streak is not the gambler's fallacy if it follows from a rule written down earlier and independent of the outcome of the previous decision - the parameters of such rules belong to section 9 and are set there. There is a second, more serious limit: the whole calculation above assumes independent trials, and a trader's trades are not independent. Several decisions taken on the same day, on correlated instruments, under the same market conditions, are in practice one bet spread across several orders. A losing streak may then reflect one common cause rather than a sequence of independent draws - and in that case neither the gambler's fallacy nor its mirror image is the right description of the situation.
In the record this shows up as a correlation: position size, or the interval before the next decision, changes with the length of the current streak. The rule that bypasses it is trivial in form and hard in execution - position size follows from a rule in which the outcome of the previous decision does not appear. It is triggered by every determination of size, that is by every order, and the action is reading the parameter off the rule rather than off the memory of recent results. The completion criterion looks different here from procedures triggered by an event, because this is a standing rule rather than a procedure with a beginning and an end: it ends only when it is deliberately changed, and the change is dated and written down before the streak it concerns. The audit trail: the size of every position can be reconstructed from the rule and from parameters recorded before the streak, without looking at the results of previous decisions - if it cannot, the rule was not a rule. You will know it is not working from the appearance in the record of departures described as one-offs.
8. Hindsight bias - definition and limit
Hindsight bias is the systematic shift in the assessed probability of an event once it is known that the event occurred. Baruch Fischhoff demonstrated it in 1975[8]: participants read a description of a historical conflict and estimated the probability of four possible endings. The groups told that one of the endings had actually occurred assigned it a clearly higher probability than the group without that information - despite being instructed to judge as if they did not know the outcome. So it is not a matter of dishonesty or of a convenient memory: knowing the outcome changes what a person considers to have been foreseeable in advance.
In trading the consequence is concrete: the record of what the trader "knew" is created after the position is closed, and is trimmed by that outcome. An evening note does not register the conviction held before the decision, only a reconstruction of it. That is why a record made after the outcome is unfit for judging the process, however detailed and honest it may be.
What this is not. Knowledge acquired later is not hindsight bias. If information appeared after the trade that had not existed before, then learning something from that information is ordinary learning, not a distortion. The distortion begins only when the trader claims to have had that knowledge before the decision.
I stop at the definition here. Applying hindsight bias to settling specific trades, the process-and-outcome matrix and the criteria for judging decisions without knowing the result are the scope of the article on outcome versus process (10.4). The construction of a record that makes such a control possible, and the review with the outcome hidden, belong to the article on the journal.
9. The blind spot: why knowledge alone is not enough
The last of the mechanisms discussed concerns all the previous ones. In 2002 Emily Pronin, Daniel Lin and Lee Ross described a phenomenon they called the bias blind spot[11]: in studies involving students and airport travellers, participants compared their own susceptibility to well-known cognitive biases with that of other people. For every bias studied they assigned themselves a lower susceptibility than others. Importantly, the description of the mechanism had been presented to them beforehand - familiarity with the phenomenon did not remove the asymmetry in judgement.
The same is visible in the estate agent study from the section on anchoring: the valuations moved with the listing price, and asked about their premises the participants did not mention it. Not because they were lying. A person knows the content of their thoughts, but not all the processes that produced them, so introspection alone is not a reliable test of their own biases.
Check when you use the names of biases
The blind spot is not an accurate judgement that somebody else is making a mistake. Another person's distortion is sometimes visible and sometimes named correctly; recognising it in others may be one of the skills developed by learning these concepts. The blind spot is solely the asymmetry: the same person, the same material, the same distortion - and a systematically lower rating of susceptibility when the person being rated is themselves. Nor is it modesty or a lack of self-belief: what is measurable is the difference between two ratings, not the level of either. And it is not a mistake in a particular trade, because the blind spot is a claim about a way of judging, not about a single decision.
