Goal

Build a production-ready V5 trading system for XAUUSD that is robust to market costs and suitable for live paper/live trading. Focus on risk controls, walk-forward validation, realistic fills, and a live inference API.

Key Requirements

Realistic backtests with slippage, commissions, partial fills, and order types
Walk-forward training & validation to reduce overfitting
Modular model artifacts: tree-only + Keras artifacts (same as V4) but with versioning
Live inference API with strict risk manager, position sizing, and circuit-breakers
Paper-trade adapter to test on real exchange simulator/broker sandbox
Monitoring & alerting (P&L, drawdown, latency, model drift)

Milestones

V5 spec + scaffolding (this doc + FastAPI scaffold)
Realistic backtester improvements
Walk-forward validation pipeline
Ensemble retrain with robust techniques
Live inference + paper-trade adapter
Monitoring, HF model card, and deployment

Acceptance Criteria

Backtest Sharpe > target (configurable) on OOS sets with realistic costs
Max drawdown within risk budget
Live inference latency < 100ms per request (p95) on target hardware
Automated rollback on model drift or large drawdown

Notes

Start from existing V4 codebase for feature engineering and model saving. Reuse safe serialization approach. Incrementally add safety features and validate each milestone with tests.