research-strategy_skill

Overview

This skill automates the full research workflow for developing new trading strategies: search, implement, backtest, evaluate, decide, record, and loop. It requires using the market-intel-assistant for the search stage and integrates with a main script that handles evaluation, decisions, and logging. The goal is continuous, autonomous research with minimal manual intervention.

How this skill works

The skill starts by using market-intel-assistant to discover 1–2 simple candidate strategies. It scaffolds a test strategy file, runs a backtest runner, then the main daemon automatically evaluates results against predefined metrics, makes decisions, moves or modifies files, and records outcomes. Completed cycles report results and immediately continue the loop to the next candidate.

When to use it

• When you want an automated pipeline for researching many small strategies quickly.
• When you need a repeatable, auditable process for implementing and testing ideas.
• When you want continuous discovery and rapid iteration without manual gating.
• During exploratory research to surface promising strategy candidates for production.
• When you need a consistent evaluation standard across experiments.

Best practices

• Always start searches with market-intel-assistant — the pipeline depends on it.
• Follow the existing strategy interface (example: vwap.py) when implementing test_xxx.py to ensure compatibility.
• Keep test strategies simple and focused to reduce backtest time and noise.
• Let the main script handle evaluation, decision rules, moving files, and logging rather than manual edits.
• Monitor logs (research_workflow.log or research.log) and set alerts for failures or anomalies.

Example use cases

• Continuously harvest simple crypto strategies from web research and validate them automatically.
• Prototype a variant of VWAP and let the pipeline decide whether to promote it to production.
• Run bulk experiments overnight: the system searches, implements tests, backtests, evaluates, and repeats without manual steps.
• Maintain an experiment archive: every candidate is logged and moved according to objective rules for later review.
• Operate a sub-agent that spawns research sessions and iterates through candidate strategies autonomously.