CN112561702A - Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine - Google Patents

Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine Download PDF

Info

Publication number
CN112561702A
CN112561702A CN202011504526.5A CN202011504526A CN112561702A CN 112561702 A CN112561702 A CN 112561702A CN 202011504526 A CN202011504526 A CN 202011504526A CN 112561702 A CN112561702 A CN 112561702A
Authority
CN
China
Prior art keywords
asymmetric
support vector
vector machine
stock
generalized
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011504526.5A
Other languages
Chinese (zh)
Inventor
吴青
张恒昌
高小凤
王凡
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Posts and Telecommunications
Original Assignee
Xian University of Posts and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian University of Posts and Telecommunications filed Critical Xian University of Posts and Telecommunications
Priority to CN202011504526.5A priority Critical patent/CN112561702A/en
Publication of CN112561702A publication Critical patent/CN112561702A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/067—Enterprise or organisation modelling

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • General Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Development Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Physics & Mathematics (AREA)
  • Marketing (AREA)
  • General Physics & Mathematics (AREA)
  • Finance (AREA)
  • Game Theory and Decision Science (AREA)
  • Accounting & Taxation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Technology Law (AREA)
  • Educational Administration (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a generalized stock price prediction method based on a multitask asymmetric adjacent support vector machine, which solves the problem that the existing stock market prediction method has poor prediction precision and robustness due to the fact that cross-correlation information among stock indexes and the distribution characteristics of a data set are ignored, and the prediction precision and the robustness are poor. The prediction method mainly comprises the following steps: acquiring a plurality of stock ticket transaction daily market data and preprocessing the stock ticket transaction daily market data to obtain a data set; constructing an asymmetric square epsilon insensitive loss function, and adjusting a hyper-parameter to better adapt to the distribution characteristic of a data set; establishing a generalized stock price prediction mathematical optimization model based on the multitask learning asymmetric adjacent support vector machine according to the multitask learning hypothesis; converting the original planning problem of the stock price prediction model based on the generalized multi-task learning asymmetric adjacent support vector machine into a dual planning problem under the KKT condition; and solving the dual planning to obtain a decision function of the stock price prediction model.

Description

Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine
Technical Field
The application relates to a data processing method suitable for business and financial trend information prediction, in particular to a stock price prediction method.
Background
The steady development of the stock market is the focus of attention of the national government and investors, and analyzing the running trend of the stock market index and the fluctuating change of the stock price is the subject of extensive scrutiny in the academic and financial world. To date, a large amount of research has been carried out by scholars at home and abroad on stock market prediction, but because the stock market has the characteristics of complex nonlinearity, high dimensionality, noise sensitivity, instability and the like, reasonable and accurate fluctuation prediction analysis of the stock market is very difficult.
The existing stock market prediction method is mainly divided into three categories: market technology analysis, ground plane analysis and quantitative analysis methods. The market technology analysis method excavates the periodic rule of the stock price by researching the form curve of the stock price, thereby prejudging the operation trend and fluctuation range of the stock price index. The method needs to follow the complex combination change of 'quantity, price, time and space' elements, is difficult to avoid the hysteresis of time, and easily causes misjudgment of market trend when different technical indexes deviate from each other. The basic surface analysis method mainly evaluates the profitability of the enterprise from three dimensions of macroscopic market economy, industry development prospect and enterprise development potential, needs practitioners to master complete financial market economic theory and research analysis method, can only give fuzzy directional decision on the price of the stock from a qualitative angle, and is quite limited when applied to the stock market which is gradually improved in China. The quantitative analysis method mainly comprises the steps of establishing a mathematical optimization model for fluctuation change of stock price indexes, continuously correcting the model by back test of historical data, and analyzing short-term fluctuation trend of future stock prices by means of the optimization model. The method for constructing the model and quantitatively analyzing not only can fully utilize historical data information, but also can more clearly reflect the law of stock price fluctuation. Typical quantitative analysis methods at present mainly include time series mining, complex neural networks, fuzzy neural systems, machine learning, and the like.
