diff --git a/docs/architecture.svg b/docs/architecture.svg new file mode 100644 index 00000000..8cb348e7 --- /dev/null +++ b/docs/architecture.svg @@ -0,0 +1,158 @@ + +Predix data flow architecture +Full pipeline from Qlib data source through R&D loop, factor and model tracks, strategy generation, portfolio optimization, to live trading. + + + + + + + + + + +Qlib data (1-min EUR/USD) +2020–2026 · 96 bars/day + + + + + +R&D loop (rdagent fin_quant) + + +Propose +LLM + + + +Coding +CoSTEER + + + +Running +Docker + + + +Feedback +LLM + + + +Record +Pickle + +Bandit selection → factor track or model track + + + + + +every N factors · auto or CLI + + + +Factor track +Hypothesis → FactorCoSTEER +FactorRunner → FactorFeedback +Output: result.h5 +MultiIndex DataFrame +IC / Sharpe metrics + + + +Model track +Hypothesis → ModelCoSTEER +ModelRunner → ModelFeedback +Output: PyTorch preds ++ mlflow logs +LSTM / Transformer / CNN + + + + + + + +Strategy generation pipeline + + +Load top factors +by |IC| + + + +LLM strategy +code gen + + + +OHLCV backtest +signals eval + +Optuna: 10 → 15 → 5 trials · Sharpe ≥ 1.5 · DD ≥ −0.30 · WR ≥ 0.40 + + + + + +Portfolio optimization +Mean-variance · Risk parity · Black-Litterman + + + + + +Live trading (closed-source) +ftmo_live_trader.py · FTMO signals + + +External services + + +llama.cpp +LLM inference + + +Docker +sandbox + + +Optuna +Bayesian opt + + +Qlib +backtest engine + +