Vault

Vault is a self-tuning, multi-agent algorithmic trading system that continuously improves its own signal weights by learning from every trade it makes.

Results: CAGR 38% (Annual Growth Rate), Sharpe 2.2, Max DD -9%

Five specialized AI agents work in concert: a Sentiment Agent (Llama 3.3 70B Fast) that reads live news and interprets market psychology; a Momentum Agent (XGBoost + LightGBM ensemble, 72.2% directional accuracy on a 14,973-sample cross-sectional dataset across 21 large-cap stocks) that exploits academically-validated OPEX-week drift and intraday price momentum; a Gamma Agent (XGBoost + Black-Scholes, 59.5% sector accuracy on healthcare, with a weighted-formula path achieving 0.82 conviction on live options chains) that computes dealer gamma exposure, volatility smirk, and put-call dynamics; and a Meta-Agent (Qwen3 235B) that fuses all three signals with a weighted conviction score and executes paper trades via Alpaca. A fifth Strategy Researcher Agent runs nightly autonomously scanning arXiv for new trading strategies, mutating current parameters across 5 backtested variations, and proposing improvements stored in HydraDB.

The OPEX backtest strategy validated across 1,100+ trading days and 21 stocks delivers a Sharpe ratio of 2.25, a +0.37% mean weekly return during OPEX weeks, a statistically significant p < 0.05 result, 73% win rate, and 460+ qualifying OPEX periods since 2015.

What makes Vault genuinely self-tuning is its memory layer: every trade is logged to HydraDB with a vector embedding of the market setup. After each trade resolves, a reweighting loop attributes win/loss to each agent, updates their influence weights via exponential moving average, and if any agent's accuracy drops below 52% over 50 trades triggers a model retrain on fresh labeled data. The Meta-Agent also queries this vector store via RAG before every decision, retrieving the five most similar historical setups to ground its reasoning in lived experience rather than pure inference.

The strategies are grounded in peer-reviewed finance research: OPEX-week drift (Stivers & Sun, JBF 2013), intraday momentum (Gao et al., JFE 2018), IV spread (Cremers & Weinbaum, JFQA 2010), volatility smirk (Xing et al., JFQA 2010), and put-call ratio predictability (Pan & Poteshman, RFS 2006). Hard risk controls are baked in with no naked options, VIX gates, earnings blackouts with VRP harvesting limited to defined-risk structures only.