TWM561277U - Computing apparatus for processing price images of a financial instrument - Google Patents

Abstract

A computing apparatus is disclosed for processing price images of a financial instrument. A Convolutional Neural Network (CNN) model is utilized to retrieve a set of history feature values from each of multiple history price images and a set of target feature values from a target price image. Each of the history price images is pre-assigned with one of N classification labels to classify all sets of the history feature values into N label sets. A clustering analysis algorithm is utilized for clustering the target feature values respectively with each of the N label sets of history feature values to determine the set of target feature values is clustered respectively in which one of history clusters within each of the N label sets of history feature values. In each of the N label sets of history feature values, a clustering feature distance representative is calculated between the set of target feature values and the sets of the history features within the same one of history clusters. In accordance, N clustering feature distance representatives are obtained and at least an operation signal is output accordingly.

Description

Computing device for image processing of financial commodity prices

The present invention relates to a computer device, and more particularly to an arithmetic device for image processing of financial commodity prices.

Program trading technology is an analysis tool for various financial indicators such as moving averages, Stochastic Oscillator or Relative Strength Index (RSI), together with trading strategies for specific financial commodity price trends. A software program for analyzing and automating transactions, automatically trading in securities or derivative financial products based on preset entry/exit conditions.

Executing trading programs on computer equipment to replace the human brain to determine trading opportunities and operational strategies does not necessarily ensure positive performance in the long run; even after a large number of financial commodity historical price data for computer simulation to verify trading results, based on human trading strategies The deductive software program is still inevitable. The problem may be that the trading strategy itself cannot be applied to all market changes, or it may be due to a software bug written by humans. Furthermore, historical price data and its trends may not be fully resolved, so that only partial data is used in the software program, and the inconspicuous variables are ignored. The use of analytical results with insufficient objectivity as a two-pole plan for judging whether to enter or not is also one of the reasons for the loss of reliability of program trading techniques.

In an embodiment, the present invention discloses an arithmetic device including a main control circuit, a storage unit for storing a code, and a processing unit operatively electrically connected to the storage unit through the main control circuit. The processing unit extracts and executes the code from the storage unit via the main control circuit, and extracts a set of historical feature values from a complex historical price image of a financial product through a convolutional neural network model, and each historical price image has been separately predicted. One of the N classification labels is assigned, and the group historical feature values are classified into N label groups, where N is an integer greater than or equal to 2. The processing unit extracts a set of target feature values from the target price image of one of the financial commodities using a convolutional neural network model. The processing unit uses a group analysis algorithm to participate in the set of target feature values of the target price image in the N tag group historical feature values of the N tag group historical price images classified by the N classification tags, one by one. Grouping the historical feature values of each tag group to determine the set of target feature values of the target price image, respectively, and grouping them into one of the complex historical clusters in each tag group historical feature value. The processing unit further calculates a group feature distance representative number between those historical feature values of the N sets of the target eigenvalues in the N sets of the tag groups to obtain N group feature distance representative numbers. The processing unit outputs at least one operation signal according to the representative number of the group feature distances.

In an embodiment, the processing unit may assign one of the N classification labels to the target price image according to the group feature distance representative numbers and output the operation signal accordingly.

In an embodiment, when the processing unit outputs the operation signal according to the representative distance of the group feature distances, the multi-investment decision trust degree is calculated according to the representative number of the group feature distances, and the operation signal is output according to each, and the investment decision trust degree = each group feature The distance representative / all group feature distances represent the sum of the numbers.

In one embodiment, the classification tag is selected from the group consisting of an entry tag, a non-entry tag, and a combination of a plurality of non-advance tags.

In an embodiment, when N=2, the historical feature values are classified into an entry label group and a non-entry label group by the processing unit, and when two group feature distance representative numbers are obtained, the processing unit according to the two The smaller one, assigns the entry label or non-entry label of the associated label group to the target price image.

In an embodiment, if the grouping feature distance representative number of the approach label group is small, the approach label is assigned to the target price image by the processing unit, and the processing unit further calculates an investment decision ratio corresponding to the approach label group, and the investment decision Ratio = the sum of the group feature distance representative number / the two group feature distance representative numbers of the non-entry tag group.

In an embodiment, if the grouping feature distance representative number of the non-entry label group is small, the non-entry label is assigned to the target price image by the processing unit, and the processing unit further calculates an investment decision ratio corresponding to the non-entry label group. , Investment decision ratio = the sum of the group feature distance representative number of the approach tag group / the distance representative of the two group feature distances.

In one embodiment, each group feature distance representative number is selected from a plurality of group feature distances or separately calculated, and each group feature distance is selected from an Euclidean distance, a Manhattan distance, a Chebyshev distance, and a Minkowski distance. The combination of the definition of Euclidean distance, cosine of the angle and the distance of Bric Curtis.

