Time Series Prediction with Recurrent Neural Networks (RNN)
Minh Khoi Nguyen Do, Vuong Ho, Chuqin Wu, Qianwen Fu
Project Page | Paper | Data
Abstract: We look to apply deep learning techniques in the subject of time series analysis and analyze their performance in this type of task, and possibly compare their accuracy against autoregressive methods, which are algorithms specially developed for time series data. Specifically, we look at whether neural networks can be applied in the financial sectors by using a variety of networks such as bidirectional RNN, GRU, and LSTM to predict future prices of Microsoft stocks, Bitcoin, and similar datasets with varying levels of features. We have also prepared and generated artificial time series datasets that simulate reality to see how deep learning can be applied in this kind of task. We find that all four of the networks we experimented with perform relatively well on our datasets and that some networks perform better at one dataset than the other.
This project was done as part of Fall 2021 CSC 240 Data Mining @ the University of Rochester.
The codebase is tested on
- Python 3.8
- PyTorch 1.7.1
External libraries:
- pandas (fast, powerful, flexible and easy to use open source data analysis and manipulation tool)
- PyTorch (An open source deep learning platform)
- scikit-learn (Machine Learning in Python)
For this project, we used 6 datasets, 4 real-world datasets including
- Microsoft stock price (Stock market)
- S&P500 Index (Stock market)
- Crude oil prices (Natural resources)
- Bitcoin price (Cryptocurrency)
along with 2 artificially generated datasets. Note that we have already generated and put them in dataset directory along with the other real-world datasets.
To generate an artificial time-series dataset, we first used the Pandas package to create a dataframe ranging from 2000-01-01 to 2021-12-31, and random numbers generated by Numpy were assigned to the dataframe. Then, the seasonal decomposition method from statsmodels package was applied to the dataframe with a period of 365 days, resulting in trend, seasonality, and residual. The trend was calculated by moving average with a window length of 365, so the first half-year and the last half year will disappear. We extracted the trend data as our final artificial dataset since it is more similar to the normal financial time-series data. A total of 7307 timesteps was generated.
The following figure displays an example of our algorithm in action.
We use 4 deep learning models for our analysis here:
These model implementation are also available in main.ipynb, along with the training driver file. Although we did not separate them into .py file, it can easily be extracted from the appropriate section of the main.ipynb notebook.
Navigate to main.ipynb (the main notebook), install the appropriate libraries, and run the notebook. Everything should be printed out along with metadata of the training.
├── README.md
├── dataset - Contains datasets, which will be used by the main file
│ ├── Artificial_trend.csv
│ ├── Artificial_trend1.csv
│ ├── coin.csv
│ ├── microsoft.csv
│ ├── oil.csv
│ └── sp500.csv
├── docs
│ └── * - For this README
├── paper - Main paper along with source tex file
│ ├── **
│ └── paper.pdf - Final paper
├── prospectus
│ └── **
└── src
├── analysis - Analysis of dataset
│ ├── Artificial_data_analysis.ipynb
│ ├── Bitcoin_Oil_data_analysis.ipynb
│ └── Microsoft_SP500_data_analysis.ipynb
├── datagen - Algorithmic generation of dataset
│ ├── Artificial_series_1.ipynb
│ └── Artificial_trend.csv
└── model
├── implementation - Core implementation of RNN models
│ ├── GRU.ipynb
│ ├── LSTM_wu.ipynb
│ ├── biRNN.ipynb
│ └── simple RNN.ipynb
├── main.ipynb - Main notebook, this is where all the magic happens
└── train.py - Training driver
Any kind of enhancement or contribution is welcomed.
@unpublished{
mkvhcwcfu2021,
title={Time Series Prediction with Recurrent Neural Networks (RNN)},
author={Minh Khoi Nguyen Do and Vuong Ho and Chuqin Wu and Qianwen Fu},
year={2021},
url={https://github.com/goodudetheboy/TimeSeriesWithNN}
}
MIT License
Copyright (c) 2021 Vuong Ho, Minh Khoi Nguyen Do, Chuqin Wu, Qianwen Fu
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.