The Journal as a Bias Detector

Recording the trade in a form that can be counted

A journal works when its fields are the ones that can be compared across trades, and there are not many of them. Before the trade: the setup by name, the entry trigger, the stop, the target or trail rule, and the size. After it: the result in R, and whether the plan was followed. Plus one line that is easy to omit and does most of the diagnostic work — the state the trade was taken in, in plain words. A record of what happened without that line is a description of prices; with it, it is a description of a person, which is the thing being studied. The tag is the important half. “Followed the plan” and “impulse” are the two categories that matter at the start, because the difference between them is the difference between a method that can be improved and a behaviour that can be removed. Anything more granular than that is a refinement: once two categories have been separated and priced, the next question — which setups are profitable, which times of day are worse, which states produce the worst sizes — can be asked of a dataset that can answer it. Attempting those refinements first produces a spreadsheet with no usable comparison in it. There is a second reason the tag has to be recorded before the result, and it is not about honesty. A plan noted after the outcome is a reconstruction: the stop that was “really” intended becomes the one that would have been better, and the setup becomes the one that is now visible. The whole value of the exercise is that the standard was written down when the outcome was unknown, so the comparison between the plan and the action is a fact rather than a memory. Memory edits itself to protect the ego, which is why the journal is the only instrument in this subject that can be trusted to disagree with its author. One quarter, two kinds of trade — Planned: 24 trades at +0.6R: +14.4R · Impulse: 16 trades at −0.9R: −14.4R · Net for the quarter: 0.0R ← · Share of trades that were impulse: 40% · The same forty trades at +0.6R: +24R ← · What the impulses cost: 24R — a quarter of the account at a 1% unit The sample is small enough that a lucky quarter can invert the sign, so the tag matters more than the average until there are a few hundred observations. The conclusion to draw from forty trades is which category to keep, not which parameter to tune.

Reading the record without fooling yourself

The first discipline in reading a journal is to slice by process before slicing by outcome. A quarter sorted by R produces a list of the biggest winners and losers, which is a list of variance; the same quarter sorted by whether the plan was followed produces the finding that matters. That ordering is counterintuitive, because the biggest number in the account is always the outcome, and it is the process slice that explains where the outcome came from. The second discipline is to keep the sample honest by refusing to change a parameter on evidence a single quarter can supply. Sixteen impulse trades produced a −0.9R expectancy with a wide confidence interval, and a trader who responds by adjusting the size rule after ten trades has changed the system on noise. The right reading of a small sample is directional rather than precise: the impulses lose, which is enough to justify removing them, and everything else waits. The third discipline is to review the record on a schedule rather than in a mood. A monthly or quarterly review of the tagged trades, read as a document rather than as a screen, produces the same finding every time and makes it impossible to ignore — which is the property the journal has that memory does not. The point is not to feel accountable to a spreadsheet. It is that the record answers the question the other lessons raise: not “am I disciplined?” but “what is the discipline worth in R, on my own trades?” • Slice by process before outcome: the tagged groups explain where the result came from. • Do not change a parameter on one quarter of data; read directionally, not precisely. • Review on a calendar, as a document rather than a screen. • The journal answers: what is the process worth in R on my own trades? The most common way a journal becomes useless is to fill it only after a bad day. The record then measures the worst trades and reports a system that loses, which is a description of when the journal is opened rather than of the trading. Every trade goes in, or the sample is a selection.

How many trades before the pattern is real

A journal becomes dangerous at exactly the point it becomes useful. Once you can count your trades, you can also count patterns that are not there — four good trades after following a particular checklist, three bad ones after trading before the open. With a small sample, almost any rule will appear to work, because random sequences contain runs and clusters by construction. Before acting on a pattern, ask how many observations it rests on and how many patterns you looked at to find it. If you examined eight habits and one of them shows a strong result, that is roughly what you would expect from noise alone. The practical thresholds are modest and worth writing down. A pattern built on fewer than about thirty trades of the same kind is a hypothesis, not a finding, and it belongs in the journal as something to keep counting rather than as a rule to trade. The related trap is that the categories start small: an impulse trade is rare, so its sample accumulates slowly — which means the cost estimate you compute in this lesson’s lab is a rough magnitude, not a precise number, and quoting it to two decimal places of R is false precision. The distinction that keeps the whole exercise honest is between an **outcome** and a **decision**. A well-planned trade that loses is a good decision; an impulse that happened to work is a bad one. If the journal records only profit and loss, it will teach you to repeat lucky behaviour. If it records whether the trade was planned, whether the size was the planned size, and whether the exit followed the rule, it can score the decision separately from the result — and that separation is the entire reason to keep the record at all. • Small samples contain patterns by construction; a rule with eight candidates tested will look good somewhere. • Under about thirty same-kind trades, treat a pattern as a hypothesis to keep counting. • Score decisions separately from outcomes, or the journal teaches luck. • Record whether the trade was planned and whether the size was the planned size.

