gdax-orderbook-ml

Application of machine learning to the Coinbase (GDAX) orderbook using a stacked bidirectional LSTM/GRU model to predict new support and resistance on a 15-minute basis; Currently under heavy development.

Model Structure (visual):

General project API/data structure:

General Project Requirements

• Anaconda environment strongly recommended
  • see requirements.txt for pip, or environment.yml for Anaconda/conda
    • Jupyter Notebook
    • Python, Pandas, Matplotlib, MongoDB, PyMongo, Git LFS
    • Scipy, Numpy, Feather
    • Keras, Tensorflow, Scikit-Learn
• Python client for the Coinbase Pro API: [coinbasepro-python] (https://github.com/danpaquin/coinbasepro-python)
• CUDA/CUDNN-compatible GPU highly recommended for model training, testing, and predicting

Tensorflow/Keras local GPU backend configuration (Nvidia CUDA/cuDNN)

Local GPU used to greatly accelerate prototyping, construction, and building of ML model(s) for this project, especially considering the nature of the dataset & machine learning model complexity.

• Requirements to run tensorflow with GPU support
• Nvidia GPU compatible with with CUDA Compute 3.0
  • Nvidia CUDA 9.0
  • Nvidia cuDNN 7.0 (v7.1.2)
    • Install cuDNN .dlls in CUDA directory
    • Edit environment variables:
      • C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0
  • pip uninstall tensorflow && pip install --ignore-installed --upgrade tensorflow-gpu
    • Default tensorflow install is CPU-only; install CUDA and cuDNN requirements, then uninstall tensorflow and reinstall tensorflow-gpu (pip install --ignore-installed --upgrade tensorflow-gpu)

Project/File Structure

Latest notebook file(s) with project code:

9_data_pipeline_development.ipynb:

• Development of data pipelines and optimization of data from MongoDB instance to ML model pretraining
• Removal of deprecated packages + base package version upgrade (i.e. Pandas)
• Development groundwork for automation pipeline for automated hourly data scrape, cycling, and training for model through segregated instance or live online-based model
• Usage of in-line markdown cells in-notebook for readability and consistency
• Even further refinement to program structure
• Function scope and structure & function creation for common operations
• Parsing of raw data into 4 separate l2 update (4 consecutive 15 minute l2update segments)

9a_model_restructure.ipynb:

• Notebook used for further development of model in different formats, and testing of reduced complexity models
  • Keras Sequential() Model
  • Keras Functional API
  • Raw Tensorflow

8_program_structure_improvement.ipynb:

• Previous notebook with proof-of-concept output results
  • Several function calls via API and multiple packages required are deprecated
  • Use as reference for updated development files/notebooks

6_raw_dataset_update.ipynb:

• Notebook file used to scrape/update raw_data for both MongoDB and csv format, 1 hour of websocket data from GDAX
  • L2 Snapshot + L2 Updates without overhead of Match data response (does not have Match data; test data has Match data and adds significant I/O overhead)

Folder/Repository Structure

• 'gdax-python' and 'gdax-ohlc-import' are repositories imported as Git Submodules:
  • After cloning the main project repository, the following command is required to ensure that the submodule repository contents are pulled/present: git submodule update --init --recursive
  • .gitmodules file is file for submodule parameters
• 'model_saved' folder:
  • Contains .json and .h5 files for current and previous Tensorflow/Keras models (trained model and model weight export/import)
• 'documentation' folder:
  • 'rds_ml_yu.revised.pptx' is a powerpoint presentation summarizing the key technical components, scope, limitations, of this project.
  • 'design_mockup' folder:
    • Contains diagrams, drawings, and notes used in the process of model and project design during prototyping, testing, and expansion.
  • 'design_explanation' folder:
    • Contains 8 pages of detailed explanations and diagrams in regards to both project/model structure and design.
  • 'previous_revisions' folder:
    • Contains previous/outdated versions of readme documentation and powerpoint presentations documenting the nature of this project
• 'saved_charts' folder:
  • Output of generate_chart() for candlestick chart with visualized autogenerated support and resistance from autoSR()
  • Screenshot of model layer structure in text format
  • Graphviz output of model layer structure
• 'test_data' folder:
  • Only has 10 minutes of scraped data for testing, development, and model input prototyping (snapshot + l2 response updates)
• 'raw_data' folder:
  • 1 hour of scraped data (snapshot + l2 response updates)
    • l2update_15min_1-4: 1 hour of l2 updates split into four 15-minute increments
    • mongo_raw.json: 1 hour of scraped data from the gdax-python API websocket in raw mongoDB format
• 'raw_data_10h' folder:
  • 10 hours of scraped data:
    • l2update_10h, request_log_10h, and snapshot_asks/bids_10h
    • 10 hours of scraped data in raw mongoDB export (JSON): mongo_raw_10h.json
  • Data in .msg (MessagePack) format currently experimental/testing as alternative to .csv format for I/O operations
• 'raw_data_pipeline' folder:
  • Contains data in .feather format as part of data pipeline(s) implementation and development
• 'archived_ipynb' folder:
  • Contains previous Jupyter Notebook files used in the construction, design, and prototyping of components of this project.
    • Jupyter Notebook (.ipynb) notebook files 1-5 & 7
    • Each successive notebook was used to construct and test whether at each "stage" if a project of this kind of scope would even be technically possible.
  • Successive numbered notebooks generally improve and are iterative in nature on previous notebook files for this project.

Misc. Technical Reference

• Git LFS (Large File Storage)
• MongoDB raw export to JSON
• Git Submodules
  • Git submodule ignore = dirty parameter

Publications and Journals referenced for model structure and design

• How to Construct Deep Recurrent Neural Networks
• Training and Analysing Deep Recurrent Neural Networks
• Where to Apply Dropout in Recurrent Neural Networks for Handwriting Recognition?
• Dropout improves Recurrent Neural Networks for Handwriting Recognition
• Speech Recognition with Deep Recurrent Neural Networks
• Recurrent Dropout without Memory Loss
• Deep Stacked Bidirectional and Unidirectional LSTM Recurrent Neural Network for Network-wide Traffic Speed Prediction

License

- gdax-orderbook-ml: BSD-3 Licensed, Copyright (c) 2018 Timothy Yu
- coinbasepro-python: MIT Licensed, Copyright (c) 2017 Daniel Paquin 
- autoSR() function adapted from nakulnayyar/SupResGenerator, Copyright (c) 2016 Nakul Nayyar (https://github.com/nakulnayyar/SupResGenerator)