QuantForge: Architecture Analysis of a Production-Grade Multi-Asset Quantitative Trading Engine

This article analyzes the design philosophy and architectural implementation of the QuantForge quantitative trading engine, with its core viewpoint being "Alpha is fragile; the robustness of infrastructure matters more." The engine supports multi-asset classes including forex, commodities, stock indices, and cryptocurrencies, and has core features such as asset-specific independent modeling based on XGBoost, a seven-layer governance framework, forward validation mechanism, and replayable state architecture. The original project is maintained by manuelhorvey and was released on the GitHub platform on May 31, 2026 (link: https://github.com/manuelhorvey/QuantForge).

In the financial quantitative field, most developers focus on signal generation and model optimization but overlook infrastructure robustness. QuantForge treats runtime execution as a state machine and systems engineering problem, with core design decisions including:

• Forward Validation Mechanism: Assets must pass extended window backtesting before entering live trading to avoid data leakage;
• Training/Service Symmetry: Use the same feature builder for training and inference to eliminate training-serving bias;
• Replayability: Execution decisions and state changes are persisted as append-only logs, supporting historical replay and auditing;
• Asset Isolation: Each asset runs an independent model and engine instance; a single point of failure does not affect other assets.

The QuantForge research universe includes over 30 assets (forex, commodities, stock indices, cryptocurrencies), which need to pass strict forward validation to enter live trading:

1. Validation Process: 5-fold cross-validation with a 3-year training window + 1-year forward test window; multi-dimensional scoring (directional consistency, IC, win rate, institutional robustness); bidirectional validation; dynamic calibration of stop-loss/take-profit multiples.
2. Asset State Levels:
   • GREEN: Passed full validation, IC and win rate meet standards;
   • YELLOW: Partially passed, indicators are at critical levels;
   • RED: Failed or suspended from trading. Currently, 15 assets are in GREEN/YELLOW levels (e.g., BTCUSD, EURGBP, Gold GC).

Asset-Specific Modeling: Each selected asset uses an independent binary XGBoost classifier with the binary:logistic objective function, 300 trees, maximum depth of 2 (strong regularization), learning rate of 0.02, and removes the HOLD state to avoid unstable behavior. This design isolates risks and simplifies the problem space. Feature Engineering:

• Alpha Features: Volatility-adjusted arbitrage, multi-period momentum (21/63/126/252 days), Z-score mean reversion, volatility regime behavior, day-of-week effect, cross-asset macro momentum (DXY, VIX, SPX, WTI);
• Inference Features: EMA spread, ADX(14), RSI(14), Bollinger Band Z-score.

QuantForge adopts seven independent configurable governance layers with an aggregation strategy of "worst wins":

Governance Layer	Trigger Frequency	Scope	Main Function
Exposure State Machine	Per tick	Per asset	Exposure scaling control
Feature Stability	Per retraining	Per asset	Validity penalty
Meta Annotation	Per signal	Per asset	Position scaling
Macro Regime Overlay	Weekly	Global	Exposure and stop-loss adjustment
Liquidity Regime	Per signal	Per asset	Exposure and suspension logic
PSI Drift Detection	Per cycle	Per asset	Penalty and suspension
Portfolio Drawdown	Global	Portfolio level	Global throttling
Even if technical signals are strong, if the macro environment is unfavorable or the portfolio drawdown is large, the system will reduce positions or suspend trading.

Execution Engine:

• Unified Entry: All trades are centrally controlled via _can_enter() to prevent inconsistent logic, duplicate entries, and other issues;
• Lifecycle: PolicyDecision → FillResult → AttributionRecord, execution products are stored in append-only mode;
• Parallel Orchestration: Each AssetEngine runs independently; failures do not affect the global system. State Persistence: Uses WAL-mode SQLite with features including O(1) appending, 5-table normalization, replay-oriented design, periodic checkpoints, deterministic recovery, and backward-compatible JSON snapshots. The replayable architecture supports decision reproduction, strategy change validation, event sequence verification, A/B testing, etc.

Bitcoin Special Management: BTCUSD is configured with an independent HighVolSatellite engine, with an AUM cap of 5% and a volatility target of 40%, supporting macro gating participation, crisis regime suppression, and portfolio-aware throttling. Summary: QuantForge is a practice of evolving from academic research to a production-grade system. Key innovations include forward validation-driven asset selection, asset-specific independent modeling, seven-layer governance framework, replayable state architecture, and training/service symmetry. It provides a reference for developers of production-grade quantitative systems; its modular design and detailed documentation make it a high-quality open-source project.