TWM560648U - Deep-learning turning point prediction system based on historical track of price rise and drop for financial products - Google Patents

Abstract

This creation provides a turning point prediction system for deep learning through the historical track of financial commodity fluctuations, including database, computing server cluster and display. The database stores information on the ups and downs of at least one financial commodity. The computing server cluster is configured to receive the ups and downs historical trajectory information, and then use the recurrent neural network to supplement the long-term and short-term memory as a model training and establish a prediction model, and use the prediction model to generate a turning point prediction curve of the financial commodity, the turning point The forecast curve has red dot up or green dot turn down information. The display is provided to display the turning point prediction curve for the financial item. Thereby, the prediction model can repeatedly train and correct the original data of the time attribute, and further strengthen the feature quantity of the model with the characteristic expression meaning.

Description

Through the historical track of financial commodity fluctuations, the turning point prediction system for deep learning

This creative department is related to the field of machine learning, especially regarding a turning point prediction system for deep learning through the historical track of financial commodity ups and downs.

At present, the stock market analysis robots mostly use the forecast of tomorrow's stock price or tomorrow's ups and downs as their starting point. These forecasting methods are analyzed, for example, through the comprehensive information of the K-line chart, that is, the opening and closing, the highest and lowest big data cross-parameter comparison and approximation are analyzed. By. However, even though these stock market analysis robots use the concept of machine learning, they still do not deviate from the meaning of human beings to design feature quantity first, and then learn the machine model, that is, adjust the weight through each output performance. However, the concept of "classification" of information has a completely different expression between the cognition of humans and machines. It is the characteristic attribute that creates the "classification" in combination with the experience of the majority, and it is still common to present the dilemma. Poor prediction results. In recent years, although deep learning in machine learning has been given great attention and attention, how to "strongly solidify" the machine for self-learning of big data information, further find the feature quantity with classification meaning, is the desire of machine learning experts The topic of improvement, especially in the face of various information fields and attributes, how to combine the experts in the information field to express the characteristics of the meaning of the feature as self-inquiry and strengthen it. The learning and establishment of the model is still arduous. .

In view of this, the team feels that its current practices are not perfect and exhausted its mental and painstaking thoughts. Through the combination of machine learning expertise and financial expertise, a The historical track of financial commodity ups and downs is the turning point prediction system for deep learning. Further, using the recurrent neural network and long-term and short-term memory as the model training, the output result generated by the accompanying time parameter is feedback-adjusted, thereby strengthening the curing tool. The feature expresses the feature quantity of meaning, and weakens unnecessary information in machine cognition, and then establishes a turning point prediction system with great reference value.

In view of the above problems, the purpose of this creation is to further model the original data with time-varying properties through the recurrent neural network and the parameter weight adjustment training framework for long-term and short-term memory, thereby realizing a reference system with great reference value.

In order to achieve the above objectives, this creation proposes a turning point prediction system for deep learning through the historical track of financial commodity ups and downs, including: a database, a computing server cluster and a display. The database stores information on the ups and downs of one of the financial products. The computing server cluster is connected to the database by telecommunication, and the computing server cluster receives the ups and downs history trajectory information, and uses the recurrent neural network to supplement the long-term and short-term memory as a model training and establishes a prediction model, and uses the The prediction model generates a turning point prediction curve of the financial product, the turning point prediction curve having at least one red dot or at least one green dot information, wherein the red dot means that the turning point prediction curve will exhibit an upward turning performance, wherein the green dot system It means that the turning point prediction curve will show a downward turning performance. The display is also telecommunications connected to the computing server cluster for displaying the turning point prediction curve of the financial product. Thereby, the prediction model can feedback and adjust the output result generated by the time parameter, strengthen the feature quantity with the characteristic expression meaning, and weaken the unnecessary information in the machine cognition, thereby establishing the turning point prediction with great reference value. system.

In a preferred embodiment, in order to enable the machine to recognize and determine the discovery of the turning point behavior, the model training system generates the ups and downs of the historical trajectory information by a transaction mode. The value of the rate of return is used to adjust the weight and path.

As mentioned above, in order to enable the machine to recognize and judge the discovery of the turning point behavior, the model training system supplements the ups and downs historical trajectory information with the trading mode to generate the high and low returns value for adjusting the weight and path. In a preferred embodiment, wherein the transaction mode is that the red dot is continuously purchased when the red dot appears, the green dot is sold all at the same time and the backhand is emptied at the same time; when the green dot appears without holding the share, the empty spot is emptied until the red dot appears backhand Do more.

