The trading journal – two records of the same decision, and what each of them settles
A note after a position closes shows what happened. To check the basis of the decision and whether execution followed the plan, a record made before the order is also needed. These two kinds of information complement each other, but should not share one field: knowing the outcome can change a later justification. This article compares two records of the same trade, organises data into three time layers, and shows how to calculate completeness and documented compliance. We also examine the limits of inference: a missing entry, a short post-loss interval or a high self-rating does not alone establish an emotional cause.

Record the decision checkpoints, including the decisions to do nothing
A list of orders can show what was requested. It cannot, by itself, show which required actions were missed or which unsuitable cases were correctly left alone. To ask those questions, define the observation set before inspecting the outcomes. Otherwise the journal selects its own evidence by recording only the occasions when something happened.
Our hypothetical replay contains 20 decision checkpoints selected beforehand. Its written exercise rule requires an entry request when the recorded condition is met and a skip otherwise. This binary instruction belongs only to this replay; it is not a recommendation to trade every signal in a live account. The recorded action is a request or a skip, not proof of a completed fill. Earlier evidence permits assessment at 18 checkpoints. At 2, the required pre-decision condition data are missing.
Keep the condition and the action in separate columns
| Condition before action | Checkpoints | Entry requests | Skips |
|---|---|---|---|
| Met | 6 | 4 | 2 |
| Not met | 12 | 2 | 10 |
| Unknown from available evidence | 2 | 1 | 1 |
The table contains 7 entry requests and 13 skips, for 20 actions in total. Among known eligible cases, 4 requests follow the exercise rule and 2 skips do not. Among known ineligible cases, 10 skips follow it and 2 requests do not. There are therefore 14 documented compliant actions and 4 documented deviations across the 18 classifiable checkpoints.
The last row remains unknown. An entry with missing condition data is not automatically a violation, and a skip with missing data is not automatically compliant. The action is known; its relation to the rule is unresolved. Keep that distinction visible rather than turning “not assessable” into whichever category improves the monthly headline.
Name the denominator before displaying a percentage
Assessment coverage for this particular check is 18/20=90%. Documented compliance among classifiable checkpoints is 14/18≈77.78%. It is not the same as a proved compliance rate across every checkpoint. With the 2 unknown cases unresolved, actual compliance over the whole exercise could lie between 14/20=70% and 16/20=80%. These are logical bounds for the recorded sample, not a confidence interval or a forecast.
An entry-only review of the classifiable requests would show 4/6≈66.67%. That answers a narrower question and misses both the correctly rejected cases and the required actions that were skipped. There is no contradiction between the percentages. Their populations differ. None measures profitability, and a perfectly followed rule can still be a poor rule. The exercise supplies no returns, execution prices or evidence of a trading advantage.
Correct the record without pretending the correction existed beforehand
If original pre-decision data are later recovered, append them with their source and recovery time, then recalculate the relevant version of the assessment. A later recollection is a different kind of evidence. Do not overwrite it into a field labelled “recorded before action”, or silently replace the earlier reported denominator. Preserve the original row and the reason for the revision.
A practical review can start with the unknown and deviating rows: which field prevents classification, which written condition was not followed, and what other records could settle the issue? These are questions about the record and the procedure. They do not identify an emotional cause or show that a change will improve returns. Keep decision evidence, execution evidence and retrospective interpretation distinguishable. The detailed lesson examines hindsight, self-reports, deviation codes and review cadence; this small matrix first makes the missing non-trades visible without calling every absence an error.
Three exercises with worked answers
The exercise data are hypothetical. Choose an answer, then check the calculation.
Continue to evidence layers and journal reviews
The full lesson separates entries made before a decision, during execution and after the outcome. It develops deviation codes, completeness and process metrics, self-report limitations, selection of examples and review protocols without promising that documentation creates an edge.
1. The same trade, two records
Below is one decision described twice. The values are illustrative – they show how the fields are laid out, not data from anybody's account. Both records concern the same order, both were produced by the same person and both took her roughly the same amount of time. What differs is the moment they came into being and what the fields are about.
Clock convention for the illustrations. All clock-time examples in this lesson use 12 March 2026 and one synchronised model clock in UTC (UTC+00:00). The dated table entries identify the instants; shorter time references in the prose use that same date and clock. Separate illustrations remain independent cases, not one combined session.
For your own record. Save the full calendar date and time as a UTC instant, retaining the original source timestamp, its clock identifier (an IANA zone or documented server offset) and the offset valid on that date. Verify synchronisation and timestamp meanings for the note and platform history before comparing them. Normalising a display to UTC does not correct clock drift; retain any known uncertainty, and leave an ordering unresolved if the clock difference could change it. For day and session grouping, record the chosen zone or fixed offset and the cutoff, apply dated daylight-saving rules where relevant, and group both sources under that same convention.
