Code & Data for the stock prediction model in our paper: Modeling the Momentum Spillover Effect for Stock Prediction via Attribute-Driven Graph Attention Networks.

Python 3.7.6 & Pytorch 1.5.1

$ python main.py --device=0

Make sure that the GPU is used to reproduce our experiments.

All the preprocessed data can be found at ./data.

The selected 198 tickers can be found at ./raw_data/stocks.txt

The raw transcational data is about 1.31Gb. Link and preprocess code will be released soon.

The news data are not provided in our repository due to copyright issues.

We use financial news from Reuters and Bloomberg over the period from 2011 to 2013, released by Ding et al. ("Using structured events to predict stock price movement: An empirical investigation." EMNLP. 2014.)

The Loughran-McDonald Master Dictionary (https://sraf.nd.edu/textual-analysis/resources/) is used to extract sentiment from financial articles.

Firm relations can be found at ./raw_data/relations/*. The Company Relations are collected from S&P Capical IQ (https://www.capitaliq.com/). The first row is the target stock tickers, and following rows are firms that has specific relation with that frim.

[email protected]