KS
← Back to projects
CompletedJun 15, 2024 — updated Apr 5, 2026

X_QUANT: AUTONOMOUS QUANTITATIVE TRADING PLATFORM

299K+ line autonomous trading system with 20 AI agents (Claude Sonnet 4), 5-member Investment Council, regime-adaptive fusion, and 9 layers of risk control — achieving Sharpe 1.708 across 664 instruments.

lines 299K+sharpe 1.708
pythonmachine-learningquantitative-financeai-agentsclaudeibkrfastapiperplexity
View on GitHub →

Overview

X_Quant is a fully autonomous quantitative trading system that runs 24/7 with zero human intervention. Every day, it analyzes 664 instruments across 15 years of hourly data, consults 20 AI agents modeled after legendary investors, deliberates through a 5-member Investment Council, and executes trades via Interactive Brokers — all within a 9.9-minute pipeline.

The system achieves an out-of-sample Sharpe Ratio of 1.708 with a CAGR of 19.71% and MaxDD of just 6.32% — a Calmar Ratio of 3.117 (world-class by institutional standards). Validated through 12-fold walk-forward across 664 symbols over 15 years, with 7/7 hard gates and 10/11 medium gates passed.

Architecture — 4 Layers of Intelligence

┌─────────────────────────────────────────────────────────┐
│         LAYER 1: QUANT ENGINE (MetaEnsemble v56)        │
│   40 atomic alphas → 10 strategies → 6 sub-ensembles   │
│   Regime detection: 6 market states (vol + trend)       │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│          LAYER 2: 20 AI AGENTS (Claude Sonnet 4)        │
│   11 legendary investors + 9 specialist analysts        │
│   9 data sources per ticker · batch=2 · ZERO 429s       │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│          LAYER 3: INVESTMENT COUNCIL                     │
│   5 members + Chairman · 3-stage deliberation           │
│   Peer review · Perplexity fact-check · final vote      │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌──────────────────────┬──────────────────────────────────┐
│   LAYER 4: FUSION    │        RISK CONTROL (9 layers)   │
│   Regime-dependent   │   Kill switches (daily/weekly)   │
│   blend 90/10→70/30  │   Circuit breakers + watchdog    │
│   LLM veto system    │   IBKR heartbeat + slippage      │
└──────────────────────┴──────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│              IBKR EXECUTION + MONITORING                 │
│   xquant-ctl.sh · 4 LaunchAgent services · 24/7         │
│   Web terminal (React + FastAPI + WebSocket)             │
└─────────────────────────────────────────────────────────┘

Layer 1: Quant Engine

ComponentDetail
40 atomic alphasMomentum, mean-reversion, RSI, volatility, Fama-French, cross-asset, residual, FIP, structure
10 base strategiesXS momentum, residual, dual, FIP, TS, composite, reversal, RSI, VolPrice, Structure
6 sub-ensemblesRP12, BestOfBest, ConsBOB, HRP, RiskParity, MinVariance
Regime detection6 states: BULL_QUIET, BULL_VOLATILE, BEAR_QUIET, BEAR_VOLATILE, SIDEWAYS, CRISIS

Regime is detected from SPY vs. SMA200 and 20-day annualized volatility. Every downstream decision adapts to the current regime.

Layer 2: 20 AI Agents

Every ticker gets analyzed by 20 independent agents, each running Claude Sonnet 4 with a distinct investment philosophy.

11 legendary investors: Buffett (value/moat), Burry (contrarian/short), Munger (quality + mental models), Graham (net-net), Druckenmiller (macro/flows), Cathie Wood (innovation), Lynch (GARP), Fisher (scuttlebutt), Ackman (activist value), Jhunjhunwala (emerging market value), Pabrai (concentrated value).

9 specialist analysts: Sentiment, Regime, Risk Manager, Portfolio Manager, Valuation, Technical, Quant Interpreter, Fundamentals, Growth.

Each agent receives enriched context from 9 data sources per ticker: IBKR 15Y hourly OHLCV, FRED macro (yield curve, credit spread, CPI), SEC EDGAR insider trades, Finnhub news + earnings, FinBERT NLP sentiment (Apple Silicon GPU), and Perplexity Sonar web-grounded research.

Rate-limit safe: 10 batches of 2 agents, 25s delay = ~5.6 min, zero 429 errors.

Layer 3: Investment Council

3-stage deliberation modeled after a hedge fund investment committee:

  1. Independent Analysis — each of 5 members reviews the ticker with full context: 20 agent opinions, Perplexity fact-check, 20 days OHLCV, active strategies. Temperature diversity: 0.2 (conservative) → 1.0 (quantitative).
  2. Peer Review — anonymous cross-critique of all opinions. Ranking of strongest arguments.
  3. Chairman's Synthesis — final BUY/SELL/HOLD decision with stop-loss and take-profit. Must justify alignment or override of agent majority.

Perplexity fact-checks agent claims against real-time web data (1h TTL) before the council deliberates.