12 March 2026, GBP/USD, long position, outcome: protective order hit. Comment: I entered too early, I ran out of patience, I need to work on my discipline.
Record B is the same trade logged as it went along. The first group of fields was created before the order was placed; the second during the life of the position; the third after it was closed. The times are not decoration here – they are the only thing that separates a description of the decision from a description of how it is being interpreted.
| UTC timestamp | Layer | Recorded content |
|---|---|---|
| 2026-03-12T09:41:00Z | before the decision | Condition from the catalogue: reaction to the 1.2640–1.2655 zone after an earlier breakout. Condition met |
| 2026-03-12T09:41:00Z | before the decision | Hypothesis: the zone holds, move towards 1.2740 |
| 2026-03-12T09:41:00Z | before the decision | Degree of conviction: 6 out of 10 |
| 2026-03-12T09:41:00Z | before the decision | Contradicting information: direction on the daily interval opposite to the one assumed |
| 2026-03-12T09:41:00Z | before the decision | Exit plan: invalidation 1.2608, target 1.2740, target not to be moved |
| 2026-03-12T09:41:00Z | before the decision | Context: second decision of the day, the previous one closed at a loss 47 minutes earlier |
| 2026-03-12T09:44:00Z | execution | Entry 1.2648, protective order placed before entry |
| 2026-03-12T11:26:00Z | execution | Target moved from 1.2740 to 1.2691. No recorded rationale – code N4 |
| 2026-03-12T13:10:00Z | after the outcome | Protective order hit |
| 2026-03-12T13:20:00Z | after the outcome | Comment: entry condition and invalidation level unchanged; the only deviation is the target move at 11:26 |
Now the most important part: the set of questions somebody will put to this data three months from now, and which of them each record answers.
| Question put to the data | Record A | Record B |
|---|---|---|
| Was the entry condition met before entry, or added afterwards? | cannot be settled | settled by the 09:41 entry |
| Was the exit plan changed while the position was open? | unknown | yes, at 11:26 |
| How much time had passed since the previous loss? | unknown | 47 minutes to the decision record, 50 to entry |
| How many times this month did the same deviation occur? | cannot be counted | the number of entries carrying code N4 |
| Does "I ran out of patience" recur in other entries? | it is not a field, it is a sentence | it does not occur – there is an event with a time |
| Were the entry and the execution compliant with the rule? | a question with no answer | entry yes, execution with one deviation |
| How many decisions were taken this month and what was their combined result? | settles it | settles it |
The last row is there so that the table does not pretend to be something it is not. The set of questions was chosen against a single criterion – whether they can be answered without looking at the outcome – and with that choice record A loses in advance, because almost nothing it contains was created before the outcome was known. It does not lose everything, though: the number of decisions, their distribution over time and the aggregate result it settles just as well, and sometimes better – it is shorter, so it is more often actually kept. The selection is therefore deliberate, and I am stating it openly here, because choosing questions to fit a conclusion known in advance is exactly the operation that section 9 of this article names and warns against. What would be a stretch is only a set presented as the complete list of questions worth putting to a journal.
The difference is not simply one of volume. Record B can use six fields, while record A fits into two lines and does not answer the first six questions in this table. Completion time depends on the tool and condition; it is not a measured result in this example. Three features matter more: recording time, what the fields describe and the coding method.
First: the moment. The six fields timestamped 09:41 were created before the outcome existed, so the outcome could not have shaped them. The same set of information written down at 13:20 is already something else, however completely honest the author may be.
Second: what the fields are about. In record B the fields describe events – a level, a time, a change to an order, a condition met or not met. Each of them can be settled by somebody who does not know the outcome. In record A the only field apart from the result contains a judgement, and a judgement can be compared with nothing except another judgement.
Third: coding. "I moved the target for no reason" written out in words is one sentence in one entry. The same event tagged with code N4 is a row in a summary that will tell you, after two months, whether it happens once in twenty decisions or once in three.
"I ran out of patience" is an example of a causal hypothesis added after the outcome was known. It may start a review, but without a behavioural definition it does not establish what changed. Compare it with the condition, decision time and execution: was a signal missing, was the time window breached, or did the action follow the plan? The outcome alone does not choose an answer. Keep the comment field separate from fields that let such events be counted across a series.
2. Why an entry written after the outcome is a different kind of data
The popular justification is "memory lies", and it is both discourteous and imprecise. Closer to the truth is the statement that a recollection is not a replay of a recording but a reconstruction assembled at the moment of remembering – out of what is left and out of what is known now. When "now" contains the outcome of the trade, the outcome enters the reconstruction as raw material.
