Agentic Market Making

A Rust-based experimental market-making research engine for studying inventory-aware quoting, public quote replay, and learned policy selection.

Status

This is a learning and research project, not a production trading system. The current focus is a clean Rust paper engine with a reproducible Python research loop.

The project now includes agentic proof-of-concept policies: Python trains small policy models from paper-session results, exports JSON models, and Rust executes those learned decisions during replay.

Agent Loop

The project has a complete research loop for adding learned policy agents:

public quote data
  -> Rust paper execution and policy gate
  -> CSV window outcomes
  -> Python model training
  -> JSON model artifact
  -> Rust model loading and policy execution
  -> reports and validation

Rust owns execution, accounting, fills, fees, inventory, and PnL. Python owns research workflows: data collection, model training, evaluation gates, and reports. JSON model files are the boundary between the two.

Future ML agents can follow the same pattern: train from policy-gate window outcomes in Python, export a compact JSON model, add a Rust policy variant that validates and scores the model, then evaluate it through the same policy gate before treating it as useful.

Requirements

• Rust toolchain with Cargo.
• Python 3. The active research harness uses the Python standard library.

What It Does

• Replays public top-of-book quote data through Rust paper-session policies.
• Compares static, adaptive, hybrid, selector, learned-selector, linear-agent, and bandit-agent policies.
• Models fills, fees, inventory, cash, mark-to-market PnL, and drawdown.
• Trains a small Python logistic-regression gate and exports it back to Rust.
• Trains a multi-action linear policy agent and exports it back to Rust.
• Trains an executable LinUCB contextual-bandit agent and exports it back to Rust.
• Evaluates policy robustness across multiple fill assumptions.
• Includes small checked-in public quote datasets for reproducing the current report.

Quickstart

python3 research/run_research_pipeline.py

For quick report regeneration from existing policy-gate outputs:

python3 research/run_research_pipeline.py --skip-policy-gates

The checked-in SOLUSD quote captures are enough to reproduce the current offline report. Collecting fresh public data is optional.

Run the learned selector against live public quotes in paper mode:

python3 research/run_paper_live_report.py configs/runs/kraken_solusd_learned_selector_maker_fee_paper_live_long.json \
  --run-id solusd_demo

Current Result

Latest seven-dataset research report:

• Public quote samples: 1440
• Replay windows: 48
• learned_selector configured utility: 0.000572
• selector configured utility: 0.000521
• adaptive configured utility: 0.000177
• linear_agent liquid-fill utility: 0.004511
• linear_agent is functional but mixed: it wins liquid-fill sensitivity and lags configured/conservative assumptions.
• bandit_agent is executable now, but remains research-only: configured utility -0.003747.

This is a small-sample research result, not evidence of a live trading edge.

Latest live-paper demonstration: 300 public quote samples, 70 paper fills, final inventory -0.2000, and max drawdown 0.0059.

Other Runs

cargo run -- run configs/runs/sample_replay.json
cargo run -- run configs/runs/sample_replay_sweep.json
cargo run -- run configs/runs/sample_paper_session.json
python3 research/collect_quote_dataset.py --pair SOLUSD --samples 120 --interval-seconds 2 --evaluate

Verify

cargo test --quiet
cargo clippy --quiet -- -D warnings
python3 -m py_compile research/*.py research/archive/*.py
python3 research/run_research_pipeline.py --dry-run

Docs

• Final research report
• Model assumptions
• Experiments and sweeps
• Live paper demo
• Research protocol
• Research workflow
• Config guide

Development Note

This project is developed as a learning and research effort with AI assistance for code generation, refactoring, and documentation. Design decisions, review, testing, and project direction are handled by the author.

Roadmap

• Preserve the final research result bundle.
• Improve the contextual-bandit agent only if more data shows it can beat simpler selectors.
• Keep improving the learned policy only when new data exposes a concrete weakness.