In one embodiment, each group feature distance representative number is selected from the group consisting of an arithmetic mean, a geometric mean, a median, a maximum value, and a minimum value of a complex group feature distance.

In one embodiment, the historical price images respectively correspond to a plurality of consecutive trading days within a historical sampling period, and the number of these historical price images is the same as the number of consecutive days of the consecutive trading days of the historical sampling period.

In one embodiment, each historical price image corresponds to a complex historical price of the same one sample time length.

In one embodiment, the convolutional neural network model is selected from one of the neural network models such as InceptionV3, Xception, VGG16, VGG19, ResNet, InceptionResNetV2, MobileNet, or any combination thereof.

In one embodiment, the group analysis algorithm is selected from the group consisting of attractor clustering, K-Means grouping, K-Medoids grouping, hierarchical grouping, or any combination thereof.

In an embodiment, the historical price image and the target price image are selected from the K-line chart, the US line graph and the point map, and the moving average indicator map, the oscillator index graph, the trend indicator graph, the moving average graph, and the exponential smoothing similarities and differences. Moving average chart, stochastic indicator chart, relative strength indicator chart, trading momentum indicator chart, trading intention indicator chart, William indicator chart, ups and downs ratio indicator chart, trading over test indicator chart, trend index chart, psychological line indicator chart, momentum A combination of indicator maps, drop indicators, VIX volatility indicator maps, and any other images derived from the price of the financial instrument.

In an embodiment, the operation signal corresponding to the financial product is output according to the representative number of the group feature distances, and the corresponding program is selected from executing the transaction, assigning one of the N classification labels to the target price image, or grouping the groups The feature distance representative number is input to a reinforcement learning model as a plural learning material.

In summary, by applying the overall technical solution of the embodiments of the present invention, the computing device uses the convolutional neural network model to completely extract the historical data value of the historical price map and the target price map of the financial product and the historical characteristic value implied by the trend graph. And the target feature value, and further obtain the relatively objective N group feature distance representative number through the cluster analysis algorithm, calculate the quantitative and objective investment decision trust degree as the system decision basis, avoiding the prior art failing to fully analyze the price data and Trends, as well as the use of poorly objective analysis results to determine the timing of entry, leading to the use of local bias data and ignoring some of the variables, it is difficult to maintain long-term performance and reliability and other technical issues.

The new computing device (Computing Apparatus) is a kind of computer equipment, and the hardware configuration can be implemented in a single machine or in the following manner: a computer device including a plurality of independent processors/memory boards, and a plurality of computers that can be serially connected to each other a device, a computer device externally connected with a graphics operation display card, or a separate graphics card configured with a graphics processing unit (GPU); the hardware configuration and the code installed in the built-in or external storage medium can perform the implementation of the new implementation In the example, the operation process of the computing device performs feature extraction and group analysis and correlation operations on a historical/target price image of a financial product (for example, the K-line diagram of FIG. 2A). Financial commodities can be, for example, various securities of domestic and foreign markets such as stocks, bonds, currencies, commercial papers, and derivative financial commodities such as futures, futures, options, warrants, and the like.

Referring to FIG. 1, which is a hardware block diagram of an arithmetic device in an embodiment of the present invention. In this example, the computing device 100 mainly includes a processing unit 110, a storage unit 120, an input unit 130, a display unit 140, a network unit 150, and a main control circuit 160. The processing unit 110, the storage unit 120, the input unit 130, the display unit 140, and the network unit 150 are electrically connected to each other through the main control circuit 160 to transmit data and signals. The processing unit 110 is a processor having a single or multiple operation processing cores, for example, implemented by an integrated circuit, and performs arithmetic operations according to user instructions preset or received by the system.

The storage unit 120 is a memory for storing data data, and includes a built-in or external storage medium that is directly connected to the processing unit 110 by a dedicated circuit or accessible by the processing unit 110 through the main control circuit 160. The storage unit 120 may include One or any combination of the following storage media: such as random access memory, flash memory, read-only memory, erasable programmable read-only memory, electronic erasable rewritable read-only memory, temporary As the arithmetic device 100, a memory, a hard disk, a portable hard disk, a CD-ROM, and the like are stored. The storage unit 120 stores an operating system (OS) 121 and a code 122 executable by the processing unit 110, and the processing unit 110 can run the operating system 121 by using the memory resource, execute the operation related to the code 122, and execute the loading. Other applications, either in accordance with a basic input/output system (BIOS) (which may be stored in the storage unit 120 or otherwise stored in a separate wafer set connection master circuit 160), activate/control various functional elements. In another example, the storage unit 120 can be built in the processing unit 110. In an embodiment, the code 122 for executing the new operational flow can be built into the central processing unit or the graphics processor firmware (CPU). /GPU Firmware) or Basic Input Output System (BIOS).

