An event-driven trading engine for quantitative trading and reinforcement learning. Designed from the ground up with AI at its core, rather than "bolting on" ML modules to a traditional quant system.

Rust core for high-performance, Python interface for RL training, one codebase for the complete pipeline from backtesting to production.

Online Documentation · Examples

Design Philosophy

• AI First: RL environment and backtesting engine share the same data structures, zero difference between training and production
• Rust Core: Nanosecond timestamps, deterministic matching, zero-cost abstractions, backtesting throughput > 1M events/sec
• Python Front: Gymnasium-compatible interface via PyO3, directly compatible with Stable-Baselines3 / Ray RLlib
• Full Pipeline: Backtest → Train → HPO → Walk-forward → Track → Register → Deploy, all built-in
• 100% Open Source: Apache-2.0 license, no enterprise edition, no feature restrictions

Features

Backtesting Engine

• Multi-Level Matching: L1 basic matching → L2 order book → L3 multi-asset crossing
• Impact Models: Almgren-Chriss permanent/temporary impact + probabilistic latency + tiered fees
• Deterministic Replay: SimulatedClock + crossbeam-channel bounded 100K event queue
• Columnar Storage: Arrow/Parquet, 1M tick read/write < 15ms
• Streaming Engine: Real-time tick ingestion via StreamDataSource trait, CSV replay / exchange simulation / future WebSocket; StreamingStrategy trait for tick-driven strategies; StreamingMetrics tracks equity curve, Sharpe, max drawdown, win rate in real time
• Paper Trading: Simulated exchange with configurable slippage, fill probability, and partial fill ratio; deterministic RNG seeding for reproducible tests

RL Environment

• Gymnasium API: Discrete / continuous / hybrid action spaces
• Reward Functions: PnL / Sharpe / Sortino, based on unified ReturnHistory
• Vectorized: VecEnv supports multi-environment parallel rollout
• PyO3 Bindings: maturin packaging, 6 submodules

Training Pipeline

• Hyperparameter Optimization: Optuna integration + NSGA-II multi-objective + Pareto frontier + early stopping
• Rolling Forward Validation: Purged + Embargo + leakage detection + Deflated Sharpe Ratio
• Experiment Tracking: MLflow / WandB / Local / Memory four backends
• Model Registry: SemVer + stage lifecycle + auto-archiving + rollback
• Distributed Training: Ray Actor + Parameter Server + Checkpoint fault tolerance

AI Enhancement

• LLM Agents: ReAct + Tool Calling, built-in PlaceOrder / QueryPortfolio trading tools with SafetyMode risk control
• Agent Swarm: Multi-Agent collaboration with Actor model, voting consensus, dynamic scaling
  • MarketAgent: Market analysis and signal generation
  • RiskAgent: Pre-trade risk assessment and compliance checks
  • ExecutionAgent: Order execution with TWAP/VWAP strategies
  • AuditAgent: Decision logging and compliance reporting
  • SwarmOrchestrator: Agent lifecycle management, message routing, auto-scaling
• Model Ensemble: Voting / Stacking / dynamic weighting, online Sharpe ratio monitoring for auto-adjustment
• Explainability: SHAP feature attribution + counterfactual explanations + Explainer trait built-in
• Compliance Audit: Immutable trade logs + decision report archiving

Production Deployment

• Exchange Integration: Binance / OKX REST + WebSocket (auto-reconnect)
• Risk Engine: Pre-trade checks (12ns), real-time circuit breaker, position limits
• Inference Engine: ONNX / Candle dual backends + CPU/GPU affinity pinning + batch inference

DeFi Integration (Experimental)

Note: DeFi features are experimental and under active development. APIs may change.

• EVM Chain Support: Ethereum / Arbitrum / Optimism / Polygon
• DEX Integration: Uniswap V3 direct integration with optimal routing
• MEV Protection: MEV-Share for sandwich attack prevention
• Smart Contract Risk: Hybrid risk checks (off-chain fast + on-chain authoritative)
• Cross-Chain Bridge: LayerZero integration for multi-chain asset transfers

Quick Start

Install (Recommended)

# Basic install (core + data processing)
pip install axon_quant

# With ONNX inference support (onnxruntime, auto-loaded)
pip install axon_quant[onnx]

# With RL training dependencies (gymnasium, stable-baselines3, torch)
pip install axon_quant[rl]

# Full install
pip install axon_quant[onnx,rl]

Verify installation:

python -c "import axon_quant; print(axon_quant.__version__)"