Since machine learning is widely applied to the field of regression prediction, quantitative analysis of the stock market by using machine learning is widely concerned by students. The support vector machine has the advantages of global convergence, insensitive sample dimension, strong generalization performance and the like, and is applied to the prediction analysis of single stock price. The support vector machine can achieve good results for the prediction of a single task, but does not take advantage of the algorithm well because it ignores the correlation between learning tasks. In the stock market, a plurality of stock price indexes have very similar historical fluctuation tracks, and the phenomenon reflects the application limitation of single-task learning. Aiming at learning tasks which are closely related to each other, multi-task learning is used for training simultaneously and capturing useful information among different tasks, so that the generalization performance of the algorithm is improved. Meanwhile, the multi-task learning can effectively relieve the problem of data shortage, so that the method is widely applied to the fields of image recognition, natural voice recognition, financial risk prediction and the like. The learners propose a multi-task adjacent support vector machine method, but a square loss function is used in an objective function of an optimization problem, so that the algorithm has strong limitation, and cannot be well applied to an actual scene.
The technical information introduced above is intended to assist the reader in quickly understanding the relevant background, the objects and concepts of the application, and therefore may contain information and considerations that do not constitute prior art that is well known to those skilled in the art.
Disclosure of Invention
The purpose of this application lies in: the method solves the problems of poor prediction precision and robustness caused by great limitation due to neglect of cross-correlation information among stock indexes and distribution characteristics of a data set in the conventional stock market prediction method.
The technical idea of the application is as follows: the cross-correlation information among the stock indexes and the distribution characteristics of the data set are mined to a greater extent, and the cross-correlation information and the distribution characteristics of the data set are better applied to stock forecasting actual scenes, so that better forecasting precision and robustness are obtained. The main links comprise: downloading a plurality of stock ticket transaction daily market data and preprocessing the stock ticket transaction daily market data to obtain a data set; constructing an asymmetric square epsilon insensitive loss function, and adjusting a hyper-parameter to better adapt to the distribution characteristic of a data set; establishing a generalized stock price prediction mathematical optimization model based on the multitask learning asymmetric adjacent support vector machine according to the multitask learning hypothesis; converting the original planning problem of the stock price prediction model based on the generalized multi-task learning asymmetric adjacent support vector machine into a dual planning problem under the KKT condition; and solving the dual planning to obtain a decision function of the stock price prediction model. The data set is divided into a training sample set and a testing sample set, a stock price prediction model is obtained through the training sample set, and the prediction precision and robustness of the method are evaluated by using the testing sample set.
The stock price forecasting method based on the multitask asymmetric adjacent support vector machine in the broad sense provided by the application mainly comprises the following steps of:
step 1: acquiring trading date market data of a plurality of stock indexes before a forecast date and preprocessing the trading date market data to obtain a data set; each stock index is a separate learning task;
step 2: constructing an asymmetric square epsilon insensitive loss function, and adjusting a hyper-parameter to better adapt to the distribution characteristics of the data set;
and step 3: according to the multi-task learning assumption, a mathematical optimization model based on the multi-task asymmetric adjacent support vector machine is established on the basis of the adjacent support vector machine;
and 4, step 4: the general information and the private information among the learning tasks are distinguished by adopting the same or different kernel functions for a general model and a private model in the multi-task learning hypothesis, so that a generalized mathematical optimization model based on the multi-task asymmetric adjacent support vector machine is obtained;
and 5: converting a generalized mathematical optimization model based on a multitask asymmetric adjacent support vector machine into a dual planning problem of the generalized mathematical optimization model by using a KKT condition;
step 6: obtaining a generalized decision function based on a multitask asymmetric adjacent support vector machine by solving a dual planning problem; taking partial data from the data set as a training sample set, and learning by using the training sample set to obtain a stock index prediction model;
and 7: and taking out data corresponding to the target stock index from the data set, and outputting a stock index prediction result through the stock index prediction model.
The application also provides corresponding two types of computer program products and equipment:
a computer device comprising a memory storing a computer program and a processor implementing the steps of the prediction method described above when executing the computer program.
A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the prediction method described above.
Compared with the prior art, the method has the following advantages that:
in theory:
(1) the method is based on a multi-task adjacent support vector machine, original data samples are gathered to a decision plane closest to the original data samples as far as possible, and the method is not designed based on the principle of interval maximization, so that the method has strong robustness;
(2) according to the method, an asymmetric square epsilon insensitive loss function is adopted in a mathematical model, and the hyperparameter is properly adjusted to better adapt to the distribution characteristics of original data, so that the generalization performance of the algorithm can be effectively improved to a certain extent;
(3) according to the method and the device, different types of kernel functions can be selected to distinguish and process the general information and the private information, and the flexibility of the algorithm is effectively improved.
In the aspect of application:
the method and the system can effectively mine stock price fluctuation information of different enterprises in the same industry, and further reasonably predict and analyze short-term fluctuation change of the enterprises; meanwhile, a plurality of complexity characteristics in the stock market are considered, the generalization performance of the algorithm can be better improved and the learning effect of the model can be improved by using the asymmetric loss function, and finally the prediction accuracy is improved. In addition, the method can provide more reliable reference for professionals, and saves time for the professionals to analyze relevant information.
Drawings
FIG. 1 is a schematic basic flow diagram of the present application;
FIG. 2 is a comparison graph of price prediction results of three existing prediction methods SVR, PSVR and MTL-PSVR and three multi-task learning methods MTL-a-PSVR, EMTL-a-PSVR (L + P) and EMTL-a-PSVR (L + G) provided by the invention for 500 historical trading days of the Shang Zheng index (SSEC);
FIG. 3 is an enlarged view of a portion of FIG. 2;
FIG. 4 is a comparison graph of the price prediction results of the three existing prediction methods and the three multi-task learning methods provided by the present invention for the deep witness (SZI)500 historical trading days;
FIG. 5 is an enlarged view of a portion of FIG. 4;
FIG. 6 is a comparison graph of the price prediction results of 500 historical trading days of the founder board index (CHINEXTC) by the three existing prediction methods and the three multi-task learning methods provided by the present invention;
FIG. 7 is an enlarged view of a portion of FIG. 6;
FIG. 8 is a comparison graph of price prediction results for 500 historical trading days of the small plate finger (SZSME) by the three existing prediction methods and the three multi-task learning methods provided by the present invention;
FIG. 9 is an enlarged view of a portion of FIG. 8;
FIG. 10 is a schematic view of a usage scenario of the present invention;
fig. 11 is an internal structural view of a computer apparatus.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
As shown in fig. 1, the stock price prediction method based on the multitask asymmetric adjacent support vector machine in the present application includes the following basic steps:
step 1: downloading trading date market data of a plurality of stock indexes before the forecast date and preprocessing the market data to obtain a data set; each stock index is a separate learning task;
step 2: constructing an asymmetric square epsilon insensitive loss function, and properly adjusting parameters to enable the loss function to be better adapted to the distribution characteristics of an original data set;
and step 3: according to the multi-task learning assumption, a mathematical optimization model based on the multi-task asymmetric adjacent support vector machine is established on the basis of the adjacent support vector machine;
and 4, step 4: the common information and the private information among the learning tasks can be distinguished by adopting the same or different kernel functions for the common model and the private model in the multi-task learning hypothesis, so that a generalized mathematical optimization model based on the multi-task asymmetric adjacent support vector machine is obtained;
and 5: converting a generalized mathematical optimization model based on a multitask asymmetric adjacent support vector machine into a dual planning problem of the generalized mathematical optimization model by using a KKT condition;
step 6: obtaining a generalized decision function based on a multitask asymmetric adjacent support vector machine by solving a dual planning problem;
and 7: and outputting the stock index prediction result. Dividing a data set into a training sample set and a testing sample set; a stock index prediction model is obtained through a training sample set, and the prediction precision and robustness of the model are evaluated by utilizing a test sample set and compared with the existing method.
The stock index in the step 1 comprises a top syndrome index, a deep syndrome index, a startup board index and a middle and small board index; the historical market data comprises the previous closing price, the highest price, the lowest price, the closing price, the rising and falling amount, the rising and falling amplitude, the volume of the deal, the hand-changing rate, the amplitude, the total market value, the circulation market value, the market profitability, the market net rate, the market present rate, the market selling rate and the opening price of each trading day. In order to predict the opening index of the stock market index, it is assumed that the opening index on the current day is a dependent variable and the remaining indexes on the previous day are independent variables.
Constructing an asymmetric squared epsilon insensitive loss function as in step 2 above
Figure BDA0002844564580000051
Wherein the epsilon value determines the width of the insensitive area and represents the fault tolerance degree of the decision function to the prediction result. p is a parameter related to the asymmetric property that is used to change the shape of the loss function to adapt as much as possible to the inherent distribution properties of the original data.