In trading the consequence is that the names of the biases do not land where they could change anything. They appear in comments on other people's decisions - under somebody's chart, in a discussion of somebody's loss - and in your own summaries written after a position is closed. They do not appear in the field filled in before the order, and not because the trader is careless: a feeling before the order does not reveal all the causes of a judgement.
That is why the signal in the record has to be a number, not an impression. Set two of them side by side: how many times the name of any bias appears in your entries made after a position was closed, and how many times it appears in fields filled in before the decision. If the first is high and the second close to zero, the knowledge from this article has been spent entirely on describing the past - your own or somebody else's. This signal has the advantage of not requiring you to judge yourself; it only requires counting occurrences.
The rule that bypasses this has to be a review with the outcome covered, because only then is the record being judged rather than the result. The trigger condition is independent of the content of the entries - calendar-based or counter-based, not event-based in the sense of noticing a problem: a review after a set number of closed positions, or once a quarter. The review should not depend solely on noticing your own blind spot. An external signal, such as disagreement in a review with another person, can nevertheless trigger it; a calendar or counter also ensures that it starts independently of self-awareness. Action: read your own pre-decision entries with the outcome column covered and mark those in which the invalidation condition could not be settled without knowing the result. Completion criterion: the review ends when the list of marked entries is closed and each of them has been assigned a reformulation of the condition for next time - not when the time allotted for it runs out. The audit trail is that list itself, saved with the date of the review: at the next one you can see whether the same entries and the same phrasings come back onto it.
How you know this procedure is not working for you. The answer here is different from the other rules, and it is better said outright, because the mechanism described applies to the procedure aimed at it as well. If in successive reviews the number of marked entries is zero, that must not be read as proof that there is nothing to mark: they are marked by the same person they concern, so zero is equally consistent with the procedure simply detecting nothing. There are two ways out and both are uncomfortable - hand the review to somebody who does not know the results and did not write the entries, or accept that this measurement is uninformative and stop using it as evidence of anything. A third, namely the conviction that in your own case you can see well enough, is not offered by this article, because it is the content of the phenomenon being described rather than a way out of it.
An honest word about the state of knowledge
The literature this text rests on is not uniform in reliability, and there is no reason to hide it. Several widely repeated claims from the same family have not survived checking. The depletion of willpower as a resource - a thesis still offered in self-help books as a physiological fact - was not confirmed in a large registered replication; this is described by the article on procedures. The study of judges' decisions, from which the popular thesis about "decision fatigue" came, faces serious methodological criticism and is unfit as evidence; it is revisited by the article on physical variables. And the debunking of the hot hand, as above, was itself debunked.
I chose the distortions described in this article because they have either repeatable laboratory demonstrations, or confirmation in data on real decisions, or both. That still does not mean that in any particular trade by any particular trader any of them was at work.
10. When the label does not fit - two runs side by side
Both runs below are illustrative examples: constructions showing what settling the question looks like, not reports from anybody's account. Both begin in the same place - with the record made before the decision - because only that order allows anything to be stated. Starting from the outcome and picking a name to match it always succeeds, and for that reason proves nothing.
Run A - looks like confirmation bias, but is not
The record before the decision. 09:12. Entry condition: a break of the morning range together with an hourly candle closing above its upper boundary. Invalidation condition: price returning below the upper boundary of the range and an hourly candle closing beneath it. Stated conviction: statement band 50-60%. Information that would refute the hypothesis if it appeared: no continuation over two consecutive candles.
Execution. Entry at 10:04, in line with the condition. A protective order submitted together with the entry, at a place that follows from the range. No modifications in the record.
The run. Price moves sideways for three hours, then returns inside the range and triggers the protective order at 14:20. The invalidation condition - an hourly candle closing below the boundary - did not occur once before the order was triggered.
Assessment of the process, independent of the outcome. Entry in line with the condition, protective order in line with the plan, no modifications, invalidation condition did not occur. Process executed.
Competing explanation of the result. Ordinary noise: at a hit rate of the order of 55%, a loss is an outcome to be expected in nearly half of cases. A second one: the conditions of the day - low volatility, no continuation after the break - that is, the situation in which this particular rule loses most often.
