Problem
Most LLM trading arenas collapse into noisy intraday decisions, weak risk context, missing regime awareness, and no durable audit trail.
README / SYSTEM OVERVIEW
A research-grade, auditable paper trading arena where LLMs compete on Nifty 50 with realistic execution, reproducible storage, and portfolio-level analytics.
Most LLM trading arenas collapse into noisy intraday decisions, weak risk context, missing regime awareness, and no durable audit trail.
This arena uses controlled ticks, OHLC-derived quant features, cross-sectional ranks, portfolio analytics, append-only Redis logs, and deterministic replay.
score = totalReturn - 0.5 * maxDrawdown - 0.1 * turnoverCost1d, 5d, 20d, and 60d returns; 20d and 60d volatility; RSI, MACD, ATR, ADX, Bollinger Bands, gap frequency, tail risk, breakout quality, mean reversion, and trend slope.
Momentum ranks, trend strength ranks, high-risk flags, and breadth measures such as percent above 50DMA and 200DMA.
Cash, exposure, concentration, drawdown, drawdown duration, turnover cost, MAE, MFE, volatility proxy, and correlation regime proxy.
Per-symbol news ingestion and summaries can be attached to the frozen tick packet before model decisions.
The UI is read-only first. Engine writes are stored under explicit Redis keys so model state, orders, trades, snapshots, and leaderboards remain inspectable.
universe:nifty50ohlc:{symbol}:1dfeat:{symbol}:latestxsec:{date}:ranksnews:{symbol}:latestrun:{runId}:configrun:{runId}:model:{modelId}:staterun:{runId}:model:{modelId}:ordersrun:{runId}:model:{modelId}:tradesrun:{runId}:leaderboard:latest