The input unit 130 can be a physical or virtual keyboard, a mouse, a trackball or other device for inputting instructions or data to the user; the display unit 140 is a flat display, a projection unit or other device for outputting images, and the plurality of display pixels are respectively turned on. The display unit 130 and the display unit 140 can be integrated into one, for example, a touch screen that implements an input/output function by various technologies. The network unit 150 connects various open or closed networks, receive/transmit network signals by wire or wirelessly, so that the processing unit 110 can access network resources or perform data transmission.

The main control circuit 160 is a circuit board, a flexible circuit board, a bus bar, a bridge, an integrated circuit or any other directly or indirectly operatively electrically connected to the processing unit 110, the storage unit 120, the input unit 130, and the display unit. 140 and network unit 150 are implemented. Although the present embodiment introduces the input unit 130 and the display unit 140, in the near-end or far-end operable environment, as long as the processing unit 110 can access and execute the code stored in the storage unit 120 through the main control circuit 160, The result of the remote or near-end output execution, other units are not absolutely necessary implementation conditions.

The various exemplary units described above may implement some or all of the functions of an integrated circuit with or without memory elements, and may execute specific code or instructions from internal or external; the integrated circuit may include a central processing unit (CPU), General purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic, A discrete hardware component, an electronic component, an optical component, a mechanical component, or a combination of any of the above.

Referring to FIG. 3, it is a schematic diagram of an operation flow of an arithmetic device in another embodiment. In an example, the processing unit 110 retrieves and executes the code 122 from the storage unit 120 through the main control circuit 160, thereby executing the operation mode of the computing device 100 of the present example, and the execution result is, for example, a historical/target price image of the financial product. Perform feature extraction and group analysis to obtain the complex representative distance representative number, which can be directly used as the basic indicator of financial commodity program trading, and output the operation signal corresponding to the financial product, according to which the transaction is executed, and one of the N classification labels is assigned. The target price image is even imported into a reinforcement learning model as a training material.

In this example, the operation mode of the computing device 100 includes steps S310-S350. In step S310 of FIG. 3, the processing unit 110 executes the code 122 to respectively extract a set of historical feature values for a complex historical price image of a financial product through a convolutional neural network model, and each historical price image is pre-assigned separately. One of the N classification labels, and all the sets of historical feature values are classified into N label groups, where N is an integer greater than or equal to 2. In step S320, the processing unit 110 executes the code 122 to extract a set of target feature values from the target price image of one of the financial products by the convolutional neural network model.

Please refer to FIG. 1 , 2A, 2B, 2C and FIG. 3 , wherein FIG. 2A is a schematic diagram of a historical price image/target price image in another embodiment of the present invention; FIG. 2B is a history of another embodiment of the present invention; Schematic diagram of the price image/target price image. 2C is a schematic diagram of a historical price image/target price image in another embodiment of the present invention. In steps S310 and S320, each historical price image/target price image of the financial product used is a digital image file, and the display has a preset sample time length (for example, including the past 20 trading days of the current day). The plural historical price or target price. In an example, each historical price image/target price image corresponds to a plural historical price or target price of the same one sample time length (eg, 20 days). In an embodiment, the historical price images respectively correspond to the number of historical price images of a plurality of consecutive trading days in a historical sampling period, which is the same as the number of days of the consecutive trading days of the historical sampling period, and the target price map Like a target date from the next day of these consecutive trading days. In different illustrations, the historical price image/target price image may be stored in the storage unit 120, or stored in an external storage medium, a near-end network hard disk, an Internet-accessible cloud hard disk, etc., for processing. Unit 110 retrieves. An example historical price image may be, for example, a K-line diagram 210 of FIG. 2A; a Candlestick chart (Rosokuashi Chart) is also referred to as a yin-yang line, a Sakai line, or a candlestick diagram, and a hollow, light-colored or green bar represents Falling, the upper end is the opening price, the lower end is the closing price, and the solid, dark or red bar represents the rise, the upper end is the closing price, the lower end is the opening price, and the upper shadow/lower shadow line ends represent the highest/lowest price respectively. The K-line chart 210 presents the price of 20 consecutive trading days, so the K-line chart 210 is a 20-day K-line chart. In another example, the historical price image/target price image may be, for example, the US line graph 220 of FIG. 2B; the Open-High-Low-Close chart (OHLC chart) represents the change in stock price with the erected line. The vertical line shows the price difference between the highest price and the lowest price. The horizontal line on the left represents the opening price, and the horizontal line on the right represents the closing price. In another example, the historical price image/target price image may be, for example, the point graph 230 of FIG. 2C; the point and figure chart is a circle "O" and a fork "X" to indicate that the price rises and falls. A chart that focuses on the performance of a price at a price point, without recording the change in price over time or the volume of the transaction. When the price rises to a range of units of price (or unit of price), it is used. An "X" indicates that when the drop reaches a magnitude of one grid, it is represented by an "O". An example (single trading day) historical price or target price may be selected from the combination of opening price, closing price, highest price, and lowest price; in contrast, historical prices represented by K-line chart 210 and US line 220 include opening dates. Price / closing price / highest price / lowest price, point chart 230 can show the opening price / closing price / grid value. The historical price image and target price image of this case can be selected from the K-line chart, the American line chart and the point chart, and can also be selected from the moving average indicator chart, the oscillator index chart, the trend indicator chart, the moving average chart, and the exponential smoothing. Similar and similar moving average chart, stochastic indicator chart, relative strength indicator chart, trading momentum indicator chart, trading intention indicator chart, William indicator chart, ups and downs ratio indicator chart, buying and selling super test indicator chart, trend index chart, psychological line indicator chart, A combination of a momentum indicator chart, a drop indicator, a VIX volatility indicator chart, and any other image derived from the price of the financial instrument.

