Well....but keep in mind that an indicator is a compression of information with extreme prejudice......somebody decided that there are 32 candlestick patterns, or 4 minor Elliott waves in a Sorbinsky retracement or a MA with lengths of A, B and C as opposed to x, y and z. We should hold a memorial service for the thousands of units of information arrogantly discard in the process of creating famous TAs.
Well again, would you give away your secret sauce in a contest? I would not and I do not. I employ a deep learning network to predict correlations and momentum/mean reversion properties and I get sufficiently good results to make the work that was put into this justifiable. And I never claimed making money trading is easy. I said utilizing time series data and defining discrete outcomes using a classification algorithm is easy. In the end you draw your own conclusions, I simply pointed to the fact that a simple classification algorithm can indeed be fed with time series data and can produce reliable and stable results. Obviously a lot depends on the input data, the labeling of the input data, the design of the network, hyper parameters, and many more. If this was all easy then the results AlphaGo produced would not be such remarkable.
Again, it depends on where you believe the information content lies. If not in tick data it is absolutely prudent, and probably a must, to discard most every tick and only sample prices each minute, only, or utilize every 100th tick. A moving average in fact does nothing else. I never said that common type indicators may add value, I merely said that sometimes eliminating data points, where one believes no valuable information reside, makes a lot of sense. You can't easily operate with networks that take as input millions of perceptrons.
Sure, but to a major degree your preferred time granularity defines that....if you are trying to predict Wheat prices two weeks into the future, then using tick data would be silly.
I am glad we agree that it still requires a thinking and experienced human to define and structure neural networks
Of course....if fact prior to the availability of quantum computing, information conceptualization is just as important as your predictive software. The variety there is truly interesting ranging from adaptation of quantum physics and string theory to scanning social media for word mentions. Is the market a self referential and quasi-aware ecological system or just the sum of the actions of a few millions humans and independent computers at any given instant? How long does information persist in the system: does an event response structure in a given instrument that occurred in late 2009 effect price unfoldment in 2016 or is market memory determined by how many bars can be viewed on very large monitors screens at the hedge funds? You have to decide a few dozen things like this before ever touching your neural network software.
I too looked at deep learning sometime back, but decided to get my normal hand written strategies into production before venturing into deep learning based indicators. Do you see any advantage in opting for deep learning based algorithms compared to having something that we know how it works? deep learning based indicators/algorithms are black boxes in some sense... we don't know exactly how they logically work although we can trace back their weights & calculations. RNNs are notorious to train and how do you make sure that you your signals are exactly what you expect... any thoughts/ideas...
A NN or SVM overcomes the human limitation of 7. According to cognitive psychologists humans make decision best when looking at about 7 or fewer considerations.When presented with 40 things to think about the brains starts to group, summarize or ignore some until it gets them to a handful for a decision. This is the reason TAs were invented: to summarize and hide too much information. My SVM can happily process 3000 columns of data input. No need to hide, summarize or ignore anything.
Glad you mentioned AlphaGo. That project was one of the reasons I am currently looking at deep learning for trading. You seem to have taken the approach I am interested in - using the NN to determine/define the inputs that add value to a strategy. Would you say that the way you have developed this, you rely less on the need for a human to select the input to your model?