The Intelligence Behind One Decision.
A multi-agent AI platform that replicates the structure of a quantitative hedge fund — 24 specialized agents analyzing markets, debating strategies, and executing with institutional discipline.
No cherry-picked backtests
Metrics populate from paper trading simulations via the Alpaca profit factor across 5 tickers.
Paper trading simulations — no real capital at risk. Past performance does not guarantee future results.
How it works
Periodically, 24 agents run a full analysis cycle — transparent from signal to trade.
Data Collection
4 agents scan in parallel: price action & technicals, financial fundamentals, news sentiment, and macro indicators from FRED.
The Debate
5 investor bots — Value, Growth, Contrarian, Momentum, Quant — apply their frameworks. They don't agree. That's the point.
Risk Gating
Before any order: RiskManager validates VaR, position sizing via Kelly criterion, daily loss limits, and portfolio concentration. Math only, no opinions.
Execution
PortfolioManager synthesizes the debate into a final decision. ExecutionTrader places the order via Alpaca. Full audit trail saved.
Our Principles
The 10 Principles
Every agent, prompt, and architecture decision aligns with these principles. They are the constitution of QuantMind.
First, Do Not Lose Money
Capital preservation is the prime directive. Circuit breaker is sacred and independent of any engine.
Code Computes. LLMs Narrate.
Every ratio, probability, and position size is calculated in Python. The LLM interprets — never the reverse.
Measure Everything. Opine Nothing.
Every method, signal, and agent is continuously evaluated via triple-barrier outcomes, IC, Brier score, and attributed Sharpe.
Learn From Every Outcome
Thompson Sampling updates agent weights nightly. Methods that add value gain influence. Those that don't are demoted or archived.
24 specialized agents
Each agent has a distinct philosophy, LLM model, and role in the pipeline. They don't converge by design — disagreement produces better decisions.
Data Agents
Nine specialized agents scan markets in parallel — technical, fundamental, sentiment, macro, regime detection, factor analysis, and ML forecasting. No single data source or model drives the decision.
Philosophy Investors
Six distinct investment philosophies debate the same data through different frameworks. They disagree by design — the Portfolio Manager weighs their arguments, not their votes.
Decision Layer
A structured debate synthesizes opposing views. The Risk Manager enforces hard mathematical limits. The Portfolio Manager makes the final call — every decision is auditable.
Execution & Compliance
Orders go to Alpaca with bracket protection (TP/SL). A compliance engine validates every trade. An append-only audit log records everything. Performance is measured continuously.
Hedge funds are black boxes. QuantMind isn't.
Every signal, debate, risk gate, and trade is logged with a full reasoning trail. You can inspect why any agent voted the way it did.
Architecture
Dual-Engine Architecture
Two independent trading engines operate under a single risk umbrella — like a multi-strategy fund, not a single bot.
Engine 1 — Deliberative
24 agents analyze, debate, and decide. Deep fundamental, macro, sentiment, and ML analysis. 2-5 min per cycle. Horizon: days to weeks.
Engine 2 — Reactive
Real-time WebSocket stream. Candlestick patterns, confluences, and indicators. <2 seconds signal-to-order. Zero LLM. Horizon: hours to days.
Risk Umbrella
Unified risk layer over both engines. Knows every position in real time. Circuit breaker liquidates all positions if drawdown exceeds limits — no exceptions.
Method Effectiveness
Self-Learning System
QuantMind does not just trade — it measures, weights, and adapts. Every resolved trade updates its beliefs using Bayesian inference, not intuition.
Triple Barrier Labeling
Lopez de Prado methodology — every trade outcome is objectively labeled via take-profit, stop-loss, and time-horizon barriers.
Thompson Sampling
Bayesian weight allocation per agent and direction. Nightly updates over a 365-day rolling window with exponential decay.
PurgedKFold CV
Academic-grade cross-validation prevents time-series look-ahead leakage in ML models. No data snooping.
Demote & Promote
Underperforming methods are automatically demoted. Outperforming ones gain influence. The system adapts — no manual tuning required.
any LLM provider or data source without touching the agent logic.
Watch the debate in real time
Open the dashboard to see the agents debating live. Or explore the full source code — no signup required.