Agentic Market Making

A Rust-based experimental market-making engine for studying inventory-aware quoting and, eventually, agent-driven strategy control.

Overview

This project is a learning and research environment for market-making ideas. It is intentionally not a production trading system. The current engine simulates a mid-price process, generates bid/ask quotes from an inventory-aware strategy, simulates fills, and tracks cash, inventory, and mark-to-market PnL.

The long-term goal is an agentic market maker: a system where a controller can observe simulation state and adapt strategy parameters over time. The current focus is building a clean, testable simulation core before adding agent or API layers.

Architecture

src/
  main.rs              # small runnable demo
  experiment.rs        # named experiment configs and reports
  lib.rs               # reusable library entry point
  engine/
    metrics.rs         # simulation summary statistics
    simulation.rs      # deterministic simulation loop
    state.rs           # accounting state and PnL updates
  market/
    mod.rs             # quotes, fills, and market-side types
  strategy/
    market_maker.rs    # inventory-skew market-making strategy
    mod.rs             # strategy trait

Current Model

• Price follows a seeded random walk.
• The strategy quotes around mid-price with a fixed spread.
• Inventory skews quotes lower when inventory is positive and higher when inventory is negative.
• Fills occur when a noisy simulated market price crosses the bid or ask.
• PnL is marked to market as cash + inventory * mid_price.
• Simulation metrics summarize fills, turnover, inventory exposure, and drawdown.
• Named experiments compare strategy settings under the same simulation conditions.

Run

cargo run

Test

cargo test

Roadmap

• Add explicit experiment configuration.
• Improve the fill model with arrival probabilities and volatility-aware behavior.
• Track richer performance metrics such as drawdown, fill count, turnover, and inventory risk.
• Add parameter sweeps for comparing strategies.
• Add an agent/control layer after the core simulator is stable.