Keras_MLP-GRU_Regressor

Introduction

This is my contribution for the master's degree thesis. In this project I present two types of nonlinear regressors for predicting time series. The first one is implemented with a classic Multi Layer Perceptron Neural Network (MLP). The second one is implemented with a Recurrent Neural Network, Gated Recurrent Unit (GRU) scheme. At first I focused only on Forex price time series and later extended my interest in Crypto / Fiat price time series.

The purpose of this repository is to show an easy way to create a good dataset, to train the desired model (MLP or GRU) and to use it in real time forecasting.

Requirements

If you have a GPU with CUDA support I highly recommend you install it. It is not mandatory. If you do not have this kind of possibility, in the code I have indicated which lines you need to comment to correctly run the program. To set up all necessary softwares follow:

• See this guide to install CUDA
• Download and set up CUDA-Toolkit To correctly execute the program we need to install and set up:
• Python3, Tensorflow and Keras. I found a great presentation made by Jaff Hauton using Anaconda, see the guide. If you don't want install Anaconda, type the following commands (Debian/Ubuntu):
  • $ sudo apt-get update
  • $ sudo apt-get install python3.8
• verify python and pip installation:
  • $ python --version
  • $ pip --version
• If the installation is correctly done type:
  • $ pip install --upgrade tensorflow
• If you have installed CUDA, you can use tensorflow-gpu:
  • $ pip install --upgrade tensorflow-gpu
• PyWavelets is a useful python package needed to apply wavelet transform, see pywt
  NOTICE: pay attention that your GPU drivers, CUDA-Toolkit and tensorflow-gpu have a compatible version. In general you need a python version greater than 3.4.X to execute the program.

The program

The project is composed by some python scripts that allow the user to clean data, to construct a dataset, to train a neural network model and to use it. Starting from the Tick prices movements (i.e. any price movement that has occurred), it constructs a 1 minute price movements (one can choose if use Open, Close, High or Low price). Depending on the price choosen, the program will create two time series. The first one will be a sequence of the prices, the second one will be a smoothed version of the same sequence. After the data is ready, the script start to train the neural network. One can choose if train the Multi Layer Perceprton or the Gated Recurrent Unit model. After the network is trained, one can execute the prediction.

Program description

• The Data folder contains four folders:
  • Raw_Tick_Data: containes some .csv files that hold the raw tick prices. The files come from Dukascopy broker, and are downloaded from JForex platform. In the folder you can find the quotations of Bitcoin
    against US dollar (BTC/USD) related to the January and February 2018.
  • Normalized_1Min_Price: contains some pickle files with the normalized price registered into 1 minute.
  • Normalized_1Min_SmoothedPrice: contains some pickle files with the normlized prices smoothed by PyWavelets registered into 1 minute.
  • Input_Data: contains a .csv file, the input example for the trained network. It is needed to real time usage.
• The MLP_Model_Training and GRU_Model_Training folders contain both:
  • a file of configuration for the hyper parameters needed by the network (each model has its own configuration)
  • a file that contain the model itself. In the files there is the DataSet construction and the model training.
• The PrepareData folder contains two scripts which, starting from the .csv files contained in Data/Raw_Tick_Data, populate the folders Normalized_1Min_Price and Normalized_1Min_SmoothedPrice.
• The Trained_GRU and Trained_MLP folders contain the trained models (the .h5 file), and two pickle files min_val.pickle and max_val.pickle that are the max and min values related to the data used for the training. These values will be used to normalize the data for real time usage.
• The Use_MLP_Trained_Model and Use_GRU_Trained_Model folders contain scripts to load and use the trained model for real-time forecast.
• The Utils folder containes some file with utils function and some configurations needed by the project.
• The script setup.sh is the executable, responsible for run all the script needed to clean data, create dataset and train the model.
• The script execute_model.sh is the executable to load the trained network and make real time prediction.

How it functions

Download the repository and move the main directory Keras_MLP-GRU_Regressor. Open here the terminal, then run:

• $ ./setup.sh
• $ ./execute_model.sh

The first script will create all the the data needed (normalized data, trained model), then it will perform the training and the validation over the desired neural network. The second script will execute the model trained to perform a real time forecast.
NOTICE: The purpose of this project is for demonstration only. Before using it in a real context it is necessary to structure the real-time forecast data in another way (retrieve data from api, query a DB and so on ...), but it is beyond my intentions here.