Event-Driven Stock Prediction using Deep-Learning

A deep learning method for event driven stock market prediction. Deep learning is useful for event-driven stock price movement prediction by proposing a novel neural tensor network for learning event embedding, and using a deep convolutional neural network to model the combined influence of long-term events and short-term events on stock price movements

Important concept need to be learnt :

NLP - http://web.stanford.edu/class/cs224n
GloVe - GloVe- https://nlp.stanford.edu/pubs/glove.pdf
NLT - www.nltk.org
Convolutional Neural Network - http://www.aclweb.org/anthology/D14-1181

Sentiment Analysis for Event-Driven Stock Prediction

Also try to visualize the data on smaller dimension using PCA to get the good idea of the problem

Use NLP to predict stock price movement based on news from Reuters, we need the following 5 steps:

Collection of Data ( essential and tricky task )

1.1 get the whole ticker list

1.2 crawl news

1.3 crawl prices using urllib2 (Yahoo Finance API is outdated)
Train the GloVe in corpus in NLTK

2.1 build the word-word co-occurrence matrix

2.2 factorizing the weighted log of the co-occurrence matrix
Feature Engineering

3.2 Unify word format: unify tense, singular & plural, remove punctuations & stop words

3.2 Extract feature using feature hashing based on the trained word vector (step 2)

3.3 Pad word senquence (essentially a matrix) to keep the same dimension
Trained a ConvNet to predict the stock price movement based on a reasonable parameter selection
The result shows a significant 1-2% improve on the test set

Use the following script to crawl it and format it to our local file Note : We can relate the news with company and date, this is more precise than Bloomberg News

1 Crawling of Data

python crawler_reuters.py

Yahoo Finanace is also a great place to collect the data

2 Word Embedding

To use our customized word vector, apply GloVe to train word vector from Reuters corpus in NLTK

python embeddingWord.py

To use pre-trained GloVe word vectors

python embeddingWordPre.py

3. Feature Engineering

Projection of word to word vector Seperate test set away from training+validation test, otherwise we would get a too optimistic result.

python genFeatureMatrix.py

Here there is important point to note when we are separating the Cross Validation set and the Training Set. The shuffiling of data can create a very large mistake and untraceble. Consider we have news that are similar in contex but the language of news are slightly different and the got separated in training and the cross validation set. Then the error in cross validation set get biased as the model is already trained against that model so that example will be of no use and effective cross validation set reduces.

4. Training

To train a Stacked-Bidirectional GRU network to predict the stock price movement.

python model_sb_gru.py

To train a ConvoNet network to predict the stock price movement.

python model_cnn.py

5. Model

Model folder contains a trained model with 100K news on Stacked-Bidirectional GRU network. To use the model

cd model
python load.py

Accuracy of model is 96.79%. Furthur improvement like attention mechanism or cyclic learning rate is required.

alexshvid/Event-Driven-Stock-Prediction-using-Deep-Learning