Link to paper: here

Link to source: None available

Further steps: It could be an interesting exercise to try to code up this model in Keras/TF.

The authors of this paper use about 17 years of daily SP500 data, and attempts to perform a regression as opposed to a classification (direction) prediction. They split the data into chunks of days, and decide that 50 days is the best chunk size. They normalize the price in the chunk. Their results method is not terribly useful or easy to understand. They compare several variants of their own model, as well as some SVM models.

This paper is useful because it does describe in decent detail an interesting deep neural net architecture consisting of CNN + LSTM, as well as using the useful method of chunking a timeseries and normalizing the chunk. This would probably be more useful if it was converted to direction instead of regression.