Constructing the multi-task learning hypothesis in the step 3 as follows: for learning tasks which are different from each other but are associated with each other, a common model rho is shared among all the tasks0While each subtask has a private model ηtThe decision function for each subtask can then be expressed as:
Figure BDA0002844564580000052
wherein ω is0And
Figure BDA0002844564580000053
normal vector and non-linear mapping function, upsilon, respectively representing a general modeltAnd φ (-) is the normal vector and nonlinear mapping function in the private model. b0Is the general model ρ0Bias term corresponding to decision hyperplane, btRepresenting a private model η for each subtasktThe bias term of (1).
The multi-task learning in step 3 above includes T sub-tasks that are different from each other but related to each other,
Figure BDA0002844564580000054
sample attribute sets representing all learning tasks, and Y represents sample label sets for all learning tasks. Each task includes mtA sample point
Figure BDA0002844564580000055
Wherein the sample attribute value
Figure BDA0002844564580000056
Sample label value
Figure BDA0002844564580000057
In common with
Figure BDA0002844564580000058
Data points, T > 1.
The generalized mathematical optimization model based on the multitask asymmetric adjacent support vector machine in the step 3 is as follows:
Figure BDA0002844564580000061
Figure BDA0002844564580000062
Figure BDA0002844564580000063
ε≥0,ξt≥0,t＝1,2,···,T
where s is the task coupling parameter and v and C are the regularization parameters. p is an asymmetric characteristic related parameter, the value of epsilon determines the width of the insensitive area,
Figure BDA0002844564580000064
for the relaxation vector of the t-th sub-task,
Figure BDA0002844564580000065
representing by non-linear mapping
Figure BDA0002844564580000066
A dataset mapped from the original space to the t-th subtask of the high-dimensional feature space,
Figure BDA0002844564580000067
column vector representing elements all 1, btA bias term representing the tth sub-task private model.
In the above step4 kernel function in generic model under multi-task learning assumption
Figure BDA0002844564580000068
And a kernel function phi (-) in the private model can be the same kernel function or different kernel functions, so that the generalized optimization model based on the multitask asymmetric adjacent support vector machine is obtained as follows:
Figure BDA0002844564580000069
Figure BDA00028445645800000610
Figure BDA00028445645800000611
ε≥0,ξt≥0,t＝1,2,···,T
wherein
Figure BDA00028445645800000612
The data set representing the t-th sub-task mapped from the original space to the high-dimensional feature space by the kernel mapping phi (-) with the other notation being the same as in the optimization problem (3).
In step 5 above, the following equation can be obtained by the KKT condition:
Figure BDA0002844564580000071
Figure BDA0002844564580000072
Figure BDA0002844564580000073
Figure BDA0002844564580000074
Figure BDA0002844564580000075
wherein
Figure BDA0002844564580000076
Figure BDA0002844564580000077
αt，βt(T1, …, T) is the lagrange multiplier vector. The generalized dual planning problem based on the optimization model of the multitask asymmetric adjacent support vector machine can be obtained by the arrangement formula (5):
Figure BDA0002844564580000078
Figure BDA0002844564580000079
wherein Q ═ AATIs a matrix of m x m order,
Figure BDA00028445645800000710
is an m block diagonal matrix.
In the above step 6, let λt＝αt-βt，
Figure BDA00028445645800000711
Solving the optimization problem (6) to obtain the optimal lambda, and further calculating by the formula (5) to obtain omega0，υtAnd bt. To calculate the insensitive pipeline width ε in the tth subtask decision function, consider the data points at the upper boundary of the pipeline
Figure BDA00028445645800000712
And is provided with
Figure BDA00028445645800000713
The set is represented as
Figure BDA00028445645800000714
The following conditions are satisfied:
Figure BDA00028445645800000715
and searching any sample point on the boundary of the pipeline to obtain the insensitive pipeline width epsilon. Solving a generalized mathematical optimization model based on a multitask asymmetric adjacent support vector regression machine, and obtaining a decision function of each subtask as follows:
Figure BDA00028445645800000716
wherein K0(-) corresponds to a mapping function in a generic model
Figure BDA0002844564580000081
Kt(-) corresponds to the mapping function φ (-) in the private model, the upper and lower boundaries of the ε -insensitive pipeline for the t-th subtask can be represented as ft(x) + ε and ft(x)-ε。
In order to reasonably evaluate the prediction accuracy and robustness of the model, a commonly used evaluation index is selected in the step 7. Defining the number of sample points of a training sample set as l, and the number of sample points of a testing sample set as k; x is the number ofiRepresenting the value of the sample attribute, yiAnd
Figure BDA0002844564580000082
respectively represent xiThe true tag value and the predicted tag value of,
Figure BDA0002844564580000083
representing true mark of test sampleAnd averaging the label values. Specific prediction evaluation indexes are as follows:
the mean absolute error is defined as
Figure BDA0002844564580000084
Root mean square error is defined as
Figure BDA0002844564580000085
The sum of squares of errors of the predicted values is defined as
Figure BDA0002844564580000086
The sum of the squares of the deviations of the true values is defined as
Figure BDA0002844564580000087
The sum of squares of the deviations of the predicted values is defined as
Figure BDA0002844564580000088
The ratio of the sum of the squares of the errors of the predicted values to the sum of the squares of the deviations of the true values is defined as
Figure BDA0002844564580000089
The ratio of the sum of the squares of the predicted deviation to the sum of the squares of the actual deviation is defined as
Figure BDA00028445645800000810
Generally, the smaller the values of MAE, RMSE, and SSE/SST, the better the predictive effect of the model. While SSE/SST decreases SSR/SST will increase.