Build from Source

For developers who want to modify the Rust core:

git clone https://github.com/pengwow/axon_quant.git
cd axon_quant

# Build
cargo build

# Test (2300+ test cases)
cargo test --workspace

# Static analysis
cargo clippy --workspace -- -D warnings

# Build and install Python wheel
maturin build --release
pip install target/wheels/axon_quant-*.whl

Training Examples

# Random baseline
python examples/01_random_agent.py

# PPO training
python examples/02_train_ppo.py --timesteps 50000

# HPO optimization
python examples/03_hpo/hpo_single_objective.py

# Rolling forward validation
python examples/08_walk_forward/walk_forward_basic.py

📖 Detailed RL training documentation: RL Training Guide

Architecture

AXON uses Cargo Workspace to manage 21 crates, organized in 9 layers:

┌─────────────────────────────────────────────────────────────┐
│  Layer 9: Application Entry                                  │
│  ├─ axon-cli        CLI tool                                 │
│  └─ axon-python     PyO3 unified entry (axon_quant package)  │
├─────────────────────────────────────────────────────────────┤
│  Layer 8: AI Agents                                          │
│  ├─ axon-llm        ReAct agent + Tool Calling               │
│  └─ axon-explain    SHAP / Counterfactual / Decision Report  │
├─────────────────────────────────────────────────────────────┤
│  Layer 7: Model Services                                     │
│  ├─ axon-inference  ONNX / Candle / tch inference engine     │
│  └─ axon-ensemble   Model ensemble (Voting / Stacking)       │
├─────────────────────────────────────────────────────────────┤
│  Layer 6: Training Pipeline                                  │
│  ├─ axon-rl         Gymnasium env + VecEnv + Reward functions│
│  ├─ axon-hpo        Optuna hyperparameter optimization       │
│  ├─ axon-distributed Ray Actor distributed training          │
│  └─ axon-walk-forward Rolling forward validation            │
├─────────────────────────────────────────────────────────────┤
│  Layer 5: Experiment Governance                               │
│  ├─ axon-tracker    MLflow / WandB / Local / Memory tracking │
│  └─ axon-registry   Model registry (SemVer + Lifecycle)      │
├─────────────────────────────────────────────────────────────┤
│  Layer 4: Production Execution                                │
│  ├─ axon-exchange   Binance / OKX adapters (REST + WebSocket)│
│  ├─ axon-risk       Risk engine (Position / Drawdown / VaR)  │
│  ├─ axon-oms        Order management system                  │
│  └─ axon-monitor    Monitoring + Health checks               │
├─────────────────────────────────────────────────────────────┤
│  Layer 3: Backtesting Engine                                  │
│  ├─ axon-backtest   L1/L2/L3 matching + Almgren-Chriss impact│
│  └─ axon-compliance Compliance audit + Reports               │
├─────────────────────────────────────────────────────────────┤
│  Layer 2: Data Services                                       │
│  └─ axon-data       Arrow columnar storage + CSV/Parquet     │
├─────────────────────────────────────────────────────────────┤
│  Layer 1: Core Types                                          │
│  └─ axon-core       Timestamp / Price / Quantity / Order     │
│                     / Event / Queue / Portfolio / SIMD        │
└─────────────────────────────────────────────────────────────┘

Threading Model

• Core Matching Engine: Single-threaded, avoids lock contention, ensures determinism
• I/O Thread Pool: tokio runtime, handles WebSocket / REST / file I/O
• Compute Thread Pool: rayon, factor calculation / data transformation / parallel backtesting
• Event Queue: crossbeam-channel bounded 100K, zero-lock design

Data Pipeline

All AXON modules share the same Arrow RecordBatch, zero-copy passthrough, no format conversion gaps:

Data Sources (CSV/Parquet/WebSocket/Mock/Exchange API)
    │
    ▼
axon-data (schema validation / time alignment / dedup / mmap cache)
    │
    ▼
Arrow RecordBatch (memory) ──→ TradingEnv / FeaturePipeline / BacktestEngine
    │
    ▼
InferenceEngine (ONNX/Candle batch inference < 1ms)
    │
    ▼
ExchangeAdapter (Binance/OKX live trading)

Layer Descriptions

1. axon-core: Foundation of the entire system. Provides Timestamp (nanosecond precision), Price / Quantity (based on rust_decimal), Order, Event, Queue, Portfolio core types, and SIMD-accelerated normalization and order book operations.

2. axon-data: Unified data access layer. Based on Apache Arrow's RecordBatch columnar storage, supports CSV / Parquet / Mock data sources, built-in FeaturePipeline (Z-Score normalization + sliding window).

3. axon-backtest: Event-driven backtesting engine. Supports L1 (price priority), L2 (order book), L3 (dark pool / auction) three-level matching, integrated Almgren-Chriss market impact model and probabilistic latency simulation.

4. axon-exchange: Production-grade exchange adapter. Unified ExchangeAdapter trait, implemented Binance / OKX REST + WebSocket integration, built-in exponential backoff reconnection and token bucket rate limiting.

5. axon-rl: Reinforcement learning environment. TradingEnv implements Gymnasium standard interface (reset / step / render), supports continuous actions (target position ratio [-1, 1]), discrete actions (position bins), multi-objective rewards and vectorized parallel environment VecEnv.

