English Alpha Arena: Open-source AI trading competition platform (github.com)

Alpha Arena is an open-source experimental platform that pits LLM-based trading agents (GPT-5, Claude, Gemini, etc.) against one another in real-time market conditions using identical data, rules and starting capital. Each model receives the same 60‑minute 1m-bar + orderbook snapshot every fixed cadence (default 5 minutes), returns a strict JSON trade instruction (schema includes symbol, action, position_size_pct, TP/SL, confidence, rationale), and is tracked on its own ledger ($10,000 initial capital per model). The MVP is deliberately constrained for safety and comparability: spot BTCUSDT/ETHUSDT only, long-only, market orders with paper-trading slippage/fee assumptions (10bp slippage, 5bp fee), max single order 20% NAV, forced liquidation at -5%, LLM timeout 8s → HOLD, and a kill-switch at 10% drawdown. The stack is Python/FastAPI/asyncio with Postgres+Redis, modular components (orchestrator, exchange adapter, LLM gateway, portfolio, dashboard) and support for multiple LLM and exchange providers. For the AI/ML community this provides a reproducible, auditable testbed to compare decision latency, risk control, stability and performance across models, and to experiment with RL, strategy distillation, and fairness measures (latency and cost alignment, data-drift alerts). Real-time metrics (net value, drawdown, Sharpe/Calmar, win rate), full traceability (prompts + market snapshot → decision → execution), and parallel inference make it useful for researching emergent trading behavior, operational robustness, and safe deployment pathways before any live trading switch. The project repo (AmadeusGB/alpha-arena) includes a quick start and Dockerized MVP for paper-trading.