Inspired from https://github.com/fulifeng/Temporal_Relational_Stock_Ranking

Changes

1. Migrated from TF1.x to compat.v1 API symbols via tf_upgrade_v2
2. Added training checkpoints under ./training_v2/.checkpoints
3. Replaced Const Ops with Placeholder to reduce GraphDef (which saves values form Const Ops) size, so it can be loaded to Tensorboard via exported events (FileWriter) or .pb files.
4. Added summaries to tracking losses from training, validation and testing

Setup

pipenv install --skip-lock -d

Run

Use training_v2 folder to run with Tensorflow 2.x.

python relation_rank_lstm.py -m NYSE -l 8 -u 32 -a 10 -e NYSE_rank_lstm_seq-8_unit-32_0.csv.npy

Tensorboard

tensorboard --logdir ./training_v2/logs

Notes

• tf.compat.v1.Saver used for training checkpoint instead of CheckingPoint due to current optimizer (AdamOptimizer) is not trackable
• The original implementation used tf1.const for relation and rel_mask ops, which is saved to GraphDef and too big to load to Tensorboard. To circumvent, we can extract GraphDef from exported model by stripping the tensor_content field of graph node and then manually upload to Tensorboard via following command. However it is still better to replace tf1.const to tf1.placeholder

  pipenv shell
  python -c "from utils import save_strip_pbtxt; save_strip_pbtxt('./model_outputs/NYSE/{#}')" 
  

For details, see below orginal Readme.md

Orignal README.MD

Code for the Relational Stock Ranking (RSR) model and the Temporal Graph Convolution in our paper "Temporal Relational Ranking for Stock Prediction", [paper].

Environment

Python 3.6 & Tensorflow > 1.3

Data

All data, including Sequential Data, Industry Relation, and Wiki Relation, are under the data folder.

Sequential Data

Raw data: files under the google_finance folder are the historical (30 years) End-of-day data (i.e., open, high, low, close prices and trading volume) of more than 8,000 stocks traded in US stock market collected from Google Finance.

Processed data: 2013-01-01 is the dataset used to conducted experiments in our paper.

To get the relation data, run the following command:

tar zxvf relation.tar.gz

Industry Relation

Under the sector_industry folder, there are row relation file and binary encoding file (.npy) storing the industry relations between stocks in NASDAQ and NYSE.

Wiki Relation

Under the wikidata folder, there are row relation file and binary encoding file (.npy) storing the Wiki relations between stocks in NASDAQ and NYSE.

Code

Pre-processing

Script	Function
eod.py	To generate features from raw End-of-day data
sector_industry.py	Generate binary encoding of industry relation
wikidata.py	Generate binary encoding of Wiki relation

Training

Script	Function
rank_lstm.py	Train a model of Rank_LSTM
relation_rank_lstm.py	Train a model of Relational Stock Ranking

Run

To repeat the experiment, i.e., train a RSR model, downloaded the pretrained sequential embedding, and extract the file into the data folder.

NASDAQ

python relation_rank_lstm.py -rn wikidata -l 16 -u 64 -a 0.1

NYSE

python relation_rank_lstm.py -m NYSE -l 8 -u 32 -a 10 -e NYSE_rank_lstm_seq-8_unit-32_0.csv.npy

to enable gpu acceleration, add the flag of:

-g 1

Cite

If you use the code, please kindly cite the following paper:

@article{feng2019temporal,
  title={Temporal relational ranking for stock prediction},
  author={Feng, Fuli and He, Xiangnan and Wang, Xiang and Luo, Cheng and Liu, Yiqun and Chua, Tat-Seng},
  journal={ACM Transactions on Information Systems (TOIS)},
  volume={37},
  number={2},
  pages={27},
  year={2019},
  publisher={ACM}
}

Contact

[email protected]