The best documented part of this phenomenon is hindsight bias. In 1975 Baruch Fischhoff gave participants a description of a historical event and four possible endings, asking them to estimate the probability of each[1]. The group told what had actually happened assigned that ending a higher probability than the group that did not know it – despite an explicit instruction to judge as if the outcome were unknown. That is a laboratory study on students with a hypothetical stake, so I do not carry the effect size over to trading. What I do carry over is the structure: knowing the ending changes the assessment of what was foreseeable beforehand, and it does so even when a person is trying not to use it. The definition of this distortion belongs to the article on cognitive traps, and its role in evaluating trades to the article on outcome and process. What interests me here is one design consequence: a field filled in after the position is closed measures something different from the same field filled in earlier, so the two must never be kept in one column.
The second limitation concerns the question "why did I do that". In 1977 Richard Nisbett and Timothy Wilson collected a series of experiments in which participants stated the reasons for their own choices, and those reasons systematically diverged from the factor the researchers had in fact manipulated[2]. Their conclusion was more cautious than the way it is quoted today: it is not so much that people have no access to their own processes as that, when required, they supply a plausible explanation – one that sounds sensible in the given situation. The paper has serious methodological critics; in 1980 K. Anders Ericsson and Herbert Simon showed that reports given during the performance of a task, concerning the content currently being processed, can be reliable data – the problem starts with retrospection and with the question about causes[3]. That distinction is crucial for a journal, because it says outright where to put the box.
The third reason a record beats memory is the most literal one: some things are visible only in the complete set of trades, and nobody carries a complete set in their head. In 1998 Terrance Odean showed, on the accounts of a US discount brokerage, that investors realised gains more readily than losses[4]. That is the equity market three decades ago, so carrying it straight across to the retail currency market would be an extrapolation; the full description of that study, the effect size and the ban on deriving counterfactual statements from it are given by the article on the psychology of losses, which owns this whole topic. For a journal something else matters: this pattern did not come out of a survey or out of anybody's recollections. It came out of an order register, because only a register covers every decision, not just the ones that stuck in memory.
3. Three time layers and the moment an entry closes
The whole construction comes down to one decision: which fields are created before the order is placed, which during the life of the position, and which after it is closed. The layers do not differ in importance. They differ in what can be done with them.
The rule that holds this whole construction together is a single one, and it is stated harshly: layer one closes the moment the order is placed and is not open to editing. Not "should not be changed". Is not open to it. An entry timestamped 09:41 to which an explanatory sentence was added at 13:20 still carries the 09:41 stamp and looks like a record made before the decision, although it is not – and that is worse than no entry at all, because a missing entry at least announces itself. In a spreadsheet a separate tab and a habit are enough; in a tool with version history it is enough not to switch it off.
The second conclusion is practical. Layer one should be as short as possible because it is created when the entry condition occurs. A long form may be skipped, but no evidence establishes a fixed field count or week after which everyone abandons it. Check completeness and the time needed to create your own record. Minimalism should limit that cost without removing information required to assess a decision.
Trigger condition: the end of the trading week, a fixed time, the same one every week.
Action: export the order history from the platform and compute two numbers – how many positions were opened and for how many of them a layer-one entry exists with a timestamp earlier than the opening. First reconcile the source clocks and compare dated UTC instants under the clock contract above; unresolved ordering stays unclassified. Without reading the content of the entries.
Completion criterion: both numbers written into the week's row. Nothing further is done in this step.
Audit trail: a column with the share of decisions that have an entry, week by week, so that comparison is possible.
Failure criterion (a requirement separate from the form): if gaps against the previously adopted target remain after four weeks, check recording time, form layout, automation and workflow. No increase at complete coverage is not a failure. Six fields are a proposed version for counting context and deviations; the three-field entry in article 10.4 answers narrower questions, but remains useful. Removing a field changes what can be analysed and does not itself remedy missing entries. You may move recording closer to order placement or link it to an existing action. Document the change before the next review window. If gaps remain after another month, consider another recording format; that does not diagnose insufficient willpower.
The status of every time threshold in this text. Four weeks, a month, one entry in ten, half the codes – these are workshop settings, chosen so that you can check them against your own record, not research findings. No study stands behind them and nobody has shown that these particular values are the right ones; their only merit is that they are short enough for you to live to see the result and long enough to gather a dozen or so observations. You may shift them for yourself on one condition: the shift is written down before the series, not during it – otherwise the threshold starts fitting itself to whatever comes out anyway.
The three-field pre-decision entry itself – what fields it consists of and how to trigger it – is described by article 10.4; that is also where the process assessment card lives, the list of criteria that can be settled without knowing the outcome. This text deals with what comes next: how to turn individual entries into a set that can be counted after two months.
