I am still new to this but my understanding is that CNN is in fact process that starts with data filtering, feature selection and data reduction in the process of convolution and pooling. Then the NN parts kicks in and does its magic. Is the first part on CNN really responsible for its better performance than other types of NN? Why would it need extra data augmentation before that and what would be the advantage of doing it?
Well you have to ask this to the people who constantly emphasize the "cleansing, filtering, and feature selecting" as being more important than the actual network. And a NN does not "kick in and it does not do its magic". It does exactly what it is prescribed and designed to do. For anyone who is looking for shortcuts in life and trading in particular, Deep Learning Networks are not the ones one wants to embark on, it is a huge losing proposition to approach neural networks without the required level of skill and knowledge sophistication to grasp the underlying stats and math. Of course kids can play around with neural networks all day long and pretend they are cool but I would attach a probability of zero of them finding anything useful.
I don't use any type of NN for trading, nor do I plan to. However, whatever "performance" you'd like your trading system to have will have to be designed in from you. This is usually not a top-down waterfall approach, but an iterative process of trial and error where you have to rework all parts on all levels until you start seeing some results that are logical and supported by some kind of trustworthy verification. The entire system, your trading engine, will have to work in concert as a whole in order to establish any kind of consistent returns (which can be one of many aims for creating trading systems). Realistically, there are infinite possible systems and no way to know which approach will work at all or be more optimal than others. So to cut down on all the possibilities, you''ll absolutely need to use your own neural network for a while.. Until anyone shares any NN type of systems for trading, there's no real basis for debate at all, and is why the signal to noise ratio is so low here.
Good questions. "Compared to other image classification algorithms, convolutional neural networks use relatively little pre-processing. This means that the network is responsible for learning the filters that in traditional algorithms were hand-engineered. The lack of dependence on prior knowledge and human effort in designing features is a major advantage for CNNs." https://en.wikipedia.org/wiki/Convolutional_neural_network#Overview See also: https://en.wikipedia.org/wiki/Convolutional_neural_network#Distinguishing_features As someone pointed out, it is difficult to get a straight answer (even with a straight question) here.
With the way 'the crowd' attacks (or stands idly by) people that post such things...I doubt you'll see much of that.
I think you have already heard several people sharing their thoughts and specific setups. You can't realistically expect details to be divulged as people put a lot of work and efforts and skills into the inner workings of their methodologies. I believe if those who actually work on deep learning networks keep on sharing general thoughts and ideas and those who have no experience in this field (= noise) keep a little quieter rather than complaining or putting up unsubstantiated claims (such as that the input data is much more important than the actual algorithms that peruse such data) then the quality of this thread would indeed be way higher. I believe if one really scanned through the posts and filtered out the 90% noise there are still quite a number valuable information out there. You would never, for example, hear from a professional firm that they used classification algorithms for research on financial data and time series, yet I can assure you that a number of firms find such approach highly rewarding and promising. You can't expect, and I repeat myself, that those who make a living from trading to post what works for them to the lazies of their world. I think a good noise filter gets you quite far like with every other thread on this site.
Are you suggesting that somewhere, in that large collection of threads, that if I look hard enough, I'll find where someone has posted their NN, or the signals from their NN, for a week?