Layer 4: Fusion + Risk

Regime-dependent blend:

RegimeQuantAI
BULL_QUIET90%10%
BULL_VOLATILE80%20%
BEAR_QUIET85%15%
BEAR_VOLATILE75%25%
CRISIS70%30%

LLM veto system — asymmetric override (AI can brake, not accelerate):

  • Confidence >= 80%: position reduced 50%
  • Confidence >= 90% + consensus >= 70%: position reduced 75%
  • Council SELL (conf >= 70%): ticker removed

9 risk layers: daily kill switch (-3%), weekly (-5%), MaxDD (-15%), VIX>60 flatten, circuit breakers, duplicate order protection, slippage monitor (>50bps), IBKR heartbeat (5min TCP check), watchdog auto-restart.

Key Results

MetricValueStatus
Net Sharpe1.708EXCELLENT
CAGR19.71%EXCELLENT
Max Drawdown6.32%EXCELLENT
Calmar Ratio3.117WORLD-CLASS
Sortino Ratio2.369GOOD
PBO0.109PASS
DSR1.029PASS
Walk-Forward12-fold, 664 symbols, 15YVALIDATED
Hard Gates7/7APPROVED
Pipeline9.9 min, ZERO errorsVERIFIED

Daily Schedule

09:00 ET  Pre-market check + Perplexity morning briefing
14:30 ET  Priority data refresh (50 symbols from IBKR)
15:30 ET  20 AI agents analyze portfolio (~5.6 min)
15:45 ET  Council deliberation — 3 stages + fact-check (~2.5 min)
16:05 ET  Fusion + rebalancing via IBKR
16:10 ET  Snapshot: NAV, positions, P&L to journal
16:30 ET  Daily report + Telegram summary
22:00 ET  Full data refresh (664 symbols) + AutoResearch

Web Terminal

Professional dashboard (React + FastAPI + WebSocket, real-time updates every 5s):

Supervisor (NAV, regime, kill switches) · Portfolio (positions, P&L, exposure) · Signals (40-alpha heatmap) · Agents (20 agents with reasoning) · Council (decisions, dissent, fact-check) · Risk (VaR gauges, stress scenarios) · Research (strategy profiles) · AutoResearch (strategy discovery) · Journal (transaction history) · Control (kill switch panel, manual triggers)

Production Notes

  • Deployment — 4 macOS LaunchAgent services managed by xquant-ctl.sh (start/stop/restart/status)
  • Self-healing — watchdog process with auto-restart, zombie socket detection, Gateway monitoring
  • Cost — ~$8/day for full AI intelligence (20 agents + council + Perplexity)
  • Data — 664 instruments × 15 years hourly OHLCV (858 MB pickle store)
  • Testing — 13/13 critical checks PASS, end-to-end pipeline verified

Tech Stack

Python, NumPy, Pandas, XGBoost, LightGBM, CatBoost, scikit-learn, Optuna, Claude Sonnet 4 (Anthropic API), Perplexity Sonar, FinBERT, asyncio, APScheduler, Interactive Brokers API, FastAPI, React, WebSocket, PostgreSQL, FRED, SEC EDGAR, Finnhub


v1 — Original Quant Platform

The original version of X_Quant was a classical quantitative trading platform focused on ML-driven signal generation without AI agents.

v1 Architecture

┌─────────────────────────────────────────────────────────┐
│                    DATA INGESTION                       │
│   Market Data API → Normalization → Feature Store       │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│              ALPHA GENERATION (30 signals)               │
│   Technical · Microstructure · Fundamental · Cross-Asset │
│   Universal shift(1) anti-lookahead guard               │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│           ML PIPELINE v4.3 (6-model ensemble)           │
│   137 raw → 411 normalized → IC selection → ensemble    │
│   XGBoost · LightGBM · CatBoost · Ridge · Lasso · EN   │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌─────────────────────────────────────────────────────────┐
│              SIGNAL GENERATION (MetaEnsemble)            │
│   7 base strategies → 6 sub-ensembles → IR-boost        │
└────────────────────────┬────────────────────────────────┘
                         ▼
┌──────────────────────┬──────────────────────────────────┐
│   EXECUTION ENGINE   │        RISK MANAGEMENT           │
│   Almgren-Chriss     │   VaR/CVaR (4 methods)           │
│   TWAP / VWAP / SOR  │   5-level drawdown controls      │
│   IBKR Live Gateway  │   Circuit breakers + regime       │
└──────────────────────┴──────────────────────────────────┘

v1 Results

MetricValue
Lines of code90K+
Sharpe Ratio1.236
CAGR16.97%
Max Drawdown13.05%
Win Rate68.7%
Strategies18
Alphas30

v1 Validation

  • CPCV — Combinatorial Purged Cross-Validation (12 folds, 5-day purge)
  • PBO — Probability of Backtest Overfitting: 0.01 (target < 0.30)
  • DSR — Deflated Sharpe Ratio: 1.189 (target > 1.0)
  • Hansen's SPA Test — Superior Predictive Ability
  • Walk-Forward Efficiency — 71.18% (target > 70%)

v1 → v2 Evolution

Metricv1v2
Lines of code90K+299K+
Sharpe1.2361.708
CAGR16.97%19.71%
Max Drawdown13.05%6.32%
Calmar3.117
Alphas3040
Strategies1810 + 6 sub-ensembles
AI agents020
Risk layers59
Data sources19 per ticker
Instruments664
AutonomySemi-manualFully autonomous 24/7