4. The minimal set of fields and the justification for each one
Every field in the journal has to answer the question: what will I be able to settle thanks to it three months from now, and what will I fail to settle without it. A field with no such answer is dropped. That is the only criterion I hold to below – not completeness, not the elegance of the spreadsheet.
| Field | Layer | What it lets you settle | What breaks without it |
|---|---|---|---|
| Entry timestamp (full date and UTC instant; retain the original clock and offset) | 1 | whether the entry was created before the order was placed | all the remaining fields lose their status as pre-decision data – and that cannot be reconstructed later |
| Entry condition (a reference to an item in the catalogue) | 1 | whether the entry had a basis recorded earlier than the entry itself | the name of the setup starts being assigned after the fact and fits everything |
| Degree of conviction on a 1–10 scale | 1 | whether declared confidence has any relationship at all with the later hit rate | there is nothing to compare with anything; confidence stays unaccountable |
| Exit plan – invalidation level and target | 1 | whether the levels existed before entry and whether they were changed | every later change looks in the record like the original plan |
| Context of the decision – which decision of the day it is, how long since the previous one closed | 1 | whether the decision came in a sequence after a loss, and how quickly | no time gap can be computed, and that is one of the few quantities measurable directly |
| Deviation code | 2 | whether a breach occurred and of what kind | breaches stay in the comments, and comments cannot be summed |
Start with the six fields in the table. Complete the first five before sending the order, using the condition catalogue and exit plan prepared beforehand. Add the deviation code after comparing execution with the plan; enter "none" when compliance can be confirmed. Check how long this takes in your tool and whether the entry is being made only after the trade. A short form helps organise the work, but does not guarantee complete data.
Extend the form when each new field answers a specific question and the basic record remains complete. Four gap-free weeks are an example review window, rather than a condition for a valid journal. Extensions include declared state on a defined scale (layer 1), event times during the position, such as order modifications (layer 2), the outcome as a fact (layer 3), and a separate comment (layer 3). A field may be added earlier if needed to test a rule; record the schema change.
5. Deviation codes – how to record a breach so that it can be counted
A deviation described in a sentence is an anecdote. A deviation tagged with a code from a closed list is a row that will enter a summary two months from now. The difference only shows up across a series, and that is why it is so easy to dismiss at the first entry.
| Code | Observable event | What settles it |
|---|---|---|
| N0 | no deviation – execution consistent with the layer-one entry | a comparison of two records, without knowing the outcome |
| N1 | no layer-one entry with a timestamp earlier than the opening of the position | the count of entries set against the order history |
| N2 | entry taken with the catalogue condition unmet, or with no reference to any condition | the content of the "entry condition" field and the catalogue description |
| N3 | position size inconsistent with the rule in force before the session | the order history and the rule – the rule itself and its level come from section 9 |
| N4 | an exit level changed after the position was opened with no recorded rationale | the order modification history and layer two |
| N5 | a decision outside the time window recorded in your own plan | the order time and the session plan |
Codes N2–N5 are not a new list of criteria. They are the same criteria that article 10.4 gives as the process assessment card, written as a label instead of a "yes" or "no" answer. Coding adds exactly one thing to them – countability across a series – and that is the entire reason for rewriting them as codes.
Three rules make these codes countable.
The list is closed. Five or six codes, not fifteen. An "other" box may exist only temporarily: if after a month more than one entry in ten gets "other", the list does not describe your process and has to be rebuilt rather than extended with further exceptions.
In the main summary, one decision receives at most one code: the earliest established deviation. This is a convention: it loses information about simultaneous violations, but produces mutually exclusive categories that sum to the decision count. Ties between N2, N3 and N5 use the lower-numbered code: N2 before N3, N3 before N5. N1 takes precedence to mark a missing entry. It does not establish that every other condition is uncheckable; a standing limit or time window may be assessed from history. Retain an additional flag for a known violation. The numbering is not a severity scale; the selection rule aims to reproduce the same classification on recoding.
N1 marks a missing required pre-decision record, rather than automatic proof of every violation or of compliance. Some conditions remain uncheckable; others, such as a size limit or time window recorded before the session, may still be tested against the history. Count N1 separately in the main summary. If a known standing rule was breached despite the missing entry, retain an additional violation flag rather than discarding it. The flag does not change the primary code or total decision count.
Trigger condition: the closing of a position – every position, including profitable ones and including those where nothing happened at all.
Action: compare the execution with the layer-one entry and assign exactly one code from the list. Without reading the outcome; if the spreadsheet allows it, the result column is hidden at this moment.
