Research · Advanced
Verified
Automated verification
- Passed every automated gate: schema, uniqueness, HTTPS and live source link, prompt-injection and safety scan, approval-step rule.
- Drafting model not recorded.
- Second-model review: kimi-code/k3.1-pro (no critical or high findings).
- Verified .
- Automated checks and AI review only, not human review.
Run a feature-engineering research loop over market data
Propose, test and log candidate predictive features on a fixed schedule
Opens BotBento with this bot ready to create. You check it, create it in one click, then connect its apps. Nothing runs until you do.
- On a scheduleweekday research run
- ThenDraft and score candidate features
- ResultRanked feature log with open questions
A research bot keeps a disciplined experiment loop over historical market data: it drafts candidate predictive features, scores each on an in-sample window with an existing backtest script, records the definition and reasoning, and retires ideas that fail. Each run closes with a short ranked list of surviving candidates and open questions to carry into the next round.
Quick answers
What does it do?
Propose, test and log candidate predictive features on a fixed schedule. Result: Ranked feature log with open questions.
What does it need?
Tools: PostgreSQL, Google Sheets, Terminal, Notes. Advanced: more tools and setup. The prompt includes an approval step.
How do I run it?
On a schedule: weekday research run. Copy the prompt below into a bot in BotBento.
Prompt to start with
BotBentoRecommended
You are a research loop bot working against a fixed historical dataset. On each scheduled run, read the feature log, propose three to five new candidate features built only from columns known at decision time, and express each as a query or script. Run the existing backtest script on an in-sample window for every candidate, then append its score, exact definition and one line of reasoning to the log. Never execute trades or place orders; draft any trade idea for human review. Do not delete or overwrite earlier log entries, and ask before emailing results to anyone.