As stated above, in order to enable the machine to recognize and judge the discovery of the turning point behavior, the model training system supplements the ups and downs historical trajectory information with the trading mode to generate the high and low returns value for adjusting the weight and path. In the preferred embodiment, the transaction mode is that when the red dot appears, one is bought, the profit reaches 7%, and the loss reaches 7% stop loss; the red dot continues to appear in the middle, and the green dot still appears. If there is a shareholding, it will go back at the same time; if the green spot appears to have no shareholding, it will enter the market and be short-selling.

As stated above, in order to enable the machine to recognize and judge the discovery of the turning point behavior, the model training system uses the ups and downs historical trajectory information to supplement the transaction mode to generate the high and low returns value for adjusting the weight and path. In the preferred embodiment, the transaction mode is that when the red dot appears, one piece is bought, and after the profit exceeds 7%, the stop loss is moved to the position of 50% of the back file; if the profit is less than 7%, the loss is first If it reaches 7%, it will stop playing; if there is still a shareholding in the green spot, it will be played at the same time, and the backhand will be short-selled; if the green spot has no shareholding, it will enter the market and be short-selled.

This paper proposes a turning point prediction system for deep learning through the historical track of financial commodity ups and downs, and further uses the recurrent neural network with long-term and short-term memory as model training, and feedbacks the output generated by the accompanying time parameters. In order to strengthen the characteristic quantity of the characteristic expression, and weaken the unnecessary information in the machine cognition, and then establish a very reference The turning point prediction system of value.

1‧‧‧Database

10‧‧‧Financial goods

100‧‧‧ Ups and Downs Historical Track Information

2‧‧‧ Computing Server Cluster

3‧‧‧ display

Figure 1 is a block diagram of the system of the preferred embodiment of the present invention.

Figure 2 is a diagram of the predicted results of the preferred embodiment of the present invention.

In order for your review board to have a clear understanding of the content of this creation, please use the following instructions to match the drawings.

Please refer to Figures 1 and 2, which are respectively a system block diagram and a prediction result diagram of a preferred embodiment of the present invention. First of all, the author proposes a turning point prediction system for deep learning through the historical track of financial commodity ups and downs, including: a database 1, a computing server cluster 2 and a display 3. The database 1 stores at least one financial commodity 10 ups and downs historical trajectory information 100, for example, the historical trajectory information 100 of each stock in the Taiwan stock market during the tradable period. The concept of the cluster of computing servers 2 refers to a computing body that is assembled from a single to a large number of computers. In the current situation of big data analysis, in order to enhance the speed and ability of computing, there are many practices based on complex computing servers. Because this is not the technical characteristics that this creation wants to emphasize, it is not repeated here.

Further, the computing server cluster 2 is connected to the database 1 by telecommunication, and the computing server cluster 2 receives the rising and falling history track information 100, and uses the recurrent neural network to supplement the long-term and short-term memory as a model. Training and establishing a prediction model, and using the prediction model to generate a turning point prediction curve of the financial product, the turning point prediction curve having at least one red dot or at least one green dot information, wherein the red dot means that the turning point prediction curve will be presented The performance of an upward turn, where the green dot means that the turning point prediction curve will exhibit a downward turn. Adopted here Recursive neural network enables the model to model the time display, that is, the hidden layer between the input layer and the output layer will be associated with the hidden layer at another time, resulting in feedback from the hidden layer. In deep learning, in addition to continuously extending the number of hidden layers between the input layer and the output layer, the time step segmentation can also be recognized as a layer-level depth performance for the feedback effect across time. And when these time-characteristic networks are deployed, end-to-end training can be utilized with time-based backpropagation. However, in the recurrent neural network, the gradient problem of disappearing will be encountered, that is, the node will have a sharp decrease in the perception of the node in the future, and the network will not be effectively trained once the network level is too deep. Therefore, this creation is further supplemented by long-term and short-term memory as a training framework to compensate for the shortcomings of the neural network. In the training process, the memory/forgetting path, the screening path and the neglected path model will be established, thereby strengthening the feature quantity with characteristic expression meaning, and the unnecessary information in machine cognition will be weakened or ignored. In addition, during the training process, the numerical value is continuously amplified and affected, and the training process is also performed by the S function. This part is not the technical focus of this case, so it will not be described. As for the display 3, the computing server cluster 2 is telecommunicationly connected to display the turning point prediction curve of the financial product 10. Thereby, the prediction model can feedback and adjust the output result generated by the time parameter, strengthen the feature quantity with the characteristic expression meaning, and weaken the unnecessary information in the machine cognition, thereby establishing the turning point prediction with great reference value. system.