The Convolutional Neural Network (CNN) model is a neural network operation program executable by the processing unit 110, and can be stored in the storage unit 120 as part of the code 122, or stored in an external storage medium, a near-end network hard disk. The cloud hard disk accessible through the Internet is used by the processing unit 110 to capture and execute. The core part of the convolutional neural network model is the Convolutional Layer. A convolutional layer usually consists of tens to hundreds of N*N filters, each of which is adjusted during training. In order to enhance different image patterns. After the training is completed, each filter can recognize different image modes. The convolutional neural network model further includes a pooling layer, which is similar to the down sampling in signal processing, for example, selecting different windows on the image and selecting in the window range. A maximum. A convolutional neural network model is superimposed by a complex convolutional layer and a complex pooling layer, which can combine the basic corner segments of the image into the structure of the object in the image and further transform it into image features (Feature )value.

In an embodiment, the aforementioned convolutional neural network model is implemented by a VGG16 neural network model. The VGG16 neural network model has an image input format of 224*224 RGB images. The neural network model has 16 layers, including 13 layers of 3*3 convolution layers for feature learning and 3 layers of fully connected layers. Fully-connected). In addition, a 5-layer pooling layer is dispersed between the convolution layers, and the overall VGG16 neural network model is cut into 6 blocks, and the first 5 blocks are composed of a convolution layer and a pooling layer. The extracted image (maxpool) feature is performed, and the last block is composed of a fully connected layer for generating a classification result of image recognition.

Before performing the feature extraction step before step S310, the VGG16 neural network model is previously trained using more than 150 million high-resolution images that are pre-divided into 22,000 classification labels in the Imagenet database. Prepare a 20-day K-line chart of financial products with a total of 100 digital files, namely the above-mentioned 99 historical price images and 1 target price image. The VGG16 neural network model is used to perform historical eigenvalue/target features one by one. Extraction of values.

When eigenvalue extraction is performed with the VGG16 neural network model, a 20-day K-line diagram (for example, the K-line diagram 210 of FIG. 2A) (size: 480*640 pixels) is first converted into an array for storage. A normalized calculation on the data can be used to form a 480*640 array data structure with elements between 0 and 1 (normalized). The normalized array input is performed to complete the pre-trained VGG16 neural network model for feature extraction, and the first convolutional layer with 64 3*3 filters is used for feature extraction, and 64 first convolutional layer images 240 can be obtained. As shown in FIG. 2D (a schematic diagram of 64 feature images extracted by the first convolutional layer of the VGG16 neural network model in another embodiment of the present invention), the 64 first convolutional images 240 The line shape of the K-line diagram 210 of FIG. 2A is faintly visible. Then, the 64 first convolutional images 240 are input to the second and subsequent convolution layers for feature extraction again, and 512 feature extraction patterns are obtained until the thirteenth convolution layer, and the 512 images are transmitted through The pooling layer outputs 512 feature values, that is, the feature extraction step of completing the historical price image (or target price image) of the K-line diagram 210 of FIG. 2A. Then, repeating the above steps, successively extracting 99 historical price images and 1 target price image, and obtaining 99 sets of historical feature values (512 historical feature values per group) and 1 set of target feature values (total 512) Target eigenvalues), each set of eigenvalues constitutes a 100*512 array data type.

