Final project of NTUST CS5144701 Practices of Deep Learning course.

The goal of this project is to improve the performance of this paper by using RNN-LSTM model instead of EB-CNN for prediction.

Download jar from the Reverb Website Prepare newsdata.txt and run:

java -Xmx512m -jar ./test-reverb/reverb.jar reverb_pre.txt reverb_result.txt

• News dataset: financial-news-dataset includes 450,341 news from Bloomberg and 109,110 news from Reuters within the date range from 2000 to 2013.

• Stock price dataset: alphavantage API.

• Reverb
• LSTM
• CNN
• Stanford-OpenIE-Python

1. process
2. reverb & local command
3. word2Vec
4. main

1. Decide using which dataset, including stock data and news data (2018-05-20).

2. To know how to use reverb to split one news title sentence (2018-05-25).

3. Process the news data: got the title and time (2018-05-27).

4. Process the stock data: got the price (2018-05-28).

5. Reconstruct the stock price data, only save the result that compared with last day price.If today's price is smaller than last days, it will be saved on 0 and other situation will be saved on 1 (2018-05-29).

6. Using the Reverb technology to split a news title sentence of all news's data to the three part format (2018-05-29).

7. Generate the word_embedding model, using reverb-result (2018-05-30).

8. Generate embedding_matrix, using Keras's tool (2018-05-30).

9. Construct the neural network model(Event-LSTM) (2018-05-30).

10. Train and predict (2018-05-30).

11. If the predict's result dont bigger than 65%, to find and modify the errors (2018-06-01).

12. Generate the final project documents and slides (2018-06-18).