annie's diary

Inspiration

Time series predictions together with features extracted from news.

What it does

It predicts stock mid-price based on historical stock price and current news feed.

How we built it

We use an LSTM network with normalisation for time series predictions, then we use word2vec to map the words to vector spaces, and then use deep ConvNet for the classification of the words, training samples obtained from the google open source repository. Then combine the two signals and train a further deep network for developing further structures between the news and history.

Challenges we ran into

None of us were familiar with NLP problems and methods. Most of us had no previous experience programming using deep learning frameworks such as keras. Too many documentation and API to understand within 24 hours.

Accomplishments that we're proud of

It works, with some minor bugs and some numerical overflow issue.

What we learned

Teamwork! Turn APIs into practical usage. Work hard? Some NLPs, read a few papers and APIs.

What's next for LSTM-SemC-Net

Solve the numerical issues. Optimise the model, current networks compile within 5 seconds, we aim to reduce the running time under 2 seconds. Build deeper ConvNet using residual blocks for sentiment classification. We have not incorporate regularisation into the network hence the model should be overfitting and we need to address this issue. Data size too large, cannot feed them in one run (e.g. for word2vec), need better way for sequential input feed.

Built With

• keras
• python
• tensorflow