QuantPipeline

• FrameWork
  • Data
  • Math-Models
  • Simulation
    • model
    • non-model
      • Back-Test
  • Strategy
    • Cross-Section Factor(Method)
    • Time-Series Factor(Method)
    • Agent
      • Factor Signal Rules
      • Risk Signal Rules
    • Optimization Method
      • gradient
      • non-gradient
  • Portfolio
    • Optimization
      • Quadratic Programming
  • Analysis

Reference Package:

• pytimetk:https://github.com/business-science/pytimetk | easy for analysis and visulization
• merlion:https://github.com/salesforce/Merlion,machine | Automatic machine learning for time series
• darts:https://github.com/unit8co/darts | Automatic machine learning for time series
• Nixtla:https://github.com/Nixtla | Automatic classicial/machine learning model for ecosystem time series
• pyod:https://github.com/yzhao062/pyod | Outlier and Anomaly Detection
• pgmpy:https://github.com/pgmpy/pgmpy | Python library for Probabilistic Graphical Models
• pomegranate:https://github.com/jmschrei/pomegranate | Pytorch of PGM
• daft:https://github.com/daft-dev/daft | Probabilistic Graphical Model (PGM) visulization
• pyro:https://github.com/pyro-ppl/pyro | Pytorch for Probabilistic programming languages (PPLs)
• pymc:https://github.com/pymc-devs/pymc | Probabilistic programming for bayesian statistical model focusing on advanced Markov chain Monte Carlo (MCMC) and variational inference (VI)
• TorchRL:https://github.com/pytorch/rl | reinforcement learning
• acme:https://github.com/google-deepmind/acme | reinforcement learning
• RLlib:https://docs.ray.io/en/latest/rllib/index.html | multi reinforcement learning
• FinRl:https://github.com/AI4Finance-Foundation/FinRL | reinforcement learning for finance
• FinGPT:https://github.com/AI4Finance-Foundation/FinGPT| GPT for finance
• hugging face:https://huggingface.co/docs | LLM
• OpenRLHF:https://github.com/OpenRLHF/OpenRLHF
• tianshou:https://github.com/thu-ml/tianshou | reinforcement learning
• dash:https://github.com/plotly/dash | interation and visulization
• lightzero:https://github.com/opendilab/LightZero | deep reinforcement learning + Monte Carlo Tree Search
• crewAI:https://github.com/crewAIInc/crewAI

• alphagen:https://github.com/RL-MLDM/alphagen
• AlphaForge:https://github.com/DulyHao/AlphaForge
• llm4ad:https://github.com/Optima-CityU/llm4ad
• LLM:https://github.com/huggingface/open-r1
• PyGAD:https://github.com/ahmedfgad/GeneticAlgorithmPython | Genetic Algorithm in Python
• Multi-Agent Graph Framework (LangGraph & LangFlow & AutoGen):https://github.com/langchain-ai/langgraph & https://github.com/microsoft/autogen
• ragflow:https://github.com/infiniflow/ragflow | Retrieval-Augmented Generation designed by flow gui chart
• FlashRAG:https://github.com/RUC-NLPIR/FlashRAG |
• https://github.com/AI4Finance-Foundation/FinRobot
• https://github.com/TradeMaster-NTU/TradeMaster | model zoo: FinAgent
• LLaMA-Factory:https://github.com/hiyouga/LLaMA-Factory | LLM models factory
• FastChat:https://github.com/lm-sys/FastChat | platform for training, serving, and evaluating
• llama-index:https://github.com/run-llama/llama_index | agent framework
• OpenHands:https://github.com/All-Hands-AI/OpenHands
• Openmanus:https://github.com/FoundationAgents/OpenManus
• langmanus:https://github.com/Darwin-lfl/langmanus

vllm:

# embedding server
vllm serve Qwen/Qwen3-Embedding-0.6B \
--served-model-name Qwen3_Embedding \
--override-pooler-config '{"pooling_type": "CLS", "normalize": true, "enable_chunked_processing": true, "max_embed_len": 1048576}' \
--task embedding \
--model-impl transformers \
--host 192.168.110.11 \
--port 8888 \
--api-key 123456 \
--trust-remote-code

# chat server
vllm serve Qwen/Qwen3-4B-Instruct-2507 \
--served-model-name Qwen3_Chat \
--chat-template {path} \
--task generate \
--max_model_len 4096 \
--model-impl transformers \
--host 192.168.110.11 \
--port 8889 \
--api-key 123456 \
--trust-remote-code

Ideas

• 2025/6/25
  • 统计方法产出的只能是【相对定位因子(横截面)】或【时序因子】
  • 统计目的要不是relative pricing就是front running
  • Signal按由简单->完备演变: Direction -> Direction,Strength -> Distribution
  • Signal必须依赖锚定物,锚定物要不是未来时刻的自身,就是同时刻其他标的
  • Signal System要不是portfolio weight就是个series decision system
  • 如果特征能被稀疏表述，例如字典学习/PCA/clustring之类，那就能根据分块计算structure break的可能性，作为label
  • 如果特征能被稀疏表述，那time series同样能被表述为有限马尔科夫过程，为预测提供更多思路
  • GPT的训练集中于token预测，如果GPT的训练更大的chunk的预测，例如预测一个段落然后用RAG计算得分（视为一个强化学习过程），那是否推理能力会更强
  • Agent: Memory(RAG), reflection, debate(Multi-Agents), prompt(COT, TOT)