Based on Keras deep learning LSTM model predicts gold main closing price
TensorFlow the cycle of learning neural network (RNN)
Preface
This article based on the deep learning model, on the poly width by JQData financial data analysis, this paper focuses on the stock data, but in theory can be used in other ways, as for the wrap theory part, anyone guess,
Background
Quantitative trading in the stock market crash back the blame, but the trend of The Times, the future is just a matter of time, the various quantitative gods have their own methods, such as traditional catch trends (trend following) or technical indicators, high-frequency (high frequency), fundamentals, news, policy surface analysis (event driven), so that these factors have become the factor, comprehensive analysis, etc., the latest areas in the field of AI, now from the return of the ordinary, to all sorts of deep learning model, to the more recent AI play chess, play interstellar train yourself to look for regular bring human miserably abused,
Deep learning
For a-share quantitative, under the limited economic conditions and data resources (wide platform and JQData) personally with deep learning model is more appropriate, for the following reasons:
The simple regression is too limited, subjective choice of ingredients is too high
Use Qlearning similar reinforcement learning, is likely to need a lot of the work force to achieve better effect of
Online learning post often seen with deep learning to predict stock price above, I personally think this direction is to use simple way to deal with difficult problem, here had to tie up the inside of the theory of some conclusions:
The absolute price forecast was impossible
Prices are following a similar to the shape of the nested interval level structure run
Described in a computer, the father of Turing's movies, there is a plot: Turing and friends built a decoding machine, but within 24 hours, let decoding machine to carry on the exhaustive cracking time is not enough, at that time, the German cable beginning fixed format as well as the corresponding translation help crack machine greatly reduces the area crack,
Combine the above content, if bound theory describes about the price and then a certain foundation in our model is: the model to predict stock price should not be used, and should use the model to predict movements, more accurate, we predict is the trend of turning point, including more than empty and idle, and of course this is not what innovation, this itself is inflection to predict, but if is given based on the theory of wind,
Put under the theory of deep learning
According to the above ideas, under the guidance of theory of wind, assuming that we are on the line level analysis, that we need to combine below level inside look at recent trends, as well as the date line above the recent trend of more, so I can get more level joint analysis of inflection point
Back to the depth of learning, the theory of wind direction of the definition of form is CNN can easily grab, and bound theory define the direction of mechanics should be LSTM category, so CNNLSTM should be compared with the model,
Specific operation
Wide platform has many restrictions, use keras + tensorflow basically can't back to test, here recommend JQData wide background data interface, specific tutorial at https://www.joinquant.com/view/community/detail/e06d91f508f31054042b904905049d41? Type=1
How to use the said there is not much, anyway,
Here I will give two idea introduction:
The first level of more simple and crude together analysis method,
Definition on the daily online movements, such as using a MA5 segmentation graph, of course, there are other ways to split, but the result will be down,,,, such a situation,
For example, then you can put all the stock in the csi 300 10 years data to make such moves data segments, then for each data of the period, take 5 m data as a secondary trend data segment, as training data is that the level of 5 m times, each period is the turning point of line, the next thing is to keras definition CNNLSTM model starts to run, the model on the use of specific please refer to the above link and Internet content,
Slightly more complicated multiple levels of analysis at the same time
Here can use two or more models each model corresponding to the kind of situation, such as lines and 30 m, as well as the weekly and daily, ran out of the two model training, can combine the results of analysis
A bit more complex analysis
Above model after actually can serve as a basis, for later analysis such as combining to the reinforcement learning as a reference to policy making, and so on inside, but this is beyond the scope of this article,
Training data
Estimates that someone will ask specific use what data to train? If according to the theory of wind, the closing price is enough, but to look at the personal idea, I think with real-time open ended, high, etc,
Conclusion
The model described above, in theory, can detect the level of corresponding to the direction of the turning point, the next business operation need not me ha, perhaps some people say that the above method is based on the theory of wind, I don't believe it bound theory, so the insecure, I want to say is that this article only borrow bound theory theories to a more appropriate use of deep learning model, is predestined friends the human to be inspired,
Personal ideas (is discussed)
AI now very fire, and it is easy to understand, known the various quantitative methods, finally can be summarized as factor to track, I think now the AI is the way of an automatic tracking factor, the use of AI is not really what news at present stage, the theories have been put forth as early as a few decades ago, AI still follow the definition of Turing down, in other words the AI can't solve all the problems, in other words, with an AI lecturer's words, "people can do the AI could do it, people can't do things the AI also can't do", my personal speculated that if use training play chess or play games for AI to fry, effect may not be as before the abuse of human,
CodePudding user response:
The following link is about the LSTM applications:https://blog.csdn.net/mangobar/article/details/105248260
CodePudding user response:
https://blog.csdn.net/mangobar/article/details/105406137CodePudding user response:
https://blog.csdn.net/mangobar/article/details/105200963