Lumina Capital Engine v6.4.6 AI hedge fund backtests +187.3%

Lumina Capital's earlier iteration, which combined LangGraph, Qlib Alpha158, and Google TimesFM, was accepted into the awesome-quant repository's Trading & Backtesting section. That recognition appears to have encouraged the developer to move beyond a proof-of-concept and integrate live broker connectivity with Interactive Brokers, along with stricter risk controls.

The system's design reflects a deliberate split between a Python layer for reasoning and a Rust layer for low-latency execution. The Rust sentinel is benchmarked at 3.63 nanoseconds for a single position check and 3.62 microseconds for 1,000 positions, while the SQLite FTS5 memory engine retrieves past trading lessons in about 0.3 milliseconds without external vector databases or embedding APIs. These choices target the latencies that can otherwise delay stop-loss decisions during fast market moves.

On the finance side, the engine incorporates six established quantitative models, including volume clock sampling, order flow imbalance, the Amihud illiquidity ratio, Hayashi-Yoshida covariance with a positive semi-definite projection, half-Kelly sizing with hierarchical risk parity, and Black-Scholes gamma exposure. The reported five-year backtest covers the 2020 post-COVID volatility and the 2022 rate-hiking period, and the developer attributes the -12.1% maximum drawdown to the Rust sentinel, HRP diversification, and an ontology dictionary that constrains how the language model interprets raw factor values.