In addition to the above-mentioned feature extraction process of the VGG16 neural network model, the ResNet neural network model is taken as an example. The ResNet (Residual Neural Network) residual neural network model uses a Residual Unit to train up to 152 layers of neural network models. Traditional convolutional or fully connected layers will lose information or loss when information is transmitted. Through the residual unit, the input signal can be directly bypassed to the output to ensure information integrity. At the same time, the entire residual neural network model only needs to learn the residual part, and can simplify the complexity of the entire learning objective function. Before performing the feature extraction step of step S310 or S320, the Imagenet database is also used for pre-training of the ResNet50 residual neural network model. When eigenvalue extraction is performed with the ResNet50 residual neural network model (50 layers), a 20-day K-line diagram (for example, the K-line diagram 210 of FIG. 2A) (size: 480*640 pixels) is first converted into an array. For storage, the array is normalized on the data to form a 480*640 array data structure with elements between 0 and 1 (normalized). The normalized array input is performed to the pre-trained ResNet50 residual neural network model for feature extraction, and the first convolutional layer with 64 7*7 filters is used for feature extraction to obtain 64 first convolutional layers. Like 250, as shown in FIG. 2E (a schematic diagram of 64 feature images extracted by the first convolutional layer of the neural network model in another embodiment of the present invention), the 64 first convolutional images 250 is faintly visible in the line shape of the K-line diagram 210 of FIG. 2A, but is different from the first convolutional layer image 240 of FIG. 2D. Then, the 64 first convolutional images 250 are input to the second and subsequent convolution layers for feature extraction. When the 50th layer is reached, there are 2048 feature extraction patterns, and the 2048 images are output into 2048 feature values. That is, the K-line diagram 210 of FIG. 2A extracts one set of historical eigenvalues through the ResNet50 residual neural network model; 99 sets of historical price images and one set of target feature images are successively completed, and the values are composed into 100*2048 array data. Type, that is, obtain 100 sets of characteristic values (2048 per group).

The process of extracting the features of step S310/S320 using the VGG16 neural network model and the ResNet residual neural network model is disclosed above, but the present invention is not limited thereto. In one embodiment, the convolutional neural network model is selected from one of the neural network models such as InceptionV3, Xception, VGG16, VGG19, ResNet, InceptionResNetV2, MobileNet, or any combination thereof. Before performing the feature value extraction of the historical price image/target price image of step S310/320, each set of historical feature values of each historical price image has been pre-assigned one of the N classification labels, and the group history is respectively The feature values are classified into N tag groups, and N is an integer greater than or equal to 2. In one example, the classification tag is selected from the group consisting of an entry tag, a non-entry tag, and a combination of a plurality of non-advance tags.

Then, in step S330, the processing unit 110 executes the code 122, and uses a group analysis algorithm to set the target feature value of the target price image to the N tag group historical price images classified by the N classification tags. Among the N tag group historical feature values, each of the tag group historical feature values is separately grouped to determine the target target feature value of the target price image, and is respectively grouped into a complex historical cluster in each tag group historical feature value. One of them. In one embodiment, the group analysis algorithm is a program for the processing unit 110 to perform the group analysis calculation, and may be stored in the storage unit 120 as part of the code 122, or stored in an external storage medium, a near-end network hard disk, and a The cloud hard disk and the like accessible by the Internet are retrieved and executed by the processing unit 110. In different embodiments, the cluster analysis algorithm is selected from the group consisting of Affinity Propagation Clustering, K-Means Clustering, K-Medoids Clustering, and Hierarchical Clustering. One of them or any combination thereof. Among them, the attractor clustering clustering is different from other clustering algorithms in that all data points are regarded as potential clustering centers at the beginning of the algorithm, and then the information transfer between the data points is used to find the most suitable clustering center. And divide other data points into these cluster centers. There are two kinds of information transmitted at each data point in the grouping process, which are responsibility and availability. The clustering result depends on the similarity between the data point samples and the information transfer. The attractor clustering clustering analysis algorithm is mainly to find out the center of each cluster as the main purpose. Compared with attractor clustering, K-means grouping must specify the number of clusters in advance. Hierarchical grouping does not require prior grouping, but in various applications of this novel, it is necessary to plan in advance that each group has several eigenvalues. In summary, for the various embodiments of the present invention, the input and output concepts of the attractor clustering group, the K-means grouping, and the hierarchical grouping are as follows: (1) Attractor clustering grouping: input is N group 512 characteristic values, the output is M group, without specifying the number of clusters in advance; (2) K-means grouping: input is 512 characteristic values of N groups, and the output is K group, and the number of clusters K must be specified in advance; (3) Hierarchical grouping: The input is 512 characteristic values of N groups, and the output is M group. It is necessary to specify in advance that each group has K characteristic values. In addition, K-medoids grouping is to replace the mean calculation in the K-means group with the median calculation and related optimization.

