Quant Research Harness

Quantitative trading project for the COMS 4232 Advanced Algorithms final project, extended with a second research track for systematic trading experiments.

Why This Repo Exists

This repo has two tracks.

1. proposal_core

Course-project benchmark track:

• Exponentiated Gradient (EG)
• Online Newton Step (ONS)
• Approximate Universal Portfolio
• Equal Weight (1/n)
• Momentum
• Mean Reversion
• BCRP as the offline oracle benchmark

All methods share the same:

• historical market data
• long-only simplex constraints
• turnover-aware transaction-cost model
• evaluation pipeline

The primary metric is transaction-cost-aware cumulative log-wealth, with Sharpe, drawdown, and turnover as supporting diagnostics.

2. ml_extension

Extended research track:

• Rolling Ridge Alpha
• Attention Sequence Alpha
• Blended Alpha Portfolio
• deterministic transaction-cost-aware optimization

This track follows the standard quant pipeline:

• predictive signal model
• portfolio optimization
• cost / risk control
• deterministic execution path

Multi-Agent Positioning

The agent layer is for research support, not autonomous execution:

• a feature / experiment agent proposes new ideas
• a validation agent checks leakage and overfit
• a risk agent reviews concentration, turnover, and guardrails
• a report agent writes the research memo

The execution path should stay deterministic even if the research loop becomes more agent-assisted.

External Inspirations

This version was shaped using:

• the final project proposal and algorithm supplement
• Open-Finance-Lab/AgenticTrading
• TauricResearch/TradingAgents
• himself65/finance-skills

This repo stays narrower than those references. The focus here is reproducible research, proposal compliance, and an implementation path that another engineer can continue.

Project Shape

frontend dashboard (React / Vite)
        |
        v
public/project-snapshot.json
        |
        v
Python research harness
  |- proposal_core algorithms
  |- ml_extension models
  |- deterministic optimizer
  |- reporting + agent briefs

Run Locally

Install frontend dependencies:

npm install

Install Python research dependencies:

python -m pip install -r requirements-research.txt

Run the research pipeline:

npm run research:run

Run only one track if you want:

python -m research_harness.run --track proposal
python -m research_harness.run --track ml

Start the site locally:

npm run dev

Build the static site:

npm run build

Artifacts

Running the harness writes:

• artifacts/research/summary.json
• artifacts/research/research_report.md
• artifacts/research/proposal_core_metrics.csv
• artifacts/research/ml_extension_metrics.csv
• artifacts/research/daily_performance.csv
• artifacts/research/agent_feature_brief.md
• artifacts/research/agent_validation_brief.md
• artifacts/research/agent_risk_brief.md
• artifacts/research/agent_report_brief.md
• public/project-snapshot.json

The frontend reads public/project-snapshot.json, so the GitHub page reflects the latest generated run.

GitHub Pages

The Vite config uses a relative base path so the static build can be deployed cleanly to GitHub Pages. A sample Pages workflow is included in .github/workflows/deploy-pages.yml.

Recommended publishing flow:

1. Run npm run research:run
2. Verify public/project-snapshot.json updated
3. Run npm run build
4. Push to GitHub
5. Enable GitHub Pages with the included workflow

Handoff Notes

If another engineer is continuing implementation, start with:

• docs/ARCHITECTURE.md
• docs/HANDOFF.md

Important Warning

This project is for research and educational purposes. Backtests can be misleading, paper trading differs from live execution, and nothing here is investment advice.