Trade outcome versus the quality of the decision process — how to judge decisions independently of the result
A trade that follows the plan can end in a loss, while an entry made without checking the conditions can turn a profit. In this article's illustrative examples, EUR/USD turns five pips short of the target, while another position, opened against the rules five minutes before a data release, earns USD 300. If we change the plan after the first and approve the second, we are judging decisions solely by their results. We will show how to separate execution compliance from financial outcome, assess the process against prior criteria and recognise when evaluating the rules requires a much longer series of observations.

Hide the result long enough to apply the prior criteria
Suppose two reviewers receive the same order record, but only one sees the profit. If their process verdict changes solely because of that extra column, they are no longer assessing the same information set. A simple review exercise is to classify the recorded conditions first, save that verdict and then reveal the result. This is a way to expose a change in judgement, not a proven cure for outcome bias.
Our invented protocol requires a checklist recorded before submission, a volume no greater than 0.20 lot and no active entry ban. All three are mandatory; one passed condition cannot compensate for another that failed. The four decisions below are fictional, and their USD results are already net of all assumed costs. No other cash adjustments occur. We assess compliance with these rules, without assuming that the rules themselves identify a profitable strategy.
Keep an unresolved record separate from a known failure
| Decision | Prior checklist record | Volume | Active entry ban | Net result |
|---|---|---|---|---|
| A | Complete | 0.15 lot | No | −18 USD |
| B | Complete | 0.25 lot | No | +42 USD |
| C | Missing evidence | 0.10 lot | No | +11 USD |
| D | Complete | 0.20 lot | No | +9 USD |
A and D have confirmed passes for all stated conditions. Equality with the volume ceiling is permitted, so D passes. B has a confirmed volume breach. C passes volume and the ban check, but the missing checklist evidence leaves its overall status unresolved. Do not convert it into either a confirmed pass or a proved omission merely to complete a two-category chart.
Averaging the three conditions could give B a seemingly favourable score while hiding a breach of a mandatory condition. That is why this protocol uses a gate, not a compensating average. A different scoring design would need to say which criteria are optional and why. The classifications above remain the same when the result column is covered.
Reveal the money and retain both kinds of record
The full net total is −18+42+11+9=44 USD. The two confirmed-compliant decisions together lost 9 USD, the known breach earned 42 USD, and the unresolved case earned 11 USD. Those sums do not show that breaking the rule improves performance. The groups contain different decisions, and no comparison design or representative sample was supplied.
The result also does not resolve C’s missing evidence. A profitable unknown remains unknown until the relevant record is recovered. Baron and Hershey’s original outcome-bias experiments examined how knowledge of an ending changed evaluations in hypothetical medical and gambling scenarios. They did not validate this trading protocol. The reason for keeping the columns separate is to make clear which evidence supports a compliance finding and which describes the financial result.
Name the denominator before reporting improvement
Of the 4 decisions, 3 are assessable as a confirmed pass or failure, giving evidence coverage of 3/4=75%. Among those 3 assessable decisions, 2 pass, giving 2/3≈66.67% confirmed compliance. Across all 4 decisions, the share with a confirmed pass is 2/4=50%. These are different summaries of the same record, not competing answers to one question.
Report the unresolved count beside any rate. Dropping C without disclosure makes it easier to mistake a change in documentation for a change in behaviour. Keep the original verdict, the later evidence and any revised status with their dates. Compliance review can identify an execution issue while leaving rule quality open; evaluating the rule needs a separate analysis of opportunities, outcomes, costs and alternatives. The detailed lesson develops that separation. A loss should not automatically reverse a documented pass, and a profit should not automatically erase a known breach.
Three exercises with worked answers
The exercise data are hypothetical. Choose an answer, then check the calculation.
Continue to the two separate review loops
The detailed lesson preserves the illustrations and research on outcome and hindsight bias. The introductory record separates compliance, missing evidence and net results before assessing the broader rules.
1. Two trades, four possible verdicts
Start with the first one. You open a long on EUR/USD at 1.0820, after a bounce off a demand zone described in your catalogue of setups. Stop loss at 1.0780, target at 1.0900 — both placed before the click, risk at the level written into the plan beforehand. Price goes north, reaches 1.0895 and turns. Five pips short of the target. It comes back to 1.0780 and you are stopped out.
Was that a bad decision?
No. It was a decision taken according to a condition that was in force before entry, with risk limited in a way planned in advance. The fact that it fell five pips short of the target changes nothing about the information you held when you clicked. It changes only the result — and the result of a single trade is one draw from a distribution, not a verdict on the accuracy of the choice.
The second trade from the same week. Illustrative example: Friday at 14:25 in platform time, five minutes before a labour-market release scheduled for 14:30. You do not check the calendar. You go long GBP/USD on a candlestick pattern from the fifteen-minute chart, with no protective order, in a size larger than your plan allows. The data comes out weak for the dollar, sterling rises sixty pips, you close with three hundred dollars of profit.
Was that a good decision?
Also no. The entry condition was never checked, the position size did not follow from a rule, and a position without a protective order has no defined risk — you cannot say how much you could have lost, because no level existed at which you would exit. Had the data come out the other way, the only limit on the loss would have been the moment you managed to react by hand. The profit proves nothing here beyond the fact that the distribution has two sides and this time you landed on the favourable one.
| Quadrant | Illustrative example | Verdict and its justification |
|---|---|---|
| 1. Compliant process, positive outcome | Entry condition met and recorded before the click, exit levels placed, target reached without interfering with the position | Positive. The result does not raise it — it would be identical had price turned five pips short of the target |
| 2. Compliant process, negative outcome | The same trade, except that price turns short of the target and exits at the stop | Positive, unchanged. This is the quadrant where the decision to change the rules is most often taken — usually with no grounds, because a single observation says nothing about them |
| 3. Breached process, positive outcome | Entry without checking the calendar, position larger than the plan, no protective order — a profit after the data release | Negative. The result may encourage repetition of the breach, but profit does not establish safety; the missing condition and protection remain visible in the record |
| 4. Breached process, negative outcome | Entry after the move, without a condition from the catalogue, with the stop moved while the position was open | Negative. The only quadrant where both verdicts agree — and precisely for that reason the least instructive |
Those two trades sit in quadrants 2 and 3 of this matrix. Quadrant 4 is recognised without difficulty: you broke a rule and you paid, the conclusion writes itself. Quadrant 1 causes no trouble either. The problem is the other two — because there the verdict issued by the result is the opposite of the verdict issued by the process, and the result is visible immediately, whereas the process had to be described earlier for you to be able to return to it.
Quadrant 3 deserves particular attention during review: a breach is accompanied by a profit. That result may obscure the departure and encourage repetition, but the example does not establish that this will happen in a particular trader. The profit is immediate and visible, whereas the consequences of omitting a protective order cannot be assessed from this one trade. If the behaviour repeats, examine its premises and execution rather than assume carelessness or a particular learning mechanism. The favourable outcome is real; it does not establish the quality of the process.
The opposite overcorrection is equally expensive. The sentence "the result does not matter" is false: the result is information, only about something other than it appears to be, and it requires a different number of observations. We return to this in section 8, where I show a situation in which the result says a great deal about the process.
2. Outcome bias — judging a decision by its result
The phenomenon has a name and it has a study. Outcome bias is the shift in how the quality of a decision is judged under the influence of information about its result, with the information available at the moment of choice held constant. The point is not that we know more after the fact — the point is that the same decision, described identically, receives a different score simply because a different ending has been appended to it.
It is worth noting what this definition does not cover, because trading texts push three different things under the same name.
This is not inference from a small sample. Judging a set of rules after three trades instead of two hundred is a problem of statistical power, described as the law of small numbers by Amos Tversky and Daniel Kahneman in 1971[4] — the respondents there were professional researchers, and they still systematically overestimated the repeatability of results obtained on small samples. The mechanism is covered by article 10.1. You can commit both errors at once and usually do, but the countermeasures differ: for a small sample, waiting before concluding helps; for outcome bias, waiting does not help at all.
