Three months after a test, most people remember the result and forget everything that produced it: the costs they assumed, the parameter they changed, the two runs that did not work. A research journal fixes that. It is a dated record of what you planned, what you changed and what you concluded, kept so that your past experiments remain evidence rather than impressions.
What a research journal is for
A trading journal records trades. A research journal records experiments. The difference matters because an experiment can be useful even when nothing was traded and even when the idea failed.
A good journal answers four questions about any past test:
- What exactly did I expect, and when did I write it down?
- What were the inputs: data, parameters, costs, size?
- How many variations did I try before this one?
- What did I decide, and why?
If an entry cannot answer all four, it is a note, not a record.
The structure: one entry per run
Each entry describes a single run. Keep the fields the same every time so entries can be compared side by side and exported as a table later.
| Field | What to record | Why it matters |
|---|---|---|
| Entry ID and date | A sequential ID and the date written | Establishes order: plan before result |
| Parent hypothesis | The hypothesis this run tests, by ID | Groups runs that belong together |
| Mode | Exploratory or confirmatory | Separates searching from checking |
| Change from previous | The one thing that differs from the last run | Makes cause and effect readable |
| Setup | Scenario, template, parameters, fees, slippage, position size | Makes the run reproducible |
| Expected result | Written before running | Shows whether you were surprised |
| Observed result | The key metrics, with number of trades | Prevents rounding into a story |
| Trial count | How many runs exist under this hypothesis so far | Puts the best result in context |
| Decision | Keep, revise or drop, with one reason | Turns results into direction |
Two fields deserve explanation.
Mode: exploratory or confirmatory
The Center for Open Science, which promotes preregistration in research, makes a point that applies directly to strategy work: the same data cannot be used both to generate a hypothesis and to test it, and planning in advance helps separate planned from unplanned work. Both kinds of work are legitimate. Exploration is how ideas are found. But a result from an exploratory run is a lead, not a finding. Marking the mode on every entry stops you from quietly promoting one into the other.
Trial count: the number most journals leave out
If you try twenty parameter combinations and keep the best one, the best one looks better than it is. Harvey, Liu and Zhu made this argument about academic finance: with hundreds of factors tested across the literature, the usual significance bar of a t-statistic above 2.0 no longer makes statistical sense, and the hurdle has to rise with the number of tests. Bailey, Borwein, López de Prado and Zhu developed a method for estimating the probability that a selected backtest is overfit, which depends on knowing how many configurations were tried.
You do not need their formulas to apply the lesson. You only need to count. An entry that says "best of 14 runs" reads very differently from one that says "first run".
Three sample entries
Below is a short sequence from one illustrative hypothesis, recorded in the format above. All numbers are illustrative and come from synthetic sample scenarios, not market data.
Hypothesis H-07 (written 2026-10-02): After a close more than 2 standard deviations below its 20-bar mean, price returns toward the mean within 5 bars often enough to beat costs. Fail if net average return per trade is 0 or below, or fewer than 30 trades.
| Entry R-07.1 | Entry R-07.2 | Entry R-07.3 | |
|---|---|---|---|
| Date | 2026-10-02 | 2026-10-04 | 2026-10-09 |
| Mode | Confirmatory | Exploratory | Confirmatory |
| Change from previous | First run | Threshold 2.0 → 1.5 s.d. | Same as R-07.1, new scenario |
| Scenario | Range-bound | Range-bound | Volatility shock |
| Template and parameters | Mean reversion, 20-bar, 2.0 s.d., 5-bar exit | Mean reversion, 20-bar, 1.5 s.d., 5-bar exit | Mean reversion, 20-bar, 2.0 s.d., 5-bar exit |
| Costs and size | 10 bps fees, 5 bps slippage per side, 10% size | Same | Same |
| Expected | Net average per trade above 0 | More trades, smaller edge | Weaker than R-07.1 |
| Observed | 34 trades; net +0.21% per trade; max drawdown 4.1% | 61 trades; net +0.04% per trade; max drawdown 6.8% | 29 trades; net −0.47% per trade; max drawdown 11.3% |
| Trial count | 1 | 2 | 3 |
| Decision | Keep. Passed both criteria in range-bound | Drop the 1.5 threshold. Edge mostly consumed by costs | Revise. Fails in shocks (and under 30 trades); add a volatility filter as H-08 |
Notice what the sequence shows that a single result would hide. R-07.1 looks like a success on its own. R-07.2 records a variation that did not help, which is worth knowing so you do not try it again next month. R-07.3 shows where the idea breaks, and its decision spawns a new hypothesis rather than editing the old one.
Habits that keep the journal honest
Structure helps, but a few habits do most of the work.
- Write the expectation first. Fill in "Expected" before you press run. A blank expected field after the fact is a sign the entry was reconstructed.
- Change one thing at a time. If two parameters change between runs, you cannot say which one mattered.
- Record the failures. A journal of only successful runs is an advertisement for yourself. The dropped variations are what make the trial count true.
- Never overwrite. Revisions get a new entry ID. The old entry stays as it was.
- Review on a schedule. Once a week, read the decisions column only. Look for hypotheses left at "keep" with no follow-up, and for ideas you have tested many times under different names.
Where the journal lives
Paper works. A spreadsheet works. What matters is that the fields stay fixed and the dates are real. In Verdaxis, each note holds the hypothesis and assumptions, and every run is stored in that note's history with its scenario, template, parameters, fees, slippage and position size. You can export a note as JSON or CSV and keep it alongside your own journal, so the record does not depend on any single tool.
Making the first entry
Start small. Pick one idea you already have, write it as a hypothesis with failure criteria (the steps are in from trading idea to testable hypothesis), run it once, and fill in every field above. When a summary of that run includes sentences you are not sure how to weigh, what an AI trading bot can explain—and what it cannot shows how to sort them before they go into your journal.