The trades you did not take

A trading journal records the trades you executed, which is a **survivor sample** of a larger set: the setups you saw and skipped. That filter is invisible, and it is where most journals fail. If the reason you skipped a setup was a moment of hesitation rather than a rule, then every study of “what works for me” is contaminated — you are measuring the setups you were comfortable enough to take, not the setups the plan specifies. The repair is to log near-misses alongside positions: the setup that met your criteria, the price at which you would have entered, the stop, and what happened. Two things come out of that column. If the near-misses would have won and your actual trades lost, the leak is execution, not selection. If they were a mix of the trades you *should* have skipped and the ones you should have taken, the leak is that your criteria are fuzzy in exactly the region where the decision is hardest. Neither finding is visible from an executed-trades-only record. The same logic explains why a journal must record the plan before the trade, not the story after. Recounting a decision from memory produces a coherent narrative that matches the outcome, because hindsight bias is as reliable on traders as on anyone else. The pre-trade line is the only version written without knowledge of how it turned out. A win rate computed only from trades you chose to make is a win rate for the subset you liked, not for your system.

From record to change

A log that is written and never read is a diary. What turns a record into a process is the review, and the review has a design problem of its own: it has to be capable of changing behaviour without turning into a scorecard of outcomes. The first thing to know is that the act of recording changes the thing recorded. Tracking a behaviour tends to reduce it when the behaviour is unwanted, before any analysis has been done — which is a free gain, and it argues for logging the behaviour you want to change rather than only the results. Impulse trades, size deviations and rules broken are worth a column of their own precisely because writing them down is already part of the intervention. The fields that make a log countable are few: the setup, the reason for entry, the risk planned, the risk actually taken, the exit reason, the result, and — the one most often left out — whether the plan was followed. That last field is what separates process variance from outcome variance, and without it the log can only tell you how much money arrived, which is the least useful thing it could tell you. Two cadences answer two different questions and should not be mixed. A monthly review is about mechanics: was the record complete, were the written rules followed, were the sizes what the plan specified, how large was the largest deviation. A quarterly review is about the buckets: expectancy by setup, average win and loss, the distribution of mistakes, and whether the trades that were passed over would have worked. Running a P&L review monthly produces over-reaction, because at a monthly sample the result is mostly noise and the conclusions drawn from it are mostly wrong. The statistical caution established earlier in this lesson applies directly at the review. With a few hundred trades, a measured expectancy of a fifth of a unit of risk is not distinguishable from zero, so the monthly review should concentrate on what is countable — adherence, the size of the typical error, the number of unplanned trades — and leave the question of edge to the larger sample. That is not a concession; it is the difference between a review that can detect something and one that generates a story each month. The failure mode to watch for is the review that becomes a report card. If it asks whether money was made, it rewards luck and punishes correct process, and it will teach the wrong lesson whichever way the quarter went. If it asks whether the plan was followed and whether the reasoning held, then a losing quarter can be a good quarter — which is the only framing under which the record improves the trader rather than merely describing them. • Recording an unwanted behaviour is already part of changing it. • Log whether the plan was followed, or the record cannot separate process from outcome. • Monthly is for mechanics, quarterly is for expectancy — mixing them produces over-reaction. • A review that asks “did I make money” teaches the wrong lesson in both directions. One page a month is enough: the number of trades, the adherence rate, the average error in units of risk, the largest deviation, and one change to make. A review that produces ten changes produces none.

Sources

• Deliberate practice and feedback loopsEricsson, “Peak” (practice and feedback)
• Journaling trades and reviewing processSteenbarger, “The Psychology of Trading”
• R multiples and expectancyVan Tharp, “Trade Your Way to Financial Freedom”