mirror of
https://github.com/NicolasBohn/NexQuant.git
synced 2026-07-27 23:47:46 +00:00
TPTBusiness
5b1ded36ca
NEW ARCHITECTURE:
┌─────────────────────────────────────────────────┐
│ Phase 1: Factor Generation (Open Source) │
│ - Generate factors with LLM v3 prompt │
│ - Backtest each factor in Qlib Docker │
│ - Save to results/factors/ with code + desc │
│ - Continue until 5000+ valid factors │
└─────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────┐
│ Phase 2: ML Training (Closed Source - Local) │
│ - Load top 50 factors │
│ - Train LightGBM model │
│ - Validate (IC, Sharpe) │
│ - Save to results/models/ │
└─────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────┐
│ Phase 3: Portfolio Optimization (Closed Source) │
│ - Select uncorrelated factors (max corr 0.3) │
│ - Optimize weights by IC │
│ - Backtest portfolio │
│ - Save to results/portfolios/ │
└─────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────┐
│ Phase 4: Strategy Generation (Closed Source) │
│ - Generate trading rules │
│ - Add risk management │
│ - Save to results/strategies/ │
└─────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────┐
│ Phase 5: Iterative Improvement (Closed Source) │
│ - Use ML results as feedback │
│ - Generate better factors │
│ - Loop back to Phase 1 │
└─────────────────────────────────────────────────┘
FILES CREATED (Closed Source - NOT in Git):
- rdagent/scenarios/qlib/local/ml_trainer.py
- rdagent/scenarios/qlib/local/portfolio_optimizer.py
- rdagent/scenarios/qlib/local/quant_loop_advanced.py
- rdagent/scenarios/qlib/local/__init__.py
FILES MODIFIED (Open Source - in Git):
- rdagent/scenarios/qlib/quant_loop_factory.py
- .gitignore (added local/ exclusion)
GRACEFUL DEGRADATION:
- If local/ components don't exist → Standard loop
- If < 5000 factors → Standard loop
- If LightGBM not installed → Falls back
- Open source users get FULLY FUNCTIONAL system
USAGE:
# Standard (always works):
rdagent fin_quant
# Advanced (automatic if local components exist + 5000+ factors):
# Same command - factory auto-selects appropriate loop