In a preferred embodiment, in order for the machine to recognize and determine the discovery of the turning point behavior, the model training system generates the rate of return and the path for adjusting the weight and path by supplementing the ups and downs of the historical track information with a trading mode. Among them, this creation proposes three trading methods for machine cognition in this trading mode. The first one is the "plus code method". The content is that when the red dot appears, it continues to buy. When the green dot appears, it is all sold and at the same time it is backhanded; when the green dot appears, there is no shareholding. Empty until the red dot has a backhand. The second one is the Stop Loss and Suspension Method. The content is to buy one when the red dot appears, the profit reaches 7%, the loss reaches 7% stop loss; the red dot continues to appear in the middle, it does not move; Green Point still holds shares and then goes back at the same time; when the green spot appears to have no shareholding, it enters the market and empties. The other is the "stop loss and mobile suspension method", the content is to buy one when the red dot appears, after the profit exceeds 7%, the stop loss moves to the position of 50% of the return file; 7%, the loss will reach 7% first, then the stop loss will appear; if there is still a shareholding in the green spot, it will be played at the same time, and the backhand will be short-selled; if the green spot appears to have no shareholding, it will enter the market and be short-selled. As shown in Figure 2, this is the model training and analysis for the stock number (2105). The training conditions are: during the test period 2016/01/01~2017/09/19; Time Step) is ten; the trading method adopts "stop loss and mobile suspension method"; the number of transactions is 12 times, 7 times more, and 5 times; the rate of return is 63.8%. In the figure, the square symbol is the green dot; the circular symbol is the red dot, which is clearly visible to a financial expert, and the prediction about the turning point of the green dot and the red dot has a fairly accurate performance, and Produce valuable reference indicators in the real financial society.

In summary, this paper proposes a turning point prediction system for deep learning through the historical track of financial commodity ups and downs, and further uses the recurrent neural network with long-term and short-term memory as the model training, and generates the accompanying time parameters. The output result is feedback-adjusted, thereby strongly curing the feature quantity with the meaning of the feature expression, and weakening the unnecessary information in the machine cognition, thereby establishing a turning point prediction system with great reference value.

However, the above descriptions are only for the preferred embodiment of the present invention and are not intended to limit the scope of the present invention; therefore, the equivalent changes and modifications made without departing from the spirit and scope of the present invention should be Within the scope of this creation's patent.

Claims (5)

A turning point prediction system for deep learning through the historical track of financial commodity fluctuations, comprising: a database storing at least one financial commodity ups and downs historical trajectory information; a computing server cluster, a telecommunications connection to the database, and the operation After receiving the information of the ups and downs of the historical trajectory, the server cluster trains with the recurrent neural network and the long-term and short-term memory as a model and establishes a prediction model, and uses the prediction model to generate a turning point prediction curve of the financial product, the turning point prediction The curve has at least one red dot or at least one green dot information, wherein the red dot means that the turning point prediction curve will exhibit an upward turning performance, wherein the green dot means that the turning point prediction curve will exhibit a downward turning performance; A display, telecommunications connection to the computing server cluster for displaying the turning point prediction curve for the financial item. The turning point prediction system according to claim 1, wherein the model training system supplements the ups and downs historical trajectory information with a trading mode to generate a return rate value for adjusting the weight and the path. For example, the turning point prediction system described in claim 2, wherein the trading mode is that the red dot is continuously purchased when the red dot appears, and the green dot is all sold and simultaneously backhanded; the green dot is not held when it appears Shares, then short-sell until the red dot has a backhand. For example, the turning point prediction system described in claim 2, wherein the trading mode is that when the red dot appears, one piece is bought, the profit reaches 7%, and the loss reaches 7% stop loss; the red dot continues in the middle. If it appears, it will not move; if the green dot still has a shareholding, it will appear at the same time and the backhand will be emptied; if the green dot appears without holding the share, it will enter the market and be emptied. For example, the turning point prediction system described in claim 2, wherein the transaction mode is When the red dot appears, it will buy one. After the profit exceeds 7%, the stop loss will move to the position of 50% of the return file; if the profit is less than 7%, the loss will reach 7% first and then stop appearing; If there is still a shareholding, it will go out at the same time, and the backhand will be short-selling; when the green dot appears without holding shares, it will enter the market and be short-selling.