Forecasting stock prices with a feature fusion LSTM-CNN model using different representations

"The Scientist and Engineer's Guide to Digital Signal Processing," Steven W. Smith
Free at http://www.dspguide.com/

"Mathematical Techniques in Financial Market Trading," Don K Mak


"Cycle Analytics for Traders," John F. Ehlers

 

Here is a not-too-old reddit post that I've missed until it got deleted, but talking about how useless those research papers are, based on the comments:


And more comments about it:
https://news.ycombinator.com/item?id=20720095

And:


And this seems to be the original post, although not sure why someone replaced "predicting the stock market" with "predicting bitcoin":

 

Sounds about right

 

Another guy using deep learning to predict the markets. Interesting setup regarding the plots.

 

how did the author use image to predict price? in your another post in this thread, the author use BULL bar and Bear bar to train the data set, and predict if the next bar will be bull or bear bar. how could the author use image to predict price? I did not get it. what kind of image to feed into the net? the candlestick chart only shows up and down bars.

 

You have to understand how CNNs work. An OK book (if you know Python) is Grokking Deep Learning.

 

Very simple. I saw some trader asking many years ago how he can save images of various charts and then use the computer to learn how to recognize each setup and alert him when similar chart/setup happens again. Such ideas were discussed many times and may be 20+ years old.
And many people tried to fully automate this and create trading programs that would recognize images of various charts and use them as trade signals. So it's only natural to use AI for this.
Candle stick charts with up and down bars are sufficient because the idea is to allow the computer to look at the same charts that humans look at, and make similar decisions. Though there may be additional charts or data that a computer may look at, same/whatever humans look at.

Though I don't believe it can work for variety of reasons. Or at least it would need to be evolved to level of complexity that would exceed doing this without charts, just using pure data that is displayed on charts anyway.

 

if you have a pattern that is profitable, you can train NN with thousands of this pattern and recognize the pattern when it occurs in out-of-sample data sets. the question is: the author even does not know which pattern is profitable, how can he use computer to recognize winning patterns?

 

You don't need to have a good idea about which pattern is profitable. You only need to know that an image or a sequence of images is such that the following image or sequence of images allows (when traded) to generate some profit or loss. Basically, x={an image or a series of images}, y={realized profit if a trade is initialized}. Then you end up with a regression of a classification problem, e.g., in case of classification you have a few classes where each class corresponds to a particular level of profitability or loss for a given set/series of input images.

If you use CNNs you would probably need to wisely pre-process your images to benefit from convolutions.

 

Training with historical data. They wrote "We will predict the stock price after five minutes by looking at the data for the previous 30 minutes based on a minute-by-minute current time point".
Though really this can't even be a question because main part of this task/project is to find out which patterns are profitable before using them. This is what AI is training for.