This is not a rational updating of belief. If after a trade you learn something that could not have been known earlier — that the instrument behaves differently at this hour of the day, that your entry condition was defined ambiguously — the update is correct. The criterion is simple: does the new information concern the rule, or only the result.
This is not a recency effect. Overweighting the most recent observations (recency bias) concerns which data you take into account. Outcome bias concerns the criterion by which you judge them. The catalogue of the cognitive biases themselves is kept in article 10.3.
In practitioner literature the same error is sometimes called resulting — a term popularised by Annie Duke, a poker player and the author of a book on decisions under uncertainty[7]. The name is convenient and it caught on, but the register has to be flagged: this is a concept from a handbook, not the finding of a study. The popularity of the term proves nothing — the proof is the work of Baron and Hershey, not the number of people repeating a slogan.
3. Hindsight bias — why yesterday's chart is a bad teacher
Outcome bias has a younger brother that spoils learning from history even more effectively. Hindsight bias is the systematic inflation, once the result is known, of the belief that the result was predictable — including a false memory of having predicted it beforehand.
The manifestation in trading is banal and therefore dangerous. You open last week's chart. You see the breakout, the pullback, the continuation. Everything is in place, every level exactly where it should be. The difficulty is that you are looking at a chart whose right edge no longer exists. On Thursday at 11:40 that same chart ended in the middle of a candle, two other scenarios stood beside it, and the entry condition was either met or not — and which of those was visible at the time cannot be reconstructed from memory after the fact, because memory has been overwritten by the ending.
This is where the sentence that shows up most often in a weekly review comes from: "it was obvious, after all". It was not. It was one of several scenarios that looked comparable at that moment, and it became obvious only once the right-hand side of the chart had been drawn in.
Hindsight bias should not be confused with ordinary knowledge acquired later. If after two hundred trades you notice that your entry condition behaves differently in the first hour of the session, that is a conclusion from data, not an illusion — provided you do not claim you knew it before collecting the data. The line runs exactly there: between "now I know" and "I knew".
A practical safeguard is to record the information and the condition before the outcome is known. This does not guarantee impartiality, but it provides a reference independent of later recollection. Compare it with the execution history rather than rely solely on memory or a promise to be more honest in the next review.
4. Self-attribution — the profit is mine, the loss is the market's
The third mechanism does not distort the judgement of a single decision, but the reckoning of a whole series. Self-attribution is an asymmetry in assigning causes: successes are explained by internal factors (skill, preparation, feel), failures by external ones (the market, the broker, the news, bad luck).
The consequence for a trader is arithmetic, not moral. If profits go to the "skill" account and losses to the "conditions" account, then after a few dozen trades the picture of your own process is skewed in one direction regardless of what the process actually looked like. The gap between declared and actual accuracy grows with it — what article 10.3 describes as poor calibration, as opposed to plain overconfidence.
Beware the other side of the same coin: the sentence "take every loss on yourself" is not a countermeasure, only an inversion of the error. Some losses really do come from conditions you had no influence over, and attributing them to yourself distorts the reckoning just as badly. The countermeasure is not severity towards yourself, but a criterion fixed earlier than the event it is meant to judge.
5. The pre-decision record — a reference for assessment
The mechanisms described can influence evaluation once the outcome is known. That is why part of the journal should be written before the order: it preserves the information available when the decision was made. A post-close entry is also needed, but it cannot replace the prior plan. Separating these records helps assess execution without relying solely on later reconstruction.
A record of a decision has three time layers: what was created before the click, what was written down while the position was open, and what was appended after the result became known. The full construction of a journal built on this split — the complete set of fields, the breach codes and the review protocol — is described in article 10.8; here only one thing is needed from it.
The first layer precedes the outcome, so it should be retained without later alteration. Record new information or a revised plan separately, with the time and the reason. The same information reconstructed only after a position closes may already contain confidence or a rationale fitted to the ending. That need not be deliberate falsification; memory is susceptible to reconstruction. The earlier record allows discrepancies to be identified, although it does not remove every source of judgement error.