Completion criterion: the code field is filled in — including when it contains N0. An empty field means the decision is not yet coded: keep it in denominator D and report it separately. It does not automatically become N1 or N0; N1 denotes a missing pre-decision entry, while N0 requires confirmed compliance.
Audit trail: the distribution of codes over the month; what interests us is how many times each occurred, not their description.
Failure criterion (a requirement separate from the form): if after a month more than half the codes are N0 but comments describe possible breaches, compare those same decisions with the plan and order history. This calls for checking definitions and assigned codes; N0 frequency alone is not proof. If almost every decision has another code, examine whether the rule is workable during those hours and how it was executed. Nine deviations in ten decisions justify that review, but establish neither that the rule is faulty nor what caused the breaches.
6. Process metrics across a series – a worked example
The series below is illustrative: sixty decisions over six weeks, arranged so as to show the arithmetic and one trap. This is not data from any account and it should not be treated as typical – this example does not estimate what is typical.
Three quantities, each with an explicit formula. D is the number of decisions in the order history, U – the number of decisions with no pre-decision entry (code N1), S = D − U – checkable decisions, V – checkable decisions with a deviation (codes N2–N5), Z = S − V – compliant decisions.
| Week | D — decisions | U — no entry | S — checkable | V — with a deviation | Z — compliant | S/D | Z/S |
|---|---|---|---|---|---|---|---|
| W1 | 11 | 6 | 5 | 2 | 3 | 45.45% | 60.00% |
| W2 | 9 | 4 | 5 | 3 | 2 | 55.56% | 40.00% |
| W3 | 12 | 4 | 8 | 2 | 6 | 66.67% | 75.00% |
| W4 | 8 | 2 | 6 | 1 | 5 | 75.00% | 83.33% |
| W5 | 10 | 2 | 8 | 1 | 7 | 80.00% | 87.50% |
| W6 | 10 | 1 | 9 | 1 | 8 | 90.00% | 88.89% |
| total | 60 | 19 | 41 | 10 | 31 | 68.33% | 75.61% |
Now the trap this whole table was built for. The compliance rate can be computed in two ways and both are arithmetically correct.
Z/S = 31/41 = 75.61% versus Z/D = 31/60 = 51.67%
The difference is 23.94 percentage points computed additively, as the subtraction of two percentages. The same difference expressed relatively, as a ratio, is 46.34% – that is how much higher the first metric is than the second. These are two different numbers describing one thing, and mixing them is the simplest way to turn your own journal into promotional material for yourself.
The trap concerns the first metric’s denominator. Compliance among checkable decisions may be inflated if entries are more often missing for violations. It need not rise whenever data is missing: the direction depends on which decisions were omitted. Report it alongside completeness and documented compliance across the full set. The relationship among the three quantities is expressed by this identity:
Z/D = (Z/S) × (S/D) = (31/41) × (41/60) = 31/60 = 51.67%
Documented compliance Z/D is the product of compliance in the checkable subset and record completeness. The identity is exact because S cancels in the product. It does not mean that any improvement in one factor offsets deterioration in the other—their values matter. Z/D = 51.67% is a lower bound on compliant decisions, not an estimate of true compliance across all 60. With the status of 19 missing cases unknown, the actual share may range from 31/60 to 50/60, or 51.67–83.33%. Rising completeness alone does not establish better execution.
The completeness of the record rises in this series from 45.45% in the first week to 90.00% in the sixth. That is a rise of 44.55 percentage points (additively) or of 98.00% (relatively, as the ratio 1.98). Both numbers are true, they describe different things, and neither of them matters more here than the observation you can see in the table without any arithmetic: out of twenty-nine markings that are not N0, as many as nineteen – that is 65.52% – are missing entries. The main problem with this record is not rule-breaking. It is that for nineteen decisions out of sixty – roughly every third one – it is impossible to establish whether a rule was broken at all. Those nineteen gaps are at the same time close to two thirds of everything in this record that even looks like a problem: there are almost twice as many empty fields here as there are deviations of every kind put together.
7. The gap from a loss to the next decision, and the limits of that metric
This metric has one rare virtue: it is computed from the order history alone, without asking anybody how they felt. It is the difference between two timestamps. Compute elapsed time between UTC instants after reconciling the source clocks, rather than subtracting unqualified civil-time displays. It also has a flaw that appears the moment somebody starts interpreting it on its own – and that flaw is what this section is about.
