Sad, this thread looked promising but has by now completely been taken over by haters or know-nothings. Well to each his own, I bow out...but thanks to those who provided valuable content.
Sorry, I should have more precisely qualified my comments as the discussions involves, NN, Deep Learning Networks and a few other methods. My experience is with Support Vector Machines and prior to that with the old fashioned NN. Have yet to work with DL networks so feel free to discount my comments on that basis.
I attended Nvidia's GPU conference here in the SF bay area a couple of months ago. There were atleast 2 presentations related to deep learning & trading. https://mygtc.gputechconf.com/form/session-listing&doSearch=true&queryInput=&topic_selector=Finance Interestingly both are from Japan. Once person was talking about a system they have developed that can tell us if the market is going to go up or down by .1% over the next 1 hr. They have a 32 M40 GPU server farm. They claim that their system achieves a reasonable (70%+) accuracy during the trading day. They feed in the one minute candles of all the markets along with one min snapshots of the order book. The second presenter talked about a charting tool he has been working on that a user (trader) can use to automatically train his strategy... for example, the trader can mark parts of a chart and feed if into the training module... the idea here is that the trained model should be able to identify such patterns in future so that the trader can trade accordingly. Both these presenters talked about RNNs and RBNs along with some time-series techniques. I was not exposed to too many deep learning based trading ideas other than my own imagination of how things can be done... also after talking to a few folks in the company that I work for (focused on deep learning), I got an impression that deep learning is an emerging field that can be tapped into. The chief scientist at my company thinks that all the ticks/candles in the world are not sufficient to train a proper deep net... we would need a lot of data to increase the probability to about 90%+ accuracy; for example, google runs atleast 3 million images of a cat into a CNN to identify a cat from a photo... we need a very large dataset for such training... and so was his reasoning.... I think he is wrong since he doesn't know trading . Deep learning is definitely a tool that increases your alpha and it becomes more and more prevalent by the day. I think at the end of the day, its the what you want your NN to train for and the backpropogation techniques that matter. To be frank, I looked into Deep learning a bit, and decided to put it in the backburner for now. Probably get back to it in a few months...
Rohan, thank you for this most interesting post. What was name of the group with 70% accuracy please.
@conduit I know how you feel! But if we develop a thick skin, and stay focused on subject, I would really value your continued input...
https://mygtc.gputechconf.com/form/session-listing&doSearch=true&queryInput=&topic_selector=Finance check out "S6589 - Algorithmic Trading Strategy Performance Improvement Using Deep Learning" Its not a consistant 70%+ through out the day... their probability increases after trading starts; touches 80% an hour later and gradually subsides over the day. They claim that their system is still in beta state and as I understood, Mizuho securities are trying to sell this system to banks and hedge funds... who knows what the internals are, but thats what they claimed...
Well of course all the best stuff is totally free.....why on earth would a company that develops top line software ever think of charging money for it? Huge global corporations like DOW, Ford and Abbot Labs are saving tens of millions on software after they discovered free software in 2014 and stopped paying SAS, IBM and others for software.
Yep, a sure sign they can make more money selling it than they can trading it. A bit more expensive version of the special DVD of 30 Secret Trading Strategies that are guaranteed to make you rich for only $299.95 plus shipping.
On the 3 million cat images...keep in mind that if one had a somewhat magnified photos of the cat's whiskers clearly showing wisker attachment area in the face to tip a not too sophisticated model could likely identify a cat from a dog, racoon, etc from whisker length variation, density, angle to horizontal, and 6 other things using perhaps 100 photos. This is the real goal for trading...finding the 30 most predictive variables out of many thousands.