Created for Event Horizon Labs as a part of the University of Chicago Project Lab, Spring 2026

This repository provides an iterative feedback loop for autonomous agents to research, develop, and implement trading strategies. Our goal is to create a verifiable environment where agents can autonomously iterate and improve upon trading strategies.

The following diagram illustrates how the Agentic System interacts with the Exchange and Execution engines to create the strategy feedback loop:

The repository includes a sophisticated snapshot system that automatically backs up trading strategies to AWS S3:

• Automated Backups: Snapshots are created automatically when pushing to snapshots/* branches
• Manual Triggers: Create snapshots on-demand via GitHub Actions
• Comprehensive Storage: Each snapshot includes code, backtesting results, and metadata
• 30-Day Retention: Automatic cleanup prevents storage bloat
• Secure & Reliable: Stored in AWS S3, separate from git repository to prevent data loss

Store and retrieve large research datasets (up to and beyond 40 GB) for agent training and backtesting:

• Manifest-Driven Discovery: Agents fetch metadata first to understand dataset structure
• Selective Partition Retrieval: Download only the data you need (date ranges) to minimize transfer costs
• Integrity Validation: SHA-256 checksums for every file
• Docker-First Execution: Reproducible agent environments with pre-configured AWS access
• Resume Capability: Interrupted downloads continue from where they left off
• Cost Optimized: Store 40 GB dataset for ~$1-5/month with selective retrieval patterns

• Strategy code (.py files, Jupyter notebooks)
• Backtesting results (JSON, CSV, visualizations)
• Performance metrics (returns, Sharpe ratio, drawdown, win rate)
• Metadata (timestamps, commit SHAs, workflow info)

• CLAUDE.md — agent bootstrap (auto-loaded by Claude Code)
• docs/OBJECTIVE.md — the research brief
• .claude/skills/ — preloaded research skills (backtest, analysis, snapshot, evaluate)
• research/config.yaml — research hyperparameters

• docs/operator/architecture.md — system architecture and design decisions
• docs/operator/aws-setup.md — step-by-step AWS infrastructure setup
• docs/operator/troubleshooting.md — common issues and fixes

agentic-trading-system/
├── execution_algos/                 # Reusable execution algorithm module
│   └── simple_execution_strategy/
│       └── execution_algorithm.py
├── execution_algos/                      # Trading strategy implementations
│   └── sample_momentum_strategy/    # Example strategy with results
│       ├── momentum_strategy.py     # Strategy code
│       ├── requirements.txt         # Dependencies
│       └── results/                 # Backtesting results
│           ├── backtest-results.json
│           └── trade-history.csv
├── .claude/
│   ├── agents/researcher.md         # Research agent definition
│   └── skills/                      # Preloaded research skills (backtest, analysis, snapshot, evaluate)
├── .github/
│   └── workflows/
│       └── snapshot-execution-algo.yml  # Automated snapshot workflow
├── docs/
│   ├── OBJECTIVE.md                 # Research brief (canonical for agents)
│   └── operator/                    # Architecture, AWS setup, troubleshooting
├── research/
│   ├── config.yaml                  # Research hyperparameters
│   ├── program_database.json        # Append-only attempt log
│   └── NOTES.md                     # Human assumption alerts
├── CLAUDE.md                        # Agent bootstrap
└── README.md                        # This file

Read docs/OBJECTIVE.md — it contains everything an agent needs: the metatask, how to access data, execution constraints, evaluation benchmarks, the research loop, and how to save results. All numeric values come from research/config.yaml.

Method 1: Automatic (Recommended)

# Create snapshot branch
git checkout -b snapshots/your-strategy-name

# Add your execution algorithm
mkdir -p execution_algos/your-strategy-name/results
# ... add your code and results ...

# Push to trigger automatic snapshot
git add execution_algos/your-strategy-name/
git commit -m "Add trading strategy with backtest results"
git push origin snapshots/your-strategy-name

Method 2: Manual

1. Go to GitHub → Actions → "Create Strategy Snapshot"
2. Click "Run workflow"
3. Enter strategy name and path
4. Click "Run workflow" button

See .claude/skills/snapshot/SKILL.md for detailed instructions.

Snapshots are stored in S3 with this structure:

s3://bucket-name/execution_algos/
└── strategy-name/
    └── 2026-04-04T12-30-45Z-abc1234/    # Timestamp + commit SHA
        ├── code/
        │   ├── strategy.py
        │   └── requirements.txt
        ├── results/
        │   ├── backtest-results.json
        │   └── trade-history.csv
        └── metadata.json                 # Snapshot metadata

For administrators setting up the infrastructure:

1. Follow the AWS Setup Guide to configure:

   • AWS account and S3 bucket
   • IAM user with minimal permissions
   • GitHub repository secrets
   • Lifecycle policies for retention

2. The GitHub Actions workflow is pre-configured and ready to use

3. Point agents at CLAUDE.md and docs/OBJECTIVE.md

A sample Execution Algorithm is included in execution_algos/simple_execution_strategy to demonstrate:

• Strategy code structure
• Backtesting results format
• How snapshots capture everything

• AWS credentials stored securely in GitHub Secrets
• IAM user has minimal permissions (PutObject, GetObject only)
• S3 bucket is private with no public access
• Lifecycle policies automatically expire old snapshots

Expected monthly costs for moderate usage:

• S3 Storage (10-100GB): ~$2-5
• API requests: ~$0.01
• Total: $3-10/month

Well within the $10-50/month budget range.

• Snapshot retrieval workflow (download via GitHub Actions)
• Performance comparison dashboard
• Automated strategy validation
• Multi-cloud backup support

This repository is designed for autonomous agents to iterate on trading strategies. Agents should:

1. Develop strategies in the execution_algos/ directory
2. Develop execution algorithms in the execution_algos/ directory
3. Include comprehensive backtesting results
4. Use the snapshot system to preserve iterations
5. Follow naming conventions in .claude/skills/snapshot/SKILL.md

Created for Event Horizon Labs - University of Chicago Project Lab

Snapshot Retention: 30 days
Last Updated: 2026-04-04