Below are two series of gaps in minutes, twelve observations each. Illustrative values, arranged so as to show the difference between two ways of summarising the same numbers.
| Series | Successive gaps (minutes) | Median | Mean | Under 20 minutes |
|---|---|---|---|---|
| after a decision closed at a loss | 3, 5, 7, 11, 14, 19, 26, 38, 52, 95, 168, 415 | 22.5 | 71.08 | 6 of 12 |
| after a decision closed at a profit | 22, 31, 35, 44, 51, 58, 63, 77, 88, 110, 140, 190 | 60.5 | 75.75 | 0 of 12 |
The medians are 38 minutes apart, computed additively; put as a ratio, the median after a loss is 37.19% of the median after a profit. The means are 4.67 minutes apart, that is 6.57% of the smaller value. The same pair of series, two ways of summarising, two entirely different impressions. The reason is visible in the table: one gap is 415 minutes and single-handedly drags the mean upwards, while the median ignores it. With distributions of this shape – a few very small values, a tail stretching far to the right – the mean describes mostly the tail, not the typical decision.
The second series, the one after profits, is nevertheless the more important, because without it the first means nothing. Six gaps under twenty minutes sounds alarming only when you do not know what the gaps look like in the remaining cases. Perhaps after every close, whatever the outcome, the next decision comes quickly, because that is the trading style and the frequency of signals. The comparison group is free in this calculation: the data is sitting in the order history anyway.
8. Self-rating scales and apparent precision
A "state" field on a scale from one to ten looks like a measurement and is sometimes treated as one. It measures a subjective report, rather than an objective state. In 1999 Norbert Schwarz assembled a review of research on self-report from which an uncomfortable thing follows: the answer depends not only on the respondent's state but also on the wording of the question, the range of the scale, the order of the questions and whatever serves as the reference point[5]. The same person in the same state will mark a different value on a scale from zero to ten and a different one on a scale from minus five to plus five. This is research on questionnaires, mostly in survey and laboratory conditions – what I carry over from it is caution towards your own boxes, not any numerical quantity.
It does not follow that the scale is useless. What follows is what it may be used for. For comparing your own entries with each other, provided the question is worded identically every time, the scale has the same range and it is filled in at the same moment – before the decision. For checking whether entries with a high reading more often sit next to particular deviation codes. And for nothing else.
There are three things it must not be used for, and they are the most tempting ones. It is not a reading of a physiological state. It cannot be compared with anybody else's scale, because the two measure different things despite the same digits. And there is no sense in computing averages from it to two decimal places – the difference between a six and a seven is not a psychological quantity, and an arithmetic mean of ordinal labels suggests a precision the data does not contain. Counting how many times each value appeared is more honest and sufficient.
There is one more inferential trap, the one people fall into most often with this field: "most of my losses fall at state 4 and 5". Even if that is true in your data, the tabulation alone establishes nothing. On days of higher volatility, tension rises, and so does the number of signals, and so does the size of the moves – three things at once, each of which could produce the same picture in the table. Before you treat the state as a cause, check whether entries with a high reading do not cluster at one time of day, on one instrument or in a week with data releases.
9. Selecting examples – how a journal confirms a conclusion decided in advance
This is the quietest way in which a journal stops being a tool and becomes evidence in a case whose verdict has already been handed down. It does not consist in falsifying entries. It consists in choosing which of them to look at – after the outcomes are already known.
The sequence is always similar. The trader sits down to a review with a hypothesis: "I think I close winning positions too early". He opens the spreadsheet, sorts by outcome, looks at a dozen or so trades closed at a profit, finds among them several where price went further – and leaves the review with a confirmation. He did not check how many trades there were in which closing early saved him from a reversal, because he had no reason to look at them. The hypothesis chose the sample.
The remedy is procedural and unpleasantly simple: the rule for selecting entries is written down before the data is opened. There are three admissible rules and all of them can be checked after the fact.
- All decisions in the period – the simplest one, and usually workable at a few dozen entries.
- All decisions with a given code – for instance every N4 in the quarter, without exception, including those that ended in a profit.
- A random sample of fixed size – when there are too many entries; the draw is made before anything is looked at.
The inadmissible rules are the ones everybody uses: "the ones I remember", "the worst ones", "the interesting ones". Each of them selects the sample on the basis of the outcome, and a sample selected after the outcome will answer only a question asked after the outcome.
On top of that come two habits, without which even a good selection rule achieves little. The first: the number of entries skipped is reported together with the result – "of 41 checkable decisions, 41 were examined" or "30 were examined, ten skipped for want of field X, one because of an export error". Gaps are data, not a cosmetic defect. The second: the definition of a deviation is not changed during a series. Changing the definition is admissible and sometimes necessary, but it breaks the series – from that moment the new one counts, and old and new weeks are not put in one column. Without that rule any metric can be improved within a week, without anything changing in behaviour.
10. A counterexample: an entry that looks like a breach and is not one
The metrics from the previous sections share one weakness: a suspicion built on them is sometimes accurate but never decisive. Below are two courses of events, both illustrative, both concerning a decision taken after a loss. The verdict comes out the opposite way from what the time gap suggests.
| First course | Second course | |
|---|---|---|
| Gap since the losing position was closed | 4 minutes | 6 hours |
| Layer-one entry | exists, timestamped 2026-03-12T08:12:00Z, before the session | none |
| Entry condition | the second of two conditions written down in the morning, met at 2026-03-12T09:12:00Z | undefined – field empty |
| Position size | consistent with the rule in force before the session | larger than the rule implies |
| Exit levels | set before entry, not moved | set after entry |
| Code | N0 – no deviation | N1, with N3 in the background |
Four minutes after a loss looks like everything this section usually describes – and it is nothing of the sort. Both conditions were written down in the morning, both belonged to the catalogue, and the second of them happened to be met shortly after the first position was closed. Nothing in the behaviour departed from the record made before the session. The gap here is a consequence of how signals were distributed in time, not a consequence of the loss.
Six hours, in turn, looks like calm and deliberation, and is a decision the record says nothing about, beyond the fact that the position size followed from no rule at all. A long gap is not evidence of anything – any more than a short one is.
It is worth noticing what is absent from this ruling. There is no outcome in it. Both decisions could have ended in a profit or a loss and the verdict would be identical, because it was produced solely by comparing the execution with the record made before the decision. That is the practical definition of assessment being independent of the outcome: if the result column were hidden, nothing in this table would change. The reverse situation – a profit that entrenches a breach – is handled by article 10.4.
Inverting the same counterexample matters just as much. If the morning entry in the first course had shown only one condition, and the second entry had been justified by a sentence added at 09:12, the verdict would read N2 – and that despite identical behaviour on the chart. What settles it is the record, not the appearance.
11. The review protocol – cadence, questions and the risk of rumination
A journal with no fixed review is an archive. That sentence sounds like a truism right up until somebody asks what exactly is supposed to happen in that review and how you know it has ended. Without those two things a review dissolves into browsing.
I propose two cadences with different jobs. A weekly one, short, purely numerical: the three numbers from section six and nothing more. Its purpose is not to draw conclusions – with eight or twelve decisions there is nothing to draw them from. Its purpose is that the data should exist at all, and that a gap in the record should be noticed after a week rather than after a quarter. A quarterly one, longer, on material of the order of a hundred and something decisions: this is where patterns are read. The split is arbitrary and follows from simple arithmetic – at a dozen or so decisions a week, a quarter gives the first set in which the distribution of codes stops jumping from week to week.

Four questions for the quarterly review. Each has an answer in the form of a number or a sentence that can be written down.
- Three numbers. How many decisions, for how many a pre-decision entry exists, how many of those carry a deviation. Without interpretation.
- Which code dominates and whether its share has changed relative to the previous quarter – the change is given in percentage points, that is additively, and checked against the possibility that it came from a change in the number of decisions rather than from behaviour.
- Whether entries appeared that cannot be classified. If so, the list of codes has stopped describing the process and it is the list that needs changing, not the trader.
- One change, or an explicit decision that there is none. Written down, with a date. "I am changing nothing" is a full-blooded answer, and more often the right one than yet another correction to the rules.
Trigger condition: a fixed day of the week for the short review and a fixed day after the quarter closes for the long one. The date written down in advance, not "when there is time".
Action: short review – three numbers into the week's row. Long review – the four questions above, on a set chosen according to a rule written down before the data was opened.
Completion criterion: four answers written down and one change, or an explicit decision that there is none. After that the review is closed, regardless of how many entries were left unread.
Audit trail: a list of changes with dates. After a year it shows how many times a rule was changed – and a rule changed every quarter never had time to be tested even once.
Failure criterion (a requirement separate from the form): if every successive review ends with a different change to the rules, the review is reacting to noise rather than to a pattern – lengthen the period and raise the number of observations required for any change at all. If, on the other hand, the review regularly ends with none of the four answers but with another look through individual losing trades, then it is not a review and has to be rebuilt from the format up: numbers only, without opening individual entries.
That last sentence touches on a risk which barely appears in writing about journals and which in practice eats more reviews than laziness does. A review easily turns into returning to particular losses – into replaying how they unfolded, into weighing what could have been done differently, into living through them a second time. That is an activity of a different kind from counting codes across a series and, unlike it, has no natural end.
Loss aversion is one possible hypothesis, rather than proof of why someone repeatedly returns to particular entries. Brown et al.’s meta-analysis[6] and Gal and Rucker’s critical review[7] address other questions; article 10.6 covers their scope and newer qualifications. The design point for a review is enough: define the set and stopping point beforehand, separate counts from comments, and do not treat rereading one loss as a new observation. This is a proposed workflow, rather than treatment for rumination.