What this run does not prove. It does not prove that the rule is good, nor that the trader is free of confirmation bias. It proves something narrower: in this trade there is no material on which that label could be based. Had the trader added it in the evening - and that is very likely, because the trade ended in a loss - they would have attached a cause to an event in which nothing corresponds to it.
Run B - the violation is there, the mechanism remains a hypothesis
The record before the decision. The same entry condition and the same invalidation condition. Stated conviction: statement band 70-80%, with a note that the setup is "exceptionally clean".
Execution. Entry in line with the condition. At 11:40 an hourly candle closes below the boundary of the range - the invalidation condition occurred and is noted in the record. The position stays open. Between 11:45 and 13:00 three added notes appear in the record justifying the hold: two refer to external publications, one to "testing the level". Position closed manually at 13:05.
Assessment of the process. The violation is unambiguous and requires no interpretation: the condition written down before entry occurred at 11:40, the action followed at 13:05. The interval is 85 minutes, and in that time three justifications were produced that had not existed before entry. That is a fact from the record, not a judgement.
What follows from this and what does not. It follows that the process was violated and that this can be measured - the interval between the condition occurring and the action is a number that can be tracked across a series of decisions. It does not follow that the cause was an asymmetry in the selection of evidence. Several other explanations fit just as well: the trader may have been away from the terminal at the time, may have judged the condition to be badly formulated, or a reluctance to close the position at a loss may have been at work - and that is already a different mechanism, described in the article on the psychology of losses. The three added notes are a premise in favour of confirmation bias, but one trade cannot settle it.
When the label starts to have a basis - and how far that basis reaches
The pattern from a single trade has to be carried over to a series, and in carrying it over one question decides everything: which layer of time the variable splitting the decisions into two groups is taken from. The answer "by whether the position was profitable" is the first one that comes to mind, and it does not work under either of the two possible readings - which is why I write both out instead of choosing quietly.
The result after the position was closed. We consider only decisions in which the invalidation condition occurred at all. In such a subsample an interval close to zero means closing right after the adverse move, so almost always at a loss, while a large interval means the position lived on and some such positions had time to come back into profit. Profitable positions therefore land in the large-interval group mechanically, through the very way the sample is selected. The relationship comes out the opposite way round from the postulated one: a uniformly slow trader comes out clean and a fast one suspect - because the test measures the moment of closing, not the way evidence is handled. This is exactly the error the rest of the article warns against: pronouncing on the process from the distribution of results.
The paper result at the moment the condition occurred. This reading is coherent and generates no selection, but it then measures something other than it promises. An interval growing with the size of the paper loss is a description of reluctance to close a position at a loss, that is of the mechanism from the article on the psychology of losses - the same one that was not rejected among the competing explanations of Run B. The test would therefore point to a different course number than the name it was supposed to license.
What is left once the result is subtracted. What is left is a variable built solely from the record made before the decision and from two timestamps, without a single monetary value. For each of twenty decisions, four fields: whether the invalidation condition was written down before entry in observable form; whether it occurred and at what time; when the action followed; and how many justifications for holding appeared in the record between those two moments, counting only those that had not been in the pre-decision record. All four come from the record, none from an export of results. Twenty decisions are an illustrative review sample, rather than a threshold proving a regularity. Repeated records may reveal an execution problem within that sample; generalisation requires further observations and attention to conditions. The sample sizes for calibration in article 10.1 are not automatically thresholds for a different type of analysis.
This review can establish whether departures from the condition recur in the observed decisions. It does not prove a lasting regularity or its cause. If the condition occurred a dozen or so times out of twenty decisions and every time the action followed with a delay, and arguments that had not existed before entry accumulated inside the delay window, then that is a fact from the record and it needs no label at all to be disturbing. It also rejects two of the three competing explanations. A loosely formulated condition stops being in play, because a condition written in observable form either can be settled or can never be settled, and that is visible across all twenty entries at once, not only in the losing ones. Absence from the terminal drops out too: an absent trader does not add justifications, and the fourth field counts exactly those additions - if arguments accumulate in the delay window, somebody was at the record during that time.
