Lumina Capital
An open-source autonomous multi-agent quantitative hedge fund engine that strictly separates the narrative reasoning capability of LLMs from conventional quantitative computation, coordinating decisions through a multi-role investment research committee and deterministic factor models.
Timeline (last 90 days)
On Reddit, a developer open-sourced an autonomous multi-agent quantitative hedge fund engine named Lumina Capital (v6.0.0), with the code open-sourced on GitHub under the name lumina-ai-hedge-fund.
Lumina Capital adopts a layered architecture: the LLM is only responsible for narrative judgment and debate, while numerical computation is handed to traditional quantitative tools—this is the project's key design decision.
Lumina's first layer is a 7-Persona anti-sycophancy committee running on LangGraph, with the seven roles corresponding to seven investment philosophies: Buffett, Soros, Burry, Wood, Graham, O'Neil, and CIS.
Lumina's multi-agent committee mechanism contains three key design points: parallel independent analysis, bull-bear cross debate, and the CRO's veto power over tail risk.
Lumina's second layer is the Qlib Alpha158 factor library implemented in pure Python, fully rewritten in vectorized form, with LightGBM used as the meta decision filter.
Lumina uses Google's TimesFM for time series forecasting, outputting a 5-day probabilistic fan chart with an 80% confidence interval and a crash probability estimate.
Lumina has a 10-second-level high-speed sentinel thread that continuously monitors the dynamic ATR stop-loss line in the background, executing a physical exit immediately when triggered, without waiting at all for the LLM inference loop.
The Lumina system integrates a Neo4j-powered 3D causal knowledge graph and GraphRAG, used to monitor supply chain bottlenecks in real time and to provide episodic memory for agents.
The Lumina project is released under the CC BY-NC 4.0 license, and the author explicitly defines it as a simulation tool for research, education and simulation purposes, rather than a trading system that can be directly connected to live markets.
All Facts (9)
The Lumina project is released under the CC BY-NC 4.0 license, and the author explicitly defines it as a simulation tool for research, education and simulation purposes, rather than a trading system that can be directly connected to live markets.
50%UnverifiedLumina Capital adopts a layered architecture: the LLM is only responsible for narrative judgment and debate, while numerical computation is handed to traditional quantitative tools—this is the project's key design decision.
50%UnverifiedLumina's first layer is a 7-Persona anti-sycophancy committee running on LangGraph, with the seven roles corresponding to seven investment philosophies: Buffett, Soros, Burry, Wood, Graham, O'Neil, and CIS.
50%UnverifiedLumina's multi-agent committee mechanism contains three key design points: parallel independent analysis, bull-bear cross debate, and the CRO's veto power over tail risk.
50%UnverifiedLumina's second layer is the Qlib Alpha158 factor library implemented in pure Python, fully rewritten in vectorized form, with LightGBM used as the meta decision filter.
50%UnverifiedLumina uses Google's TimesFM for time series forecasting, outputting a 5-day probabilistic fan chart with an 80% confidence interval and a crash probability estimate.
50%UnverifiedLumina has a 10-second-level high-speed sentinel thread that continuously monitors the dynamic ATR stop-loss line in the background, executing a physical exit immediately when triggered, without waiting at all for the LLM inference loop.
50%UnverifiedThe Lumina system integrates a Neo4j-powered 3D causal knowledge graph and GraphRAG, used to monitor supply chain bottlenecks in real time and to provide episodic memory for agents.
50%UnverifiedOn Reddit, a developer open-sourced an autonomous multi-agent quantitative hedge fund engine named Lumina Capital (v6.0.0), with the code open-sourced on GitHub under the name lumina-ai-hedge-fund.
50%