Backtesting Skill

An AI-native backtesting skill for fast strategy idea evaluation. It provides a schema-driven workflow for AI agents and runs on top of Zipline.

Why This Exists

• Fast idea-to-backtest flow for AI agents
• Reproducible runs from constrained JSON schema
• Structured diagnostics beyond raw return metrics

This project is not a full portfolio optimization platform.

How It Works

Install as an Agent Skill (npx)

Install from GitHub (project-local):

npx skills add https://github.com/garroshub/backtesting-skill -a opencode -y

Verify the skill is installed:

npx skills list -a opencode

Optional checks and alternatives:

# Show available skills in this repo without installing
npx skills add https://github.com/garroshub/backtesting-skill --list

# Global install
npx skills add https://github.com/garroshub/backtesting-skill -g -a opencode -y

# Legacy alias (deprecated)
npx add-skill garroshub/backtesting-skill

Quick Start

Python 3.11+ is recommended.

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Validate a schema:

python scripts/run_backtest_from_schema.py --schema references/example_trend_dip_single_schema.json --validate-only

Run a single backtest:

python scripts/run_backtest_from_schema.py --schema references/example_trend_dip_single_schema.json

Run a grid search:

python scripts/run_backtest_from_schema.py --schema references/example_trend_dip_grid_schema.json

If bundle data is missing and schema allows it:

python scripts/run_backtest_from_schema.py --schema references/example_trend_dip_grid_schema.json --ingest-if-missing

Output Preview

Primary output blocks include:

• metrics
• performance_metrics
• trade_summary
• capacity_diagnostics
• risk_attribution
• practical_assessment

Grid mode adds:

• top_results
• stability_diagnostics

Supported Templates

• oversold_bounce_long_only
• sma_crossover_long_only
• trend_dip_buy_long_only

Multi-symbol equal-weight mode is available for:

• sma_crossover_long_only
• trend_dip_buy_long_only

oversold_bounce_long_only remains single-symbol strategy logic.

Data and Live Interface

• Active runtime data source: data.source = "bundle"
• Reserved data interfaces: csv, parquet, custom (replaceable by design)
• Reserved live interface: live_data (for example ibkr) is validation-only and not production live trading

See references/schema.md for complete fields and output contract.

Star History

Project Layout

• SKILL.md: skill behavior and workflow
• references/schema.md: schema and output reference
• references/example_*.json: runnable examples
• scripts/run_backtest_from_schema.py: runner
• scripts/ingest_yahoo_bundle.py: optional ingestion helper

License

MIT. See LICENSE.