Skip to content

AKKI0511/QuantTradeAI

Repository files navigation

QuantTradeAI

Give your coding agent a quant research lab, not a blank terminal.

Getting Started · Project YAML · Docs · Roadmap · Contributing

Python 3.11+ License: MIT CI


QuantTradeAI is built for coding agents — Claude Code, Codex, Cursor, and similar tools — to research trading strategies without repeatedly writing data, backtest, sweep, artifact, and deployment plumbing from scratch.

You give the research objective. The agent uses a generated workspace, config/project.yaml, and the quanttradeai CLI to run repeatable strategy experiments, compare artifacts, and recommend winners.


Why QuantTradeAI

AI agents can write code, but quant research requires repeatable infrastructure.

Without structure, agents write one-off scripts, scatter outputs across directories, and produce runs that cannot be compared or built on. Every session starts fresh with no memory of what worked.

QuantTradeAI provides:

  • A stable experiment environment — one project config drives data, features, research, and agents
  • Standardized run artifacts — every run writes structured outputs agents can read directly
  • A promotion model — backtest → paper → live, with live execution gated behind explicit human approval

Before and After

Without QuantTradeAI With QuantTradeAI
Setup Agent writes custom fetch and backtest scripts each time Agent reuses the quanttradeai CLI against one project config
Outputs Scattered across ad-hoc directories Every run writes to runs/ with standardized artifacts
Comparison No way to compare strategy variants Scoreboards and --compare are built in
Structure Research and agent code in separate scripts One config/project.yaml drives both
Safety No gate before live execution Backtest → paper → live, each requiring explicit promotion

Quickstart

Note

Package publishing is being stabilized. Until PyPI is available, use the local development setup below.

pip install quanttradeai
quanttradeai init my-lab
cd my-lab

Open my-lab in Claude Code, Cursor, Codex, or another coding agent and ask:

"Research RSI and SMA crossover strategies on AAPL and MSFT and find the best one."

The agent uses config/project.yaml and the quanttradeai CLI to run experiments, compare results on a scoreboard, and recommend a winner — without you writing any backtest code.


What init Creates

my-lab/
├── config/project.yaml                    # canonical project config — start here
├── AGENTS.md                              # tells coding agents how to use the CLI and YAML
├── CLAUDE.md                              # Claude Code-specific context and guidance
├── .claude/skills/quanttradeai-research/  # skill pack for Claude Code agents
└── .quanttradeai/workspace.yaml          # workspace-level metadata and defaults

The coding agent reads these files to understand how to use the framework correctly — you do not need to explain the workflow to it each session.


What the Agent Can Do

  • Create strategy labs — initialize multi-agent projects from templates (rule, model, llm, hybrid)
  • Run parameter sweeps — expand YAML-defined grids into parallel backtest variants
  • Compare scoreboards — rank runs by Sharpe ratio, PnL, or other metrics
  • Inspect artifacts — read summary.json.run_result and scoreboard.json from any run
  • Promote backtests to paper — move winning runs forward through explicit gates
  • Generate deployment bundles — emit local runners, Docker Compose, or Render worker configs
  • Keep live trading gated — live mode requires human acknowledgement at every promotion step

Artifact-Based Research

Every run writes machine-readable artifacts. Agents read these to decide what to do next — no manual result-parsing required.

Artifact What it contains
summary.jsonrun_result High-level outcome: winner, ranked candidates, failures
scoreboard.json Ranked metrics across all sweep variants
results.json Per-variant metrics for batch and sweep runs
metrics.json Full metrics for a single run
resolved_project_config.yaml Exact config used for the run — fully reproducible

Safety Model

Experiments start in backtest. Agents cannot self-promote to live trading.

Mode Gate
Backtest Always available — no credentials needed
Paper Replay-backed simulation — requires explicit promote from a passing backtest
Live Requires --acknowledge-live from a human, plus validated streaming and risk config

Caution

Live mode and broker-backed execution are fully opt-in. QuantTradeAI does not guarantee profitability.


Current Capabilities

Capability Status
Strategy labs, sweeps, and research runs Supported
Rule, model, LLM, and hybrid agents Supported
Backtest, paper (replay-backed), and live modes Supported
Scoreboards and run comparison Supported
Research-to-agent promotion pipeline Supported
Deployment bundles (local, Docker Compose, Render) Supported

For the full list, see the docs.


Local Development

git clone https://github.com/AKKI0511/QuantTradeAI.git
cd QuantTradeAI
poetry install --with dev

Initialize a workspace and open it in your coding agent:

quanttradeai init my-lab
cd my-lab

Dev commands:

make format   # auto-format
make lint     # run linters
make test     # run test suite

Documentation

Get startedGetting Started · Docs

ConfigureProject YAML · Config Overview

ReferenceAPI Docs · Roadmap · Contributing


License

MIT. See LICENSE.

About

A quant research lab built for AI coding agents to test, compare, and deploy trading strategies from one CLI workspace

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages