Price graphs: Utilizing the structural information of financial time series for stock prediction

This repository is the official PyTorch implementation of the experiments in the following paper:

Junran Wu, Ke Xu, Xueyuan Chen, Shangzhe Li, Jichang Zhao. Price graphs: Utilizing the structural information of financial time series for stock prediction

Information Sciences

arXiv

Installation

Install PyTorch following the instuctions on the official website. The code has been tested over PyTorch 1.7.0+cu110 version.

Then install the other dependencies.

pip install networkx pyunicorn

Data preparation

• Unfold data

tar -xvzf data.tar.gz

• Change directory to code

cd code

• Transform stock prices to graphs

python price_graph.py

• Calculate collective influence (CI) for graph nodes

python price_ci.py

• Price graph embedding

python price_embedding.py

• Generate dataset for train and test

python dataset.py

Trainer Usage

Train:

usage: trainer.py [-h] [-e EPOCH] [-b BATCH] [-ts TIMESTEP] [-hs HIDDENSIZE] [-y YEARS [YEARS ...]] [-sn SEASON] [-dr DROPRATIO] [-s SPLIT] [-i INTERVAL] [-l LRATE] [-l2 L2RATE] [-t]

Train the price graph model on stock

optional arguments:
  -h, --help            show this help message and exit
  -e EPOCH, --epoch EPOCH
                        the number of epochs
  -b BATCH, --batch BATCH
                        the mini-batch size
  -ts TIMESTEP, --timestep TIMESTEP
                        the length of time_step
  -hs HIDDENSIZE, --hiddensize HIDDENSIZE
                        the length of hidden size
  -y YEARS [YEARS ...], --years YEARS [YEARS ...]
                        an integer for the accumulator
  -sn SEASON, --season SEASON
                        the test season of 2019
  -dr DROPRATIO, --dropratio DROPRATIO
                        the ratio of drop
  -s SPLIT, --split SPLIT
                        the split ratio of validation set
  -i INTERVAL, --interval INTERVAL
                        save models every interval epoch
  -l LRATE, --lrate LRATE
                        learning rate
  -l2 L2RATE, --l2rate L2RATE
                        L2 penalty lambda
  -t, --test            train or test

An example of training process is as follows:

python trainer.py -e 1000 -l 0.001 -hs 32 -ts 20 -b 256 -dr 0 -i 50 -s 30 -l2 0 -y 2017 2018 -sn 1

An example of test with existing models is as follows:

python trainer.py -e 1000 -l 0.001 -hs 32 -ts 20 -b 256 -dr 0 -i 50 -s 30 -l2 0 -y 2017 2018 -sn 1 -t --test_2019

Base Code Repo

VG algorithm is adopted from https://github.com/pik-copan/pyunicorn

CI algorithm is adopted from https://github.com/zhfkt/ComplexCi.

Struc2vec implementation is adopted from https://github.com/shenweichen/GraphEmbedding.

DARNN implementation in PyTorch is adopted from https://github.com/ysn2233/attentioned-dual-stage-stock-prediction

Citation

If you use this code for you research, please cite our paper.

@article{wu2022price,
	title = {Price graphs: Utilizing the structural information of financial time series for stock prediction},
	journal = {Information Sciences},
	volume = {588},
	pages = {405-424},
	year = {2022},
	issn = {0020-0255},
	doi = {https://doi.org/10.1016/j.ins.2021.12.089},
	url = {https://www.sciencedirect.com/science/article/pii/S0020025521013104},
	author = {Junran Wu and Ke Xu and Xueyuan Chen and Shangzhe Li and Jichang Zhao},
	keywords = {Stock prediction, Complex network, Time series graph, Graph embedding, Structure information},
}