In addition, if the reference model is replaced by the support vector machine with least squares support vector machine in the above solution of the present application, two disadvantages are generated:
(1) the kernel function in the least square support vector machine must meet the positive definite condition in the Mercer's theorem, so the alternative has limited kernel functions available in the using process. The kernel function in the adjacent support vector machine does not need to meet the limiting condition, and the method is provided based on the adjacent support vector machine and does not need to meet the limiting condition;
(2) if the mathematical model of the stock price prediction method of the alternative multi-task least square support vector machine does not use the insensitive loss function of the asymmetric square epsilon but uses the traditional square epsilon loss function, the problems or data sets which can be predicted by the method can be greatly reduced, because the data sets corresponding to the problems to be predicted are usually asymmetrically distributed.
The technical effect of the present application is verified by a specific example below.
The upper syndrome index (SSEC), the deep syndrome index (SZI), the creation board index (CHINEXTC) and the middle small board index (SZSME) are the most important four stock market indexes in the stock market in China. The experimental data in this example is derived from the east wealth Choice financial data terminal, where the subdata set corresponding to each stock index includes 500 historical trading day data from 1/9/2014 to 18/9/2016. The four stock indexes are influenced by many internal or external factors to show synchronous operation trend, so that the stock indexes can be regarded as four related but different subtasks. In order to verify the correlation between the four stock indexes, the relationship between them was analyzed by Spearman's correlation coefficient method. Each transaction date data is taken as a sample point and mainly comprises 17 indexes: front closing price, highest price, lowest price, closing price, rising and falling amount, rising and falling amplitude, volume of transaction, hand-changing rate, amplitude, total market value, circulation market value, market profit rate, market net rate, market present rate, market selling rate and opening price. In order to predict the opening index of the stock market index, it is assumed that the opening index on the current day is a dependent variable and the remaining indexes on the previous day are independent variables. Table 1 shows the correlation between the four large index points. As can be seen from Table 1, the pairwise correlation coefficient of each finger with the other three fingers is greater than 0.6. Analysis results show that the four large-strand fingers have strong internal correlation and can be predicted by adopting a multi-task learning method.
TABLE 1
Correlation coefficient SSEC SZSC CHINEXTC SZSME
SSEC 1.0 0.7695 0.6888 0.7389
SZSC 0.7695 1.0 0.9839 0.9969
CHINEXTC 0.6888 0.9839 1.0 0.9906
SZSME 0.7389 0.9969 0.9906 1.0
For the generalized stock price index prediction method based on the multitask asymmetric adjacent support vector machine, which is provided by the application, the method comprises the following steps:
when K is in the decision function0(. phi) and Kt(ii) when the same kernel function is selected, obtaining a stock price index prediction method (MTL-a-PSVR) based on the multitask asymmetric adjacent support vector machine;
when K is in the decision function0(. phi) and KtWhen different kernel functions are selected, a linear kernel function, a polynomial kernel function, a linear kernel function and a radial basis kernel function are respectively selected to obtain two generalized stock price index prediction methods EMTL-a-PSVR (L + P) and EMTL-a-PSVR (L + G) based on the multitask asymmetric adjacent support vector machine.
