Analyzing and trading on its own, an AI agent fleet
ArcAlpha runs a fleet of independent AI agents that analyze, decide, and trade across US, Korean, and crypto markets — governed by multi-LLM intelligence and hierarchical risk control.
No human in the loop
Simulation = live, 99% shared code
Crash-safe recovery
13+
AI/ML systems
8+
Trading strategies
3
Markets unified
24/7
Autonomous operation
Platform
Built for autonomous markets
Every layer — from intelligence to execution — is designed to run continuously, safely, and without a human in the loop.
Multi-LLM Intelligence
Gemini, OpenAI, and Anthropic are routed per task — analysis, decisioning, and self-reflection — through a Mixture-of-Experts orchestrator.
Autonomous Agents
Each asset is managed by an independent worker running its own analyze-decide-execute cycle with checkpointed state.
Multi-Market Coverage
US equities, Korean stocks, and crypto — unified under one risk and execution layer.
Hierarchical Risk Control
Layered guardrails from per-agent limits to portfolio-wide exposure caps, wash-trade handling, and leader-locked execution.
Competition & Evolution
Agents compete in performance tournaments; top performers earn more capital while underperformers are paused automatically.
Real-Time Autopilot
Dynamic strategy allocation adapts to market regimes, with live monitoring across every worker and portfolio.
Architecture
One pipeline, end to end
Market data flows through intelligence and decisioning, clears every risk gate, and only then reaches a broker — with state checkpointed back on every cycle.
How it works
One loop, running without you
Every agent repeats the same disciplined cycle around the clock — fetching fresh data, reasoning over it, acting, then saving its state so nothing is lost on restart.
01
Analyze
Each worker pulls cached market data and news every cycle, then routes it through a Mixture-of-Experts panel and a financial RAG layer for context.
02
Decide
Gemini, OpenAI, and Anthropic are selected per task to produce a decision with an explicit confidence score — with a rule-based engine as fallback.
03
Execute
Approved orders are routed by the OrderManager to the right brokerage, with position sizing, leverage caps, and day-trade rules enforced first.
04
Checkpoint
State is written to MongoDB after every cycle through a 9-state machine, so workers recover exactly where they left off after any restart.
Intelligence
An intelligence stack, not a single model
ArcAlpha layers specialist reasoning, retrieval, simulation, and learning systems behind every decision.
Mixture of Experts
Four specialists — macro, fundamental, quant, and sentiment — are weighted and aggregated per task by an expert router.
Financial RAG
Filings, earnings, news, and research are embedded into a vector store and retrieved to ground each decision in real context.
Market Regime Detection
A statistical model classifies the market into bull/bear and stable/volatile regimes so strategies adapt to conditions.
Monte Carlo Simulation
Thousands of simulated paths estimate Value-at-Risk, expected shortfall, and stress outcomes before capital is committed.
Reinforcement Learning
A PPO agent learns from a rich state space to refine trade timing and sizing over time.
Evolution Tournaments
Agents compete in performance tournaments with a Hall of Fame; winning traits propagate while weak ones retire.
Coverage
Three markets, one control layer
Unified data, execution, and risk across every venue ArcAlpha trades.
US Equities
Live US stock and ETF trading through Alpaca, with real-time quotes and order routing.
Korean Equities
Direct access to KOSPI and KOSDAQ via the Korea Investment & Securities API.
Crypto
Round-the-clock trading across major cryptocurrencies on a cash, non-marginable basis.
News & Macro Data
Sentiment and signals from Finnhub, NewsAPI, and FRED macro series feed every analysis.
Risk
Institutional-grade risk, enforced automatically
Guardrails sit between every decision and the market — no trade reaches a broker until it clears them.
VaR & Expected Shortfall
95% and 99% Value-at-Risk and CVaR quantify tail risk continuously.
Drawdown & Concentration
Live max-drawdown tracking and Herfindahl concentration limits keep exposure balanced.
Smart Position Sizing
Kelly-criterion and volatility-based sizing scale every position to conviction and risk.
Leverage & PDT Caps
Hard leverage ceilings, single-position limits, and pattern-day-trader rules are enforced.
Emergency Stop
A breach of any limit triggers immediate position closure across affected agents.
Synthetic-Data Guard
Live trading is blocked the moment mock or synthetic data is detected in the pipeline.
Strategies
A strategy for every regime
Allocation shifts dynamically as conditions change.
Profit Maximization
Dynamic stop-loss and take-profit bands that adapt to volatility.
Momentum Scalping
Tight targets and short holds that ride intraday momentum.
Evolving Profit
Risk parameters that learn and adapt from realized outcomes.
Ultra Scalping
High-frequency execution for fast, small, repeatable gains.
Statistical Arbitrage
Pairs trading and mean reversion on co-integrated assets.
Rule-Based Fallback
A deterministic engine that keeps trading when LLMs are unavailable.
Let the agents run the desk.
Sign in to the control room — monitor every worker, strategy, and position in real time.