Cross-timezone, business-wide share price prediction using various recurrent architectures.

The goal of the project was to investigate whether stock quotes for companies from the same area, but earlier timezone carry decent predictive power if compared to the company itself or all companies from its region. Code in this repository allows the user to download data for a chosen business and then experiment with several RNN-based architectures.

This guide assumes you have Python 3.x installed on your computer. Clone the repository to a desired local directory

git clone https://github.com/jkoscialkowski/gsn-projekt.git

Then cd to this folder, create and activate a virtualenv and run

pip install -r requirements.txt

Afterwards run

python3 amphibian/fetch/downloader.py -T banking -R REGION -B banking

to download the data. You need to substitute REGION with all of the following, one-by-one: ASIA_PACIFIC, EMEIA, AMERICA.

Finally, you can run one of the main*.py files to start cross-validation. For example to train the Attention network run

python -m main_attention.py 

This will yield .csv files with Cross-Validation results. The best set of hyperparameters can then be used to train a final model which will be used for generating insights.

Amphibian comes with the amphibian.visual module which implements:

• a nice class for confusion matrix display (with precisions, recalls and accuracy presented),
• Model-Agnostic Variable Importance which computes permutational importance for a given company and shows the results on a barplot.

We advise playing around with a trained model and the visualisation tools implemented in the said module.