The above three multitask-based asymmetric adjacent support vector machine methods proposed by the present application are compared with the existing support vector machines (SVR), adjacent support vector machines (PSVR) and multitask adjacent support vector machines (MTL-PSVR). All methods 20 separate experiments were performed on each stock index dataset and table 2 gives the average of the results of the 20 experiments.
TABLE 2
Figure BDA0002844564580000101
The experimental results in table 2 show that: the stock price prediction method based on the multitask asymmetric adjacent support vector machine has high prediction precision and strong robustness on four stock index prices. As can be seen from table 2: as for the MAE evaluation index, the MTL-PSVR and the MTL-a-PSVR respectively generate the minimum mean absolute error on the two data sets, and the MAE evaluation index of the EMTL-a-PSVR (L + P) and the EMTL-a-PSVR (L + G) on the four data sets achieves better results. For the RMSE evaluation index, the EMTL-a-PSVR (L + P) achieved the smallest error on the SZSC, and the MTL-a-PSVR achieved the smallest error on all three remaining data sets. For SSE/SST metrics, MTL-a-PSVR gave the best results on all four datasets. For SSR/SST evaluation indexes, EMTL-a-PSVR (L + P) and EMTL-a-PSVR (L + G) achieve the best results on four stock index datasets. Experimental results show that the asymmetric squared epsilon insensitive loss function is generally applicable to more data sets. By selecting proper parameters, the prediction precision and the robustness of the multi-task learning model can be further improved.
Fig. 2-9 show the predicted curves of the six methods for the market indexes of the four stocks. Fig. 2, 4, 6 and 8 show the predicted curves of the six methods for 500 consecutive trading days of four large fingers, respectively. In order to distinguish the prediction performance of different methods more clearly, fig. 3, fig. 5, fig. 7, and fig. 9 are sequentially enlarged partial views of regions (marked by dashed boxes) where significant prediction differences exist in fig. 2, fig. 4, fig. 6, and fig. 8. It can be seen from fig. 3, 5, 7 and 9 that the three multitask learning methods and the MTL-PSVR proposed by the present invention have ideal prediction effects on four stock indexes, and have poor prediction effects on SVR and PSVR.
A usage scenario of the present invention is shown in fig. 10, in which a terminal and a server communicate via a network. The server can provide a corresponding user-oriented webpage platform, accept the demand information input by the user, and send corresponding prediction result information to the terminal. The terminal can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices, and the server can be implemented by an independent server or a server cluster formed by a plurality of servers.
In hardware the invention is typically implemented on the basis of a computer device, which may be a server, the internal structure of which may be as shown in fig. 11. The computer device typically includes a processor, memory, network interface, and database connected by a system bus. The processor is used for providing calculation and control capability, and the memory comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium may store an operating system, a computer program, and a database; the internal memory may provide an environment for the operation of an operating system and computer programs in the non-volatile storage medium. The database of the computer device is used for storing relevant data generated by predicting the stock index. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a generalized multi-tasking asymmetric proximal support vector machine-based stock price prediction method.
Accordingly, the present invention may also be embodied directly in hardware in a computer readable storage medium, on which a computer program is stored, the computer program, when executed by a processor, implementing the above-described generalized stock price prediction method based on a multitasking asymmetric proximity support vector machine.
Technical features of the above embodiments may be selectively combined as necessary, but for brevity of description, not all possible combinations of the technical features of the above embodiments are described. However, as long as there is no contradiction between the combinations of these technical features, it should be clearly understood that the scope of the present specification is defined.

Claims (11)

1. The generalized stock price prediction method based on the multitask asymmetric adjacent support vector machine is characterized by comprising the following steps of:
step 1: acquiring trading date market data of a plurality of stock indexes before a forecast date and preprocessing the trading date market data to obtain a data set; each stock index is a separate learning task;
step 2: constructing an asymmetric square epsilon insensitive loss function, and adjusting a hyper-parameter to better adapt to the distribution characteristics of the data set;
and step 3: according to the multi-task learning assumption, a mathematical optimization model based on the multi-task asymmetric adjacent support vector machine is established on the basis of the adjacent support vector machine;
and 4, step 4: the general information and the private information among the learning tasks are distinguished by adopting the same or different kernel functions for a general model and a private model in the multi-task learning hypothesis, so that a generalized mathematical optimization model based on the multi-task asymmetric adjacent support vector machine is obtained;
and 5: converting a generalized mathematical optimization model based on a multitask asymmetric adjacent support vector machine into a dual planning problem of the generalized mathematical optimization model by using a KKT condition;
step 6: obtaining a generalized decision function based on a multitask asymmetric adjacent support vector machine by solving a dual planning problem; taking partial data from the data set as a training sample set, and learning by using the training sample set to obtain a stock index prediction model;
and 7: and taking out data corresponding to the target stock index from the data set, and outputting a stock index prediction result through the stock index prediction model.
