Predicting-Financial-Time-Series-Data-by-using-Neural-Network

In this project, we aim to predict the close price of sp500 index based on the previous 22 days from year 1950 - 2017.

This work is based on BenjiKCF's work. This result is much better since the models are optimized.

There are 2 models. Neural network and lstm-rnn.

• Filename: NN_ClosePrice_prediction.ipynb

• Filename: LSTM_ClosePrice_prediction.ipynb

NN model result

Train Score: 0.00003 MSE (0.00574 RMSE)

Test Score: 0.00009 MSE (0.00934 RMSE)

LSTM model result

Train Score: 0.00001 MSE (0.00325 RMSE)

Test Score: 0.00006 MSE (0.00765 RMSE)

LSTM has better result than regular Neural network

Future improvement:

1. Sentiment analysis will be added as features
2. More indexes and stocks will be included

Some Notes

1. CuDNNLSTM is much faster than LSTM, in my experience, it's around 5x.
2. Predicting close price is very challenging, both complex model are beaten by naive prediction.

How to set up the environment and train the model

I suggest using virtualenvwrapper, then this project will not affect others

• Set up a virtualenv with python3 (same for python2)

mkvirtualenv --python=/usr/bin/python3 nameOfyourEnvironment

workon nameOfyourEnvironment

• Install all the requirements

pip install -r requirements.txt

• Install tensorflow and keras with gpu enabled

pip install tensorflow-gpu

pip install keras

• Setup ipykernel

pip install ipykernel

python -m ipykernel install --user --name=nameOfyourEnvironment

• Open jupyter notebook, choose nameOfEnvironment kernel