I think you missed the point. You cannot directly trade with Neural Networks. Funds and individuals that use Neural Networks use them for tasks like portfolio optimization, and NOT directly trading.
Several neural networks are at the core of how I put food on the table. They work, but you need to know what you're doing. You also need to keep your eye on them... Those who say they dont work are probably the same ones who are simply trying to predict the prices of the next "bar" in a point wise fashion based on a sequence of the last bars. You're trying to characterize and control a highly stochastic dynamic process.... If you approach NNs for trading without this in mind, you've probably already wasted your time.
Yes, but I do not implement traditional neural network models. I work primarily with DNN sequence models. RNNs and GRUs and am currently exploring some very new approaches in academia as well. I never found value in Boltzman Machines, k-nearest-neighbors or the like as applied to financial trading algorithms.
Then you are not well informed. I myself use and I know of several shops that use DNNs to derive multi factor models that make valuable volatility predictions, and forecasts of velocity/momentum. Some of those models are as simplistic as classification algorithms, others orders of magnitude more complex. The current bottleneck is not finding valuable approaches but the performance of inferencing and running the algorithms in production as applied or high frequency trading. For anything of higher latency computational performance is way sufficient
Curious; as I understand (correct me), matlab coder provides a way to convert DNNs into cuda C++ code; that you can later compile & link into your trade signal generator. Doesn't that solve the latency requirement if you are trading within the 500us - 1ms range? Yes, under 10us is a different ball game. Wondering where you are drawing the line between what is acceptable and what is not in relation to production performance & latency when you use DNNs? I myself use a python inference engine that runs in parallel to a C++ trading system. After some back and forth communication, the system places a trade. I see a 5ms latency; which is acceptable for now, but may not be shortly.
So that's where it comes from. Interesting. I kept trying to figure it as some sort of acronym without any vowel hints, but it is a NN thing? K, I'll stop trying to figure it out now, haha. (I got so exhausted working on the issue in my head, I started settling on some weird form of "I am Groot.") I am now free from the bondage of trying to figure out your nick.
I will probably refrain from stating specific numbers as latency requirements are incredibly sensitive to the exact specs of each strategy approach. I do not know what Matlab does because I think it's not the ideal tool to perform DNN research but as long as the trained network with all its weights can be transfered into C++ then it will satisfy quite a number of hft algorithms, latency wise.
I will not Google that at this time, since I am shorting into a bull market, because I THINK I know better than everyone else. :-(
This logic or joke or both eluded me. And why do you short into something without price confirmation? You might get it right once for each 8 or 10 times you got it wrong, guessing the market. Not a great track record ofr success.