What it does not settle. It does not settle which mechanism did it, and it is not worth pretending that it does. Arguments accumulating fits an asymmetry in the selection of evidence, but it also fits a reluctance to realise a loss: someone who does not want to close a position in the red also writes down reasons, except that the arguments are then a consequence of the decision to wait rather than its cause. A platform export and your own notes contain the order of entries, not the causal order, so they do not separate those two runs. Of the four explanations that were in play for Run B, two therefore remain, and that is as much as this material yields.
The conclusion is thus narrower than the name one would like to attach, and I deliberately do not widen it. On the basis of their own record a trader can state that the process is being violated systematically - and that is enough to act on, because the corrective rule is the same for both mechanisms: a decidable condition, written down before entry, executed without a window for adding notes. What they cannot state is that the cause was confirmation bias. Anyone who wants to know the second of the two remaining possibilities from the inside will find it described in the article on the psychology of losses; separating the two is not something this article provides, because on the basis of a trading record alone they cannot be separated.
11. Summary and what this article does not settle
The table below has three working columns and none with costs. The column with the simpler explanation is there not for balance, but because in most cases it is the one that turns out to be right.
There are nine rows and seven mechanisms, and that is not an inconsistency. Overconfidence and the availability heuristic each leave two different traces in the record, and the table organises the material by trace rather than by the name of the mechanism - because when reviewing entries you start from what is visible, not from what it is called. The split rows carry a shared prefix, so their assignment to a single mechanism can be read off the table itself.
| Distortion | What is visible in the orders and the record | Simpler explanation of the same thing | Signal across a series of decisions |
|---|---|---|---|
| Confirmation bias | Justifications for holding a position are produced after entry; more sources checked after than before | The invalidation condition was loosely formulated and it really could not be stated whether it had occurred - or a reluctance to close the position at a loss was at work, producing the same interval and the same added notes (psychology of losses) | In the window between the condition occurring and the action, justifications systematically accumulate that were not in the pre-decision record. This signal separates a violation of the process from its absence, but does not separate the two mechanisms - neither the interval alone nor the accumulation of justifications does that (section 10) |
| Overconfidence: precision | A forecast in the form of a single value, no alternative scenario | The record format forces a single number and has no field for a range | Statements clustered in one band regardless of the situation |
| Overconfidence: statement diverging from accuracy | Stated confidence does not differentiate outcomes | The statement bands differ by setup, instrument or session time - different things are being compared | A flat share of hits across successive bands while the statement rises |
| Anchoring | Justifications for modifications refer to the entry price; protective orders at equal distances from it | The level genuinely has structural significance and coincides with the entry price | The same number comes back in the justifications for successive decisions on different instruments |
| Availability: recency | A rule changed shortly after a losing streak, with no reference to a condition written down earlier | Market conditions genuinely changed and the old observations stopped describing the present state | The date of the rule change systematically within a day of a loss |
| Availability: vividness of an event | Avoiding an instrument or a situation after one spectacular event | The event genuinely changed the conditions of trading that instrument | Withdrawals clustered around events reported in the media, not around events in your own record |
| Gambler's fallacy | Position size or the interval to the next decision changes with the length of the streak | The decisions were not independent - several orders were one bet in several disguises | A correlation between the length of the current streak and the deviation from the size written in the plan |
| Hindsight bias | A note saying "it was obvious" with no record made before the decision | The information genuinely appeared later and changed the picture then | The share of entries filled in after the position was closed |
| Blind spot | The names of biases appear in comments on other people's decisions and in your own summaries after a position is closed, not in the field filled in before the order | The pre-decision fields are simply left blank, regardless of content - no habit of recording, rather than an asymmetry in judgement | The share of entries in which any name was written down before the decision is close to zero while many names are written after it |
An exercise for the next ten decisions
For ten consecutive decisions, write one sentence before each: the piece of information that - if it appeared - would make you treat the hypothesis as refuted. Not a generality, but something about which it can be stated whether it occurred. After the tenth decision count two numbers: in how many cases that information appeared, and in how many of those you did anything about it. That is all. The exercise does not measure whether you are prone to confirmation bias, and it promises nothing about results - it checks only whether your invalidation conditions are formulated in a way that lets them occur at all.