However, the eigenvalues after grouping are scattered in the high dimension (for example, when 512 eigenvalues are extracted for each price image, they will be distributed in 512 dimensions after grouping), and it is not easy to graphically display the results after grouping. The values provide a distribution of the distribution after grouping. Please refer to FIG. 4 , which is a distribution diagram of the 35th and 501th eigenvalues of 100 sets of 512 eigenvalues in another embodiment, including 99 sets of historical eigenvalues and 1 set of target eigenvalues. A two-dimensional distribution map composed of 99+1 35th eigenvalues (X-axis) and 99+1 501th eigenvalues (Y-axis). In FIG. 4, the 35th feature value of the target feature value is 0.0403119139, and the 501st feature value is 0.457416027, which constitutes a coordinate (0.0403119139, 0.457416027), that is, the feature point 410. After the feature points 410 are grouped, they belong to the sub-cluster E. In addition to the feature points 410, there are seven other feature points (box E) belonging to the sub-cluster E. Therefore, the distribution and clustering of the eigenvalues after grouping should be clearly revealed. concept.

In the following, the sub-cluster clustering is performed, and the 99 sets of historical eigenvalues and one set of target eigenvalues taken by the aforementioned VGG16 neural network model are taken as examples to illustrate the implementation of the clustering analysis algorithm. For the clustering analysis calculus of step S330, first, since each group of historical feature values of each historical price image has been pre-assigned one of N classification labels, all the group historical feature values have been pre-classified into N label groups. Therefore, the attractor clustering cluster analysis algorithm is used to set the target eigenvalues of the target price image (for example, the 512 eigenvalues extracted by the VGG16 neural network model), and the N tag group historical prices classified by the N classification tags. Among the N tag group historical feature values of the image, one by one participates in the grouping of the historical feature values of each tag group. That is, if N=2, that is, there are a total of 2 tag group historical feature values, in the example of the previous VGG16 neural network model, for example, the first tag group has 44 sets of historical feature values (512 per group), The two tag groups have a total of 55 sets of historical feature values (512 per group), and a total of 99 sets of historical feature values are derived from the aforementioned 99 historical price images. Firstly, the group of target eigenvalues (such as the above 512 eigenvalues) and the first tag group have 44 sets of historical eigenvalues (512 per group) by clustering clustering analysis algorithm. After completion, they are attracted to the group. The sub-cluster group analysis algorithm groups the group of target eigenvalues (such as the above 512 eigenvalues) and the second tag group with 55 sets of historical eigenvalues (512 per group). Thus, assuming that the 44 sets of historical feature values of the first tag group (512 per group) are finally divided into 11 historical clusters, and 55 sets of historical feature values of the second tag group (512 per group) are finally divided into 5 historical clusters, then After grouping, the group target feature values of the target price image may be separately determined, which group of the 11 historical clusters of the first tag group, and which of the five historical clusters grouped into the second tag group are grouped.

In step S340, the processing unit 110 retrieves and executes the code 122 to calculate a group feature distance representative number between the historical feature values of the same historical cluster in the N tag groups. N group feature distances represent numbers. In an embodiment, each group feature distance representative number is selected from a plurality of group feature distances or separately calculated, and each group feature distance is selected from an Euclidean Distance, a Manhattan Distance, and a cut. The combination of the Chebyshev Distance, the Minkowski Distance, the Standardized Euclidean distance, the Cosine and the Bray Curtis Distance. The foregoing calculation of the group feature distance is intended to find the distance representative number between the center and the target feature value of each historical cluster, for example, between the feature point 410 in FIG. 4 and the other seven similar group feature points, respectively. The value is calculated using the definition of the above-described grouping feature distance. The group feature distance representative number is selected from the group consisting of an arithmetic mean, a geometric mean, a median, a combination of a maximum value and a minimum value of a complex group feature distance.

Finally, in step S350, the processing unit 110 executes the code 122, and outputs at least one operation signal corresponding to the financial product according to the group feature distance representative number. First, in an embodiment, the operation signal corresponding to the financial product is output according to the representative number of the group feature distance, and the corresponding program is selected from executing the transaction, assigning one of the N classification labels to the target price image, or grouping the groups The feature distance representative number is input to a Reinforcement Learning Model as a plural learning material. Secondly, after obtaining the representative numbers of the N cluster feature distances, it is already known that the target feature values are in the high-dimensional distribution and the distances of the clustered features of the N historical clusters, so that it is easy to understand which of the N classification labels is closer to the target feature values. One, so that the above operation signal can be output.

However, the new model can further provide a quantitative basis for decision making if needed. Please refer to FIG. 5A , which is a schematic flowchart of the operation mode of the computing device in another embodiment of the present invention. Step S350 in FIG. 3 can further provide a quantitative decision basis through calculation of an investment decision trust degree. After the step S340, the processing unit 110 executes the code 122, and outputs the operation signal according to the group feature distance representative numbers. As shown in step 351, the complex investment decision trust degree is calculated according to the group feature distance representative numbers and the output operation is performed according to the grouping feature distance representative numbers. Signal, investment decision trust = each group feature distance representative number / the sum of all group feature distance representative numbers.