A separate note about scales. The degree of conviction serves to compare your own entries against each other, not to measure anything — the limitations of self-report scales, and the apparent precision they introduce, are discussed in article 10.8.
6. The process scorecard — criteria checkable without knowing the result
To limit judgement of the process by its outcome, define criteria that a reviewer can assess without knowing the profit or loss. If an answer requires the result column, the criterion also concerns the outcome. This is a useful test for a compliance scorecard; it does not guarantee complete impartiality or replace a separate assessment of the rules.
Five questions that pass this test:
- Was the entry condition met before entry — according to the description in the catalogue, and not according to a justification appended later?
- Did the position size follow from a rule in force before the session? The rule itself and its level come from section 9 — here you check only whether it was applied.
- Were the exit levels placed before entry, and did any later changes comply with the plan?
- Was the decision taken inside a time window in which you are allowed to trade under your own plan?
- Were the premises documented before the click?
Each of these questions has a "yes" or "no" answer and each can be checked against the order history and against the first layer of the record. None requires knowing the result. That is the difference between a process scorecard and a subjective mark of "that trade was fairly good".
There is one more way of evaluating a process that looks sensible and is a trap. It is tempting, after fifty trades, to split them into compliant and non-compliant, compare the results of the two groups and treat the difference as proof that the process works. Such a comparison describes an association between compliance and outcome, rather than proving an effect of the process. It is not necessarily outcome bias when compliance criteria were fixed in advance; it does require attention to differences between the groups and cautious inference.
Breaches may coincide with losing streaks, the end of a session, larger positions or particular market conditions. Compliant and non-compliant groups may therefore differ from the moment they are formed. The direction also needs checking: earlier losses may affect later actions, breaches may affect subsequent outcomes, and both may have common causes. Splitting fifty observations into two groups cannot by itself distinguish these possibilities. Use criteria defined beforehand and account for differences in conditions.
The process metric that makes no appeal to the result is simply the frequency of breaches and its distribution over time: how many decisions out of ten carried a full set of five "yes" answers, how that number changes from week to week, after which events it drops. These are questions whose answers require no look at the result. The construction of the journal itself, the breach codes and the review protocol are described in article 10.8; designing the procedures meant to limit those breaches — article 10.7.

7. How many observations before the result means anything
Since the result is information, the question remains how much of it is needed. The answer is uncomfortable and it is worth seeing it calculated rather than hearing it as advice about patience.
Take the simplest quantity a trader observes: the share of trades that end in a profit, that is the win rate. Assume that successive trades are independent and that the true win rate is constant. Both assumptions are violated in practice and I return to that below; under such assumptions, however, the arithmetic is at its most optimistic, so the conclusions are all the stronger.
The standard error of the observed win rate is:
SE = √( p × (1 − p) / n )
For a true win rate of 50%, the normal approximation gives the following ranges covering approximately 95% of observed proportions. They are rounded; an actual proportion in a finite sample takes discrete values:
| Number of trades | Standard error | Approximate win-rate range (about 95%) |
|---|---|---|
| 20 | 11.2 pp | from 28% to 72% |
| 50 | 7.1 pp | from 36% to 64% |
| 100 | 5.0 pp | from 40% to 60% |
| 200 | 3.5 pp | from 43% to 57% |
| 400 | 2.5 pp | from 45% to 55% |
| 1000 | 1.6 pp | from 47% to 53% |
Reading this table is simple. A set of rules with a true win rate of 50% can, after twenty trades, produce a reading in the broad approximate range of 28–72%; possible readings are multiples of 5 percentage points. After a hundred, a reading of 41% and a reading of 59% are both still within the norm — and that is precisely the range in which a trader usually decides to change something.
The practical question, however, is a different one: how many trades does it take to tell a 55% win rate from a 50% one? For a one-proportion test, at a significance level of 0.05 and 80% power, it works out at roughly 620 trades in the one-sided version and roughly 780 in the two-sided one. A simulation check for a one-sided normal-approximation proportion test (α = 0.05), so that the formula is not the only argument here: drawing 200,000 series of trades with a true win rate of 55% (seed 20260810), the difference is detected in 24% of runs at 100 trades, in 64% at 400 and in 88% at 800.