12. What the journal will not show
The list is shorter than the popularity of this tool would suggest, but every item on it is hard.
It will not create an edge. Carefully documented execution of a rule with no justification produces carefully documented losses. Whether a set of rules has positive expected value after costs is settled by section 8, not by a column in a spreadsheet.
It will not establish the causes of your behaviour. It will show what was done and at what time. A sentence about why remains a hypothesis – and a hypothesis put forward by somebody who already knows the outcome. The record lets you at most narrow the list of explanations, by checking the simpler ones first: time of day, instrument, volatility, number of earlier decisions in the session.
It will not replace measurement. Everything you write into the state box is self-report – useful, but not a reading off an instrument.
It will not show what was not recorded. A missing entry is not a missing event or emotion. Without evidence, do not assume missingness is random or that gaps occur only in the hardest decisions. Count them and examine their relationship with time, instrument and other available information. Keep unknown status unknown, rather than assigning compliance or a violation.
It will not settle anything about strategy on your own series. A few dozen decisions is enough to compute the completeness of the record and the distribution of codes. It is not enough to say anything about the effectiveness of the rules – the order of magnitude of the sample required is discussed by article 10.4, and methods for testing a system belong to section 8.
It is also worth knowing where this record ends and others begin. Logging predictions about the market – what is going to happen and why – is the market hypothesis log. Outcome metrics, excursions and the tools for computing them are the workshop in section 15, and a ready-made spreadsheet with the full set of columns is a separate appendix. The rules whose observance you are coding here are created in section 9 and in the article on procedures. The journal replaces none of them. It checks one thing only, but it is the only thing that checks it: whether what was settled beforehand was actually carried out.
The nearest use for these numbers appears at the moment when the distribution of codes stops being a statistic and becomes a signal – when N3 and N4 markings accumulate within one session and the gaps between decisions shorten. What to do then, how to stop further orders and on what observable conditions to come back, is described by the article on reset protocols (10.9). The trigger condition for such a protocol comes precisely from the record – and that is the only way to stop it being set by the very state the protocol is meant to interrupt.
FAQ – the trading journal and recording the process
Is a journal filled in after the session worthless?
How many fields should a journal have?
How many trades are needed before a journal shows anything?
Should emotions be recorded?
A spreadsheet or a dedicated tool?
Does high compliance with the plan translate into results?
Sources and references
- 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. Participants estimated the probability of four possible endings to a described historical event; being told the actual ending raised the probability assigned to it despite an instruction to judge without that knowledge. Laboratory study, hypothetical stake.
- Nisbett R.E., Wilson T.D., "Telling more than we can know: Verbal reports on mental processes", Psychological Review, 84(3), 1977, pp. 231–259. A review together with the authors' own studies; participants' reports on the reasons for their own choices diverged from the factor actually manipulated by the researchers. The paper has extensive methodological critics and does not prove an absence of access to one's own states, but the unreliability of retrospective explanation of causes.
- Ericsson K.A., Simon H.A., "Verbal reports as data", Psychological Review, 87(3), 1980, pp. 215–251. An analysis of the conditions under which verbal reports are reliable data: a report given during a task and concerning the content currently being processed has a different status from a report retrieved from memory and concerning causes.
- Odean T., "Are Investors Reluctant to Realize Their Losses?", The Journal of Finance, 53(5), 1998, pp. 1775–1798. Around 10 thousand accounts at a US discount brokerage, equities, 1987–1993. Gains realised more readily than losses; winning positions that were sold delivered an excess return of about +2.35% over the following year against −1.06% for losing positions that were held, that is a 3.4 percentage point difference. Equity market, not the retail currency market – carrying the conclusion across is an extrapolation.
- Schwarz N., "Self-reports: How the questions shape the answers", American Psychologist, 54(2), 1999, pp. 93–105. A review of research on self-report; answers depend on the wording of the question, the range and labels of the scale, the order of questions and the reference point adopted. Survey and laboratory context.
- Brown A.L., Imai T., Vieider F.M., Camerer C.F., "Meta-analysis of Empirical Estimates of Loss Aversion", Journal of Economic Literature, 62(2), 2024. A meta-analysis of several hundred estimates from the experimental literature; a loss aversion coefficient of 1.955 with a 95% Bayesian interval of [1.820, 2.102], with marked variation between measurement methods. The value is an average across experiments with defined stakes and does not license linear conversion of specific amounts.
- Gal D., Rucker D.D., "The Loss of Loss Aversion: Will It Loom Larger Than Its Gain?", Journal of Consumer Psychology, 28(3), 2018, pp. 497–516. A review article challenging the universality of loss aversion; some results attributed to the phenomenon have simpler explanations. Cited here as evidence that the literature is ambiguous on this point.