Second, in an embodiment, the target price image can be assigned as one of the N classification tags. Please refer to FIG. 5B , which is a schematic flowchart of the operation mode of the computing device in another embodiment of the present invention. After step S340, the processing unit 110 executes the code 122, and when the operation signals are output according to the group feature distance representative numbers, as shown in step 352, one of the N classification labels is assigned to the target price according to the group feature distance representative numbers. The image is output according to the operation signal. In an example, the processing module 110 assigns the target price image to the target price image according to the smallest representative number of the group feature distances, and the associated classification label (still one of the N classification labels). In another example, the processing module 110 calculates the investment decision trust minimum according to the representative number of the group feature distances, and the associated classification label (still one of the N classification labels) is assigned to the target price image.

FIG. 6 is a schematic flow chart showing the operation mode of the computing device in another embodiment of the present invention. In an embodiment, when N=2, the historical feature value (processed unit 110) is classified into an approach tag group and a non-entry tag group (step S311), and the approach tag group and the non-entry tag are obtained. When the two group feature distances of the group represent the number (step S341), the processing unit 1105 executes the code 122, and assigns the entry label or the non-entry label of the belonging label group according to the smaller representative number of the two group feature distances. The target price image is given (step S353). Wherein, if the grouping feature distance representative number of the approach tag group is small, the approach tag is assigned to the target price image by the processing unit, and the processing unit further calculates an investment decision ratio corresponding to the approach tag group, and the investment decision ratio = not entered The group feature distance of the field label group represents the sum of the number / 2 group feature distance representative numbers. If the grouping feature distance representative number of the non-entry label group is small, the non-entry label is assigned to the target price image by the processing unit, and the processing unit further calculates an investment decision ratio corresponding to the non-entry label group, and the investment decision ratio = The group feature distance of the field tag group represents the sum of the number / 2 group feature distance representative numbers.

In summary, by applying the overall technical solution of the embodiments of the present invention, the computing device uses the convolutional neural network model to completely extract the historical data value of the historical price map and the target price map of the financial product and the historical characteristic value implied by the trend graph. And the target feature value, and further obtain the relatively objective N group feature distance representative number through the cluster analysis algorithm, calculate the quantitative and objective investment decision trust degree as the system decision basis, avoiding the prior art failing to fully analyze the price data and Trends, as well as the use of poorly objective analysis results to determine the timing of entry, leading to the use of local bias data and ignoring some of the variables, it is difficult to maintain long-term performance and reliability and other technical issues.

The operation modes of the computing device described in the flowcharts and the steps are merely exemplified in the various embodiments, and the process execution sequence or step layering can be arbitrarily reorganized within the scope of the novel disclosure, and the specification and the drawings are not The disclosure is limited. Entity/digital or hardware/software components of various types of formats as described in the various steps and sequences are not limited by the steps or order in which they are disclosed.

Although the present invention has been disclosed in the foregoing embodiments, it is not intended to limit the present invention. Any one skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. The scope of patent protection shall be subject to the definition of the scope of the patent application attached to this specification.

100‧‧‧ computing equipment
110‧‧‧Processing unit
120‧‧‧ storage unit
121‧‧‧Operating system
122‧‧‧ Code
130‧‧‧Input unit
140‧‧‧Display unit
150‧‧‧Network Unit
160‧‧‧Master circuit
210‧‧‧K line chart
220‧‧‧United States line drawing
230‧‧‧ points chart
240/250‧‧‧ first volume layer image
S310, S311, S320, S330, S340, S341‧‧ steps
S350, S351, S352, S353‧‧‧ steps

FIG. 1 is a schematic diagram of a hardware block of an arithmetic device according to an embodiment of the present invention. 2A is a schematic diagram of a historical price image/target price image in another embodiment of the present invention. 2B is a schematic diagram of a historical price image/target price image in another embodiment of the present invention. 2C is a schematic diagram of a historical price image/target price image in another embodiment of the present invention. 2D is a schematic diagram of 64 feature images extracted by the first convolutional layer of the VGG16 convolutional neural network model in another embodiment of the present invention. 2E is a schematic diagram of 64 feature images extracted by a target convolutional layer of a ResNet convolutional neural network model in another embodiment of the present invention. FIG. 3 is a schematic flow chart showing the operation mode of the computing device in another embodiment of the present invention. FIG. 4 is a distribution diagram of the 35th and 501th feature values of 100 sets of 512 feature values in another embodiment of the present invention. FIG. 5A is a schematic flow chart showing the operation mode of the computing device in another embodiment of the present invention. FIG. 5B is a schematic flow chart showing the operation mode of the computing device in another embodiment of the present invention. FIG. 6 is a schematic flow chart showing the operation mode of the computing device in another embodiment of the present invention.