What this article does not settle
This article does not establish which mechanism is operating in a particular person. In the review proposed here, repeated traces in the record may support a hypothesis that still needs comparison with alternative explanations. This is not the only way to study cognitive biases. You can count decisions outside the plan’s conditions and check whether their frequency falls after a procedural change; the change should not be assumed to have worked. Financial results can also be measured, but improvement does not itself establish a reduction in bias and depends in part on the economic basis of the rule.
What remains is the question that opens the next stage of this material: if the outcome of a single trade does not tell you whether the decision was good, then what does it tell you - and by what criteria should a process be judged once the result is already known. That is dealt with by the article on outcome and the decision process.
FAQ - Cognitive biases in trading
Can cognitive biases be removed?
How do you tell a bias from an ordinary mistake or noise?
Is overconfidence the same thing as self-confidence?
Does the "hot hand" exist?
How many entries are needed to see your own calibration?
Is holding a losing position always a distortion?
Sources and references
- Tversky A., Kahneman D., "Judgment under Uncertainty: Heuristics and Biases," Science, 185(4157), 1974, pp. 1124-1131. Anchoring and adjustment; a laboratory experiment with a wheel of fortune, hypothetical stake. Effect size: median estimates of 25% and 45% at anchors of 10 and 65 respectively, with the anchor openly random in origin and participants fully aware of that origin.
- Tversky A., Kahneman D., "Availability: A heuristic for judging frequency and probability," Cognitive Psychology, 5(2), 1973, pp. 207-232. A series of laboratory experiments with a hypothetical stake; population: participants in university studies; dependent variable: judged frequency of events. The most frequently cited experiment from this work concerns letter positions: participants judged whether a typical English text contains more words beginning with a given consonant or more words with it in third position, and most pointed to the first position - including for letters that in fact occur more often in third place. Direction of the effect: systematic overstatement of the frequency of that class of events whose examples come to mind more easily.
- Wason P.C., "On the failure to eliminate hypotheses in a conceptual task," Quarterly Journal of Experimental Psychology, 12(3), 1960, pp. 129-140. The 2-4-6 sequence task; some participants tested their hypothesis only with examples consistent with it. A laboratory study on students, without market transactions. Direction of the effect: most participants announced the rule without making a single attempt capable of refuting it. N = 29 psychology students, checked in the original paper, pp. 131–133. Participant compensation was not established.
- Nickerson R.S., "Confirmation Bias: A Ubiquitous Phenomenon in Many Guises," Review of General Psychology, 2(2), 1998, pp. 175-220. A literature review; separation of selective search, interpretation and recall.
- Moore D.A., Healy P.J., "The trouble with overconfidence," Psychological Review, 115(2), 2008, pp. 502-517. A review together with the authors' own studies; separation of three forms of overconfidence and a demonstration that they behave differently on hard and easy tasks. Original study: N = 82 Carnegie Mellon students, quizzes of varying difficulty with real payment for performance and estimates. Full text, pp. 507–508.
- Barber B.M., Odean T., "Trading Is Hazardous to Your Wealth: The Common Stock Investment Performance of Individual Investors," The Journal of Finance, 55(2), 2000, pp. 773-806. 66,465 households, an American discount brokerage, common stocks, 1991-1996; highest-turnover households about 11.4% a year net, the average household about 16.4%, the least active about 18.5%, the market about 17.9%. Overconfidence is the authors' interpretation in that work, not a measured variable.
