wjcper2008 / rnn_garch

Estimating Value-at-Risk with a recurrent neural network (Jordan type) GARCH model

Recurrent Neural Net predicting Stock volatility

This repository contains Python code to train a recurrent Neural Network which tries to model the volatility of the daily returns of the SP500 index.

To run the code

Download the repository content by clicking "Download ZIP" and unzipping to a folder on your machine.

Download a Python 3.x interpreter from here. Or to make sure all the neccesary modules are installed in one go, download and install the Anaconda module packages, which also comes with a Python 3 interpreter. The Anaconda package can be downloded here.

When Python (and the appropriate packages) are dowloaded. Simply type one of the following commands in your command prompt:

python train_GARCH.py
python train_RNN.py 
python VaR_GARCH.py
python VaR_RNN.py

The first two scripts estimates the GARCH(1,1)-model and the Jordan Neural network with 5 hidden layers on the SP500 daily returns and saves the output in a JSON-file, named GARCH_est_.json, for the ARCH model and Jordan_est_.json for the Neural Network model.

The two scripts: VaR_GARCH and VaR_RNN produces some VaR plots which are saved in your_folder/plots.

Data

The data come from Yahoo fianance https://finance.yahoo.com/q?s=^GSPC and is located in a CSV file in the data-folder.

Paper

This code was made for a University paper. A draft version of the paper in PDF can also be found in the repository, named RNN_GARCH_paper.pdf.