How accurately a bot built from a trade journal repeats a trader's setup

An open evaluation of the “Journal → bot” feature: method, results with confidence intervals, and limitations.

Published September 17, 2026 · Bogdan Movchan

Setup

Real journals with a known ground-truth setup don't exist, so the journals are synthetic but built on real history. A hidden setup is chosen — one of 28 rules (counter-trend, trend, breakout, patterns; longs and shorts). Entries are picked on bars where it fired: spread out in time, from a single market wave, or with one wrong entry, from 3 to 20 of them. A bot is built from the entries with the same code as on the website, and its trades are compared with the real setup's trades using the same backtesting engine.

Data: 21 datasets — crypto, forex, stocks and markets with a planted setup; 4h and 1d timeframes; 25,597 trials.
Evaluation on the last 40% of history, which the bot did not see when it was built.
Metric: trade match (entry within ±1 bar) — precision, recall and F1.
Baselines: random entries at the setup's frequency and an “always enter” bot.
Intervals: cluster bootstrap over (dataset × setup) pairs.

Results

Method	Trade-match F1
Random entries at the setup's frequency	0.12
Always enter	0.24
Spread thresholds (previous method)	0.36
Statistical selection (on the website)	0.44
Ceiling: bot from all of the setup's trades	0.77

Statistical selection versus thresholds: +0.079 (95% interval +0.069 to +0.088); better across all markets and timeframes.
With 3 entries: 0.37 versus 0.21; with entries from a single market wave: 0.47 versus 0.29.
Share of the ceiling: 3 entries — 51%, 8 entries — 61%, 20 entries — 64%.
On unseen data the error grows by about 10%, equally for both methods: statistical selection adds no overfitting.

Limitations

The journals are synthetic — a pilot will show accuracy on real traders.
No gain on trend setups or with 15–20 widely spread entries.
A single wrong entry noticeably lowers accuracy for both methods.
The condition language is limited: even a bot built from all of the setup's trades matches 77%.
The “average trade result error” metric proved weak — an “always enter” bot is almost indistinguishable from the threshold method by it — so conclusions rely on trade matching only.

Reproducibility

The test harness, data loaders and analysis are kept with the platform code, and the test reruns with a single sequence of commands. The feature itself is described on the “Journal → bot” page.

The screener is always free. The bot builder, journal and analytics are part of the subscription, with a free 3-day trial.

Not individual investment advice. Trading involves the risk of losing money.