- Northcraft G.B., Neale M.A., "Experts, amateurs, and real estate: An anchoring-and-adjustment perspective on property pricing decisions," Organizational Behavior and Human Decision Processes, 39(1), 1987, pp. 84-97. Two groups of participants - professional estate agents and students - valued the same real house after inspecting it; the only manipulated piece of information was the listing price. The valuations of both groups moved with it, and the agents for the most part did not name it among the premises of their valuation.
- Fischhoff B., "Hindsight ≠ foresight: The effect of outcome knowledge on judgment under uncertainty," Journal of Experimental Psychology: Human Perception and Performance, 1(3), 1975, pp. 288-299. Population: students; a laboratory task, hypothetical stake. Participants estimated the probability of four possible endings of a described historical event; being told a reported ending raised the probability assigned to it despite instructions to judge without that knowledge. Experiment 1 used four student samples (100, 80, 87 and 100 participants); some assessed multiple events. Original full text, pp. 289–290.
- Gilovich T., Vallone R., Tversky A., "The hot hand in basketball: On the misperception of random sequences," Cognitive Psychology, 17(3), 1985, pp. 295-314. The work combines a survey of fans, an analysis of field-goal and free-throw records of professional teams, and a controlled shooting experiment with college players. Direction of the result: in none of these analyses was a positive relationship found between a hit and the next hit - hence the classic challenge to belief in a hot streak.
- Miller J.B., Sanjurjo A., "Surprised by the Hot Hand Fallacy? A Truth in the Law of Small Numbers," Econometrica, 86(6), 2018, pp. 2019-2047. A demonstration of bias in the estimation method used in the 1985 work and a re-analysis of the same data. The reversal of the conclusion concerns the controlled experiment from that work, not the game data.
- Pronin E., Lin D.Y., Ross L., "The Bias Blind Spot: Perceptions of Bias in Self Versus Others," Personality and Social Psychology Bulletin, 28(3), 2002, pp. 369-381. Studies involving students and airport travellers; participants assigned themselves a lower susceptibility to cognitive biases than to others, including after being shown a description of the mechanism.
- European Securities and Markets Authority (ESMA), decision on a temporary product intervention concerning contracts for difference, 2018, together with subsequent extensions and national provisions introducing the obligation to disclose the share of losing retail accounts. The 74-89% range comes from materials of that period and concerns defined providers and periods; it is neither a universal figure nor an explanation of the causes of losses.
- Klein G., "Performing a Project Premortem," Harvard Business Review, September 2007. Practitioner literature; the source of the technique, not evidence of its effectiveness.
- Croson R., Sundali J., "The Gambler's Fallacy and the Hot Hand: Empirical Data from Casinos," Journal of Risk and Uncertainty, 30(3), 2005, pp. 195-209. An analysis of recordings of roulette play at a casino in Reno; the share of bets on the opposite colour rose with the length of a run of one colour. Real stakes, known distribution, independent trials. The observation covered 139 players and 24,131 bets over 18 hours. Original paper, pp. 196 and 200–201.
- FCA: Occasional Paper on behavioural economics and retail investor mistakes.
- OECD: International financial consumer protection and behavioural insights research.
Sample sizes and tasks are reported where checked in the original papers: for example, Wason — 29 students; Moore and Healy — 82 students; Barber and Odean — 66,465 households; Croson and Sundali — 139 players. Fischhoff’s entry distinguishes samples from repeated assessments. The records are not equally complete: compensation was not verified for [3], [7], [9] and [11], nor the population in the wheel-of-fortune example in [1]. These gaps must not be replaced with “a series of experiments” or an assumed N. The calculations were performed independently: Wilson intervals for hit proportions assuming independence; the losing streaks from 10.1 by a Markov chain, checked with 400,000 repetitions at seed 20260810 (the NumPy check differing by up to 0.15 percentage points); and 5/12 by enumerating all eight three-toss sequences. Illustrative values show the mechanism and do not come from a trader’s account.