It is worth recalling here the calculation from article 10.1: at a 50% win rate, a run of five losses in a pre-specified sequence of trades has a probability of (0.5)⁵, that is about 3%, while the probability of such a run occurring anywhere within 100 independent trades is about 81%. A losing streak alone therefore does not establish failure — it can also occur under unchanged rules.
An analogy that puts this in order: an actuary at an insurance company does not change the tariff after one large payout. He changes it when the loss ratio departs from the assumption across the whole book of policies. He is not braver than the trader — he simply has a sample of the right size from day one. After a month the trader has a dozen or so observations and the same need to draw a conclusion.
8. Two loops: judging execution and judging the rules
Execution and the set of rules need separate reviews, with different criteria, frequencies and sample sizes. Confusing them can lead to settings being changed after a few losses without checking whether the rule’s basis has changed. The distinction organises the analysis, but it does not automatically resolve every execution or strategy problem.
The execution loop runs weekly and needs no results: in it you check the five criteria from section 6 and count how many decisions carried a full set of "yes" answers. The rules loop runs on hundreds of trades, requires a series of results and statistics, and its tools — edge assessment, historical testing, simulation — belong to section 8. The second loop cannot be sped up by discipline, just as a missing edge cannot be repaired by better execution.
Separating the reviews does not mean always waiting for a large sample when new information arrives. The following counterexample shows the limit of that principle.
Separate two questions: “What was the financial outcome?” and “Did I execute the plan according to the agreed criteria?” The plan and entry conditions can be assessed before the position opens. The full execution, including later modifications and the exit, can only be assessed after those actions, by comparing the earlier record with the order history without using profit or loss as the criterion for compliance. A separate review of a longer series assesses the economic basis of the rule.
The same distinction becomes much harder to maintain when successive decisions stop being a response to the market and become a response to the previous loss. That is the subject of the article on revenge trading and tilt (10.5): what observably changes in behaviour before the sequence goes too far.
FAQ — trade outcome and decision quality
Can I judge my rules after twenty trades?
How does outcome bias differ from hindsight bias?
I made money by breaking my own rule. What should I do with that?
I issue the process verdict on myself. Isn't that useless?
Does this mean the result is irrelevant?
I trade fast, I have no time to write anything before entry.
Sources and references
- Baron J., Hershey J.C., "Outcome bias in decision evaluation", Journal of Personality and Social Psychology, 54(4), 1988, pp. 569–579. The evaluation of the same decision shifts with its outcome; medical and gambling scenarios, hypothetical stakes.
- 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. Knowing the ending raises the retrospective estimate of its probability. Experiment 1 used four student samples (100, 80, 87 and 100 participants); some assessed multiple events. Original full text, pp. 289–290.
- Guilbault R.L., Bryant F.B., Brockway J.H., Posavac E.J., "A meta-analysis of research on hindsight bias", Basic and Applied Social Psychology, 26(2–3), 2004, pp. 103–117. A meta-analysis of more than ninety studies; the effect confirmed, the average size small.
- Tversky A., Kahneman D., "Belief in the law of small numbers", Psychological Bulletin, 76(2), 1971, pp. 105–110. Systematic overestimation of what can be inferred from a small sample.
- Mezulis A.H., Abramson L.Y., Hyde J.S., Hankin B.L., "Is there a universal positivity bias in attributions?", Psychological Bulletin, 130(5), 2004, pp. 711–747. A meta-analysis of 266 studies of the asymmetry in attributing causes to successes and failures.
- Miller D.T., Ross M., "Self-serving biases in the attribution of causality: Fact or fiction?", Psychological Bulletin, 82(2), 1975, pp. 213–225. A critique of the motivational explanation of attribution asymmetry.
- Duke A., Thinking in Bets, Portfolio/Penguin, 2018. The source of the term "resulting" — practitioner literature, not research.