Bastien Girardet - Bachelor Thesis - Cryptocurrencies volatility forecasting using heterogeneous models based on multi-layer neural network adn classical ARCH family models

Installation

First we need to install pipenv

$ brew install pipenv

or

$ pip3 install pipenv 

Then provided the Pipenv file in the repository, we can install all the requirements by doing in the directory:

$ pipenv install

Scripts are to be run in the correct order

0-generate-graphs

In this notebook we generate all the graphs for this thesis concerning descriptive stats, log returns distribution, historical volatility distribution, etc..

Figures are saved in the folder /figs

1-bruteforce-ts-model (Heavy computation)

This python script bruteforce all arch-type possible models The results are saved in /models/ts/fits

2-forecast_variance_with_garch_model_eval

This file recompute the variance of each models to make forecast on the testing sample Results are saved in /mcs/data/models/TS/

3-compute-mcs-ts

We process the results from script number 2 to find the superior set of classical arch-types models. Results are saved in ./mcs/results/*-top-5*.csv/

4-find_best_model

We get back the results of the models from the metadata saved in script #1

5-forecast_variance_with_ARCH_and_plot

We plot the top models

6-forecast_vol_with_ann (Heavy computation)

Here we compute all combination possible with the different hyperpamaters (e.g. model 5 - Drop out rate 0.4 - Normal - ANN-GARCH(1,1)

7-compute-mcs-hetero

Here we compute the mcs for the heterogenous models computed in the previous script. Results are saved in /mcs/results/crypto-top-5-formatted.csv

8-compute-hetero-and-graph

We plot the top models

Libraries

Cryptov2

Own library made for the purpose for this thesis can be found in libaries/Cryptov2.py Compute various indicators