This project was developed in cooperation with Katsogiannis Giorgos for the purposes of the Deep Neural Networks course in my postgraduate studies (MSc) in Data Science and Information Technologies (NKUA).
The topic of this research project is centered around the domain of stock prices. In general, our main focus relies on understanding the data and then making predictions. With this in mind, the first part of the project contains the analysis of the time series data. In descriptive modelling or time series analysis, a time series is modelled to determine its components in terms of seasonal patterns, trends, relation to external factors, the correlation between features and other key characteristics of each series. Having obtained an understanding of the data, the second part of this research is dedicated to stock prices predictions. Specifically, we begin by tackling a one-step price forecast problem, by applying classic machine learning methods as well as deep neural architectures . During this regression problem, many experiments were conducted to find which set of features, that represent the data, is the most suitable and of course, which model manages to better fit the series after the fine tuning process. By finding which models generalize better to our data, we proceed to a 60-day price forecast as well. Moving on, we continue with the slightly different problem of the prediction of stock price movements. In this case, the regression problem is converted into a classification problem with two classes and the goal is to predict whether the price of a certain stock will rise or fall. During this set of experiments, many different architectures were applied to find which method generalizes better on this problem.
The dataset that was used to achieve the goals of this project was the CAC40 Stocks Dataset. CAC40 is a benchmark of the French stock market index.
We experiment using both neural models and classic machine learning algorithms as a baseline. Namely, we use Linear Regression, XGBoost and Support Vector Machines (SVM) as our baselines for the regression problem and Logistic Regression, Random Forests and SVM as our baselines for the classification problem. We also use 5 different deep neural networks: LSTM, GRU, Bidirectional LSTM (BiLSTM), Bidirectional GRU (BiGRU) and Linear
The code of the project is distributed to three different Jupyter Notebooks. For ease of use, we have also provided Google Colab links for all Notebooks which automatically load the dataset. In order to run these files locally, the user must also download the CAC40 Stocks Dataset separately.
We provide a Jupyter Notebook containing all the code used for data exploration and time series analysis: Deep_Learning_Project_Data_Analysis.ipynb
Having completed the time series analysis process, the next step of our project is to make predictions on our data. As we’ve already mentioned, this section contains two major experiments: the stock price forecast and the prediction of stock price movements. The Jupyter Notebook containing the code used for the regression and classification experiments is the following: Deep_Learning_Project_Predictions.ipynb
Deep_Learning_Project_Long_term_Predictions.ipynb is a Jupyter Notebook containing the code for the multiple day prediction experiment.
A short presentation with the results of this project can be found here