The obvious move is to export your trades and paste them into whatever assistant you already pay for. It works often enough to be dangerous. We ran fifteen questions across six models on a real trading journal to find out where it breaks, and the failures were not the ones we expected — the models were rarely confused. They were confidently wrong, in specific and repeatable ways.
1. They cannot add up, and they will not say so
Ask a model to total a column of two hundred rows and it will produce a number. It will look right. Sometimes it is.
In our run, two of the six models inverted a conclusion — reporting a losing group of trades as profitable — purely because they were asked to sum rows rather than read a figure that was already there. Every model answered the same question perfectly when the total was stated in the document.
Arithmetic is free in a program and unreliable in a language model. So compute every aggregate before the model sees it — net by hour, by weekday, by instrument, by size bucket — and state each one outright. The model then does what it is genuinely good at: noticing patterns across figures it can trust, and answering the follow-up.
2. Long exports get truncated, silently
A year of trading does not fit in a chat window. When the list is cut without explanation, the careful models refuse to answer and the careless ones guess — and a guess about the part of your record it could not see is indistinguishable from an answer.
Any boundary has to be stated: in the header, and again at the point of the cut. “The 500 most recent trades of 4,901” is a document a model can reason about honestly. The same 500 rows with no note is one that produces a confident description of your last two months labelled as your year.
3. Mixing automated and discretionary trading ruins the behavioural questions
This was the single best observation of the run, and it came from the models rather than from us. Two of them independently refused the behavioural questions, on the grounds that the journal mixed fixed-size algorithmic accounts with hand-sized discretionary ones.
They were right. “Do you size up after a loss?” is meaningless when half the trades come from a strategy that always trades two contracts. The average across both describes neither, and any conclusion drawn from it is noise dressed as insight.
The lesson is about the data, not the model: behavioural analysis has to be computed per account, and on the accounts where a human chose the size.
4. Identifiers hide in names
Removing the account-number column does not remove account numbers. In this journal, two account labels carried the broker number — and a model quoted one straight back in an answer.
If you are handing your record to a third party, scrub the labels as well as the columns. Journal notes are the other one to think about: they are private, often unflattering, and frequently about other people.
Why a document beats a built-in chatbot
Plenty of tools will generate an insight for you. The trouble with a generated insight is that it is true for a moment and then stale, and it cannot be argued with.
The value of asking an assistant about your trading is almost entirely in the second question: why was that day different, show me that without the automated accounts, is the pattern still there if I drop the three largest trades. That is a conversation, and it belongs in the assistant you already use, with your own history and your own preferences, not in a worse chat window rented to you inside a journal.
Questions worth asking
- Which hour of the day costs me money, and is the sample big enough to believe?
- Do I trade differently after a loss than after a win — size, timing, hold time?
- What happens to my results if you remove my best five trades? My worst five?
- Where is the gap between my gross and my net, and what is it costing me a month?
- Which of these patterns would you not act on, and why?
That last one is the most useful question in the list, and the one people never ask.
What Choptick hands over
One self-contained text document, built for exactly this. You paste it into whatever assistant you already pay for — Choptick does not rent you a chat window.
- Every aggregate pre-computed — net, win rate, profit factor and expectancy by hour, half hour, session, weekday, instrument, setup, hold time and position size. Stated, not left for the model to total.
- Every boundary declared, in the header and at the cut, so no answer can quietly describe part of your record as all of it.
- Automated and discretionary separated, with behavioural sections computed per account and on hand-sized trading only.
- Account numbers scrubbed from labels, not just from columns.
- Journal notes, grades and mistake tags left out entirely — they are private, and there is no setting that adds them.
- The same arithmetic as the dashboard, shared in code, so the document and the screen cannot disagree.
Figures in the examples are invented and illustrative. Nothing here is financial advice.