6. axon-inference: Model inference engine. Supports ONNX Runtime, Candle (pure Rust), tch-rs (PyTorch C++) three backends, with async batch inference pipeline, CPU/GPU affinity binding and model hot update capability.

7. axon-llm: Large language model agent. Based on ReAct reasoning loop, built-in "market analysis", "query portfolio", "place order" three tools, supports OpenAI-compatible backends and streaming responses.

8. axon-explain: Explainability engine. Integrates SHAP feature attribution, counterfactual explanations ("What if I hadn't bought, how would returns change") and structured decision reports, meeting compliance and strategy iteration needs.

9. axon-ensemble: Model ensemble. Provides HardVote, SoftVote, WeightedVote, Stacking, DynamicWeighted five strategies, supports online performance monitoring and automatic weight adjustment.

Repository Structure

axon_quant/
├── crates/                     # 21 Rust crates
│   ├── axon-core/              # Core types (time/types/market/order/event/queue/portfolio)
│   ├── axon-backtest/          # Backtesting engine (L1/L2/L3 matching + impact models)
│   ├── axon-rl/                # RL environment (Gymnasium + VecEnv)
│   ├── axon-hpo/               # Hyperparameter optimization (Optuna + NSGA-II)
│   ├── axon-walk-forward/      # Rolling forward validation (Purged + Embargo)
│   ├── axon-distributed/       # Distributed training (Ray)
│   ├── axon-tracker/           # Experiment tracking (MLflow/WandB/Local/Memory)
│   ├── axon-registry/          # Model registry (SemVer + lifecycle)
│   ├── axon-exchange/          # Exchange adapters (Binance/OKX)
│   ├── axon-inference/         # Inference engine (ONNX/Candle)
│   ├── axon-risk/              # Risk engine
│   ├── axon-oms/               # Order management system
│   ├── axon-monitor/           # Monitoring alerts
│   ├── axon-llm/               # LLM agent
│   ├── axon-python/            # Python bindings entry
│   └── axon-cli/               # CLI tool
├── python/                     # Python package (axon_quant)
├── examples/                   # Training example scripts
├── tests/                      # Tests (Rust + Python)
├── docs/                       # Design docs + ADR
├── scripts/                    # Build and test scripts
├── pyproject.toml              # Python packaging config
├── Makefile                    # Development commands
└── Dockerfile                  # Multi-stage build

Crate Matrix

Performance

Benchmarks

Workspace has established 50+ Criterion benches across 5 crates:

make bench                 # Full workspace, 5-10 minutes locally
make bench-cmp             # Save main baseline for PR comparison
make bench-one CRATE=axon-core BENCH=event_builder_tick   # Single bench
cargo bench -p axon-core -- impact_linear    # Direct cargo run

CI doesn't run bench (to avoid main runner performance noise). Report: target/criterion/<group>/report/index.html.

CPU/GPU Affinity

axon-inference provides affinity module for cross-platform core pinning to reduce cross-core cache misses:

use axon_inference::affinity::{AffinityPlan, pin_to};
let plan = AffinityPlan::new().with_cpus(vec![0, 1]).with_cuda(0);
pin_to(&plan)?;

Or via BatchConfig configuration (auto-called at BatchInferencePipeline::new startup):

[batch]
collect_cpu_cores = [0, 1, 2, 3]
collect_gpu_device_id = 0

Platform support: Linux / macOS full support, Windows runtime returns Err(AffinityError::NotAvailable) (use WSL2 / numactl instead).

Engineering Practices

• TDD Driven — Test first, implement later, CI enforces -D warnings
• 1200+ Tests — Unit tests + integration tests + Python scenario tests
• cargo clippy — Zero warning policy
• cargo-mutants — Mutation test coverage
• cargo-fuzz — Fuzz testing (matching engine / order book / risk control)
• Miri — Data race detection
• Loom — Deterministic concurrency testing

Documentation

• Installation & Quick Start
• AI-Native Core Design
• Strategy Development Pipeline
• LLM Agent Trading
• Production Deployment
• Traditional Strategy Migration
• API Reference
• FAQ

License

Apache-2.0

Disclaimer

This project is an open-source quantitative trading framework intended for research and educational purposes only.

• No investment advice: Nothing in this repository constitutes financial, investment, or trading advice.
• No guarantee of profits: Past performance (including backtesting results) does not guarantee future returns.
• Use at your own risk: The authors and contributors are not responsible for any financial losses incurred through the use of this software.
• Not production-ready: This software is provided "as is" without warranty of any kind. Use in live trading environments requires thorough testing and risk assessment.
• Regulatory compliance: Users are solely responsible for ensuring compliance with applicable laws and regulations in their jurisdiction.

By using this software, you acknowledge that you understand and accept these terms.