Claims (15)

An arithmetic device for image processing of financial commodity prices, comprising: a main control circuit; a storage unit for storing a code; and a processing unit operatively electrically connected to the storage unit through the main control circuit, The processing unit retrieves and executes the code from the storage unit via the main control circuit to extract a set of historical feature values for a plurality of historical price images of a financial product through a convolutional neural network model, each historical price map For example, one of the N classification labels has been pre-assigned, and the group historical feature values are classified into N label groups, where N is an integer greater than or equal to 2, and the processing unit uses the convolutional neural network model to the financial product. One target price image extracts a set of target feature values, and the processing unit classifies the set of target feature values of the target price image by the group analysis algorithm, and the N tag group histories classified by the N classification tags Among the N tag group historical feature values of the price image, each of the tag group historical feature values is separately involved in clustering to determine the target price. The set of target feature values is grouped into one of a plurality of historical clusters in each of the tag group historical feature values, and the processing unit calculates each of the N tag groups to be associated with the group of target feature values. A group feature distance representative number between the historical feature values of the historical cluster is obtained to obtain N representative group distance representative numbers, and the processing unit outputs at least one operation signal according to the group feature distance representative numbers. The computing device of claim 1, wherein the processing unit outputs the operation signal according to the group feature distance representative number, and further comprises assigning one of the N classification tags to the target according to the group feature distance representative number The price image is based on which the operation signal is output. The computing device of claim 1, wherein the processing unit outputs the operation signal according to the group feature distance representative number, and further comprises calculating the complex investment decision trust according to the group feature distance representative numbers and outputting the operation according to the grouping feature distance representative number. Signal, each investment decision trust = each of the group feature distance representative numbers / the sum of all the group feature distance representative numbers. The computing device of claim 1, wherein the classification labels are selected from the group consisting of an entry label, a non-entry label, and a combination of a plurality of non-advance labels. The computing device of claim 1, wherein when N=2, the historical feature values are classified by the processing unit into an entry label group and a non-entry label group, and two representative distance representative numbers are obtained. The processing unit assigns the presence label or the non-entry label belonging to the label group to the target price image according to the smaller of the two. The computing device of claim 5, wherein if the grouping feature distance representative number of the approach label group is small, the approach label is assigned to the target price image by the processing unit, and the processing unit further calculates a The investment decision ratio corresponds to the entry tag group, and the investment decision ratio = the sum of the group feature distance representative number of the non-entry tag group / the two group feature distance representative numbers. The computing device of claim 5, wherein if the grouping feature distance representative number of the non-entry tag group is small, the non-entry tag is assigned to the target price image by the processing unit, and the processing unit is further Calculating an investment decision ratio corresponding to the non-entry tag group, the investment decision ratio = the sum of the group feature distance representative number of the approach tag group / the two group feature distance representative numbers. The computing device according to claim 1, wherein each of the grouping feature distance representative numbers is respectively selected from a plurality of grouping feature distances or separately calculated, and each of the grouping feature distances is selected from an Euclidean Distance, Manhattan Manhattan Distance, Chebyshev Distance, Minkowski Distance, Standardized Euclidean distance, Cosine and Bray Curtis Distance A combination of definitions. The computing device of claim 1, wherein each of the grouping feature distance representative numbers is selected from the group consisting of an arithmetic mean, a geometric mean, a median, a maximum value and a minimum value of the complex group feature distance. The computing device of claim 1, wherein the historical price images respectively correspond to a plurality of consecutive trading days in a historical sampling period, and the number of the historical price images is continuous with the historical sampling period The number of days in the trading day is the same. The computing device of claim 1, wherein each of the historical price images corresponds to a plurality of historical prices of the same one sample time length. The computing device of claim 1, wherein the convolutional neural network model is selected from one of a neural network model such as InceptionV3, Xception, VGG16, VGG19, ResNet, InceptionResNetV2, MobileNet, or any combination thereof. The computing device according to claim 1, wherein the group analysis algorithm is selected from the group consisting of: Affinity Propagation Clustering, K-Means Clustering, and K-Medoids Clustering. One or any combination of Hierarchical Clustering. The computing device of claim 1, wherein each of the historical price image and the target price image is selected from the group consisting of a candlestick chart, an OHLC chart, and a point and figure chart. And the moving average indicator chart, the oscillator index chart, the trend indicator chart, the moving average chart, the exponential smoothing similarity moving average chart, the stochastic indicator chart, the relative strength indicator chart, the trading momentum indicator chart, the trading intention indicator chart, the William indicator chart, The combination of the ups and downs ratio indicator map, the buy and sell over test indicator chart, the trend index chart, the psychological line indicator chart, the momentum indicator chart, the drop indicator, the VIX volatility indicator chart, and any other images derived from the price of the financial product. The computing device of claim 1, wherein the operation signal corresponding to the financial product is output according to the group feature distance representative number, and the corresponding program is selected from executing the transaction, assigning one of the N classification labels to the target The price image, or the number of representative distance representative numbers, is input to a reinforcement learning model as a plural learning material.