2. The generalized multi-task asymmetric proximity support vector machine-based stock price prediction method according to claim 1, characterized in that said plurality of stock indices in step 1 include a top-certificate index, a deep-certificate index, a startup board index and a middle-small board index; the historical market data comprises the previous closing price, the highest price, the lowest price, the closing price, the rising and falling amount, the rising and falling amplitude, the volume of the deal, the hand-changing rate, the amplitude, the total market value, the circulation market value, the market profitability, the market net rate, the market present rate, the market selling rate and the opening price of each trading day.
3. The generalized multi-tasking asymmetric proximity support vector machine-based stock price prediction method of claim 1, wherein the asymmetric square epsilon insensitive loss function constructed in step 2 is
Figure FDA0002844564570000011
Wherein the epsilon value determines the width of an insensitive region and represents the fault tolerance degree of the decision function to the prediction result; p is a parameter related to the asymmetric property that is used to change the shape of the loss function to adapt as much as possible to the distribution properties of the data set.
4. The generalized multi-tasking asymmetric proximity support vector machine-based stock price prediction method of claim 1, wherein the multi-tasking learning hypothesis is constructed in step 3 as follows: for learning tasks which are different from each other but are associated with each other, a common model rho is shared among all the tasks0While each subtask has a private model ηtThen the decision function for each subtask is expressed as:
Figure FDA0002844564570000021
wherein ω is0And
Figure FDA0002844564570000022
normal vector and non-linear mapping function, upsilon, respectively representing a general modeltAnd φ (-) is the normal vector and nonlinear mapping function in the private model; b0Is the general model ρ0Bias term corresponding to decision hyperplane, btRepresenting a private model η for each subtasktThe bias term of (1).
5. The generalized multi-tasking asymmetric proximity support vector machine-based stock price prediction method of claim 4, wherein the multi-tasking learning in step 3 comprises T sub-tasks that are different from each other but related to each other,
Figure FDA0002844564570000023
sample attribute sets representing all learning tasks, and Y represents sample label sets of all learning tasks; each task includes mtA sample point
Figure FDA0002844564570000024
Wherein the sample attribute value
Figure FDA0002844564570000025
Sample label value
Figure FDA0002844564570000026
In common with
Figure FDA0002844564570000027
Data points, T > 1.
6. The generalized stock price forecasting method based on the multitask asymmetric proximity support vector machine according to claim 5, wherein the mathematical optimization model based on the multitask asymmetric proximity support vector machine established in the step 3 is as follows:
Figure FDA0002844564570000028
wherein s is a task coupling parameter, and ν and C are regularization parameters; p is an asymmetric characteristic related parameter, the value of epsilon determines the width of the insensitive area,
Figure FDA0002844564570000029
for the relaxation vector of the t-th sub-task,
Figure FDA00028445645700000210
representing by non-linear mapping
Figure FDA00028445645700000211
A dataset mapped from the original space to the t-th subtask of the high-dimensional feature space,
Figure FDA00028445645700000212
column vector representing elements all 1, btA bias term representing the tth sub-task private model.
7. The generalized multi-task asymmetric proximity support vector machine-based stock price prediction method according to claim 6, wherein the generalized multi-task asymmetric proximity support vector machine-based optimization model obtained in step 4 is as follows:
Figure FDA0002844564570000031
wherein
Figure FDA0002844564570000032
A data set representing the t-th sub-task mapped from the original space to the high-dimensional feature space by the kernel mapping phi (-) with other symbols having the same meaning as in equation (3).
8. The generalized multi-tasking asymmetric proximity support vector machine based stock price forecasting method of claim 1, wherein in step 5, the following equation is obtained by KKT condition:
Figure FDA0002844564570000033
wherein
Figure FDA0002844564570000034
αt，βt(T ═ 1, …, T) is the lagrange multiplier vector; and (5) obtaining a generalized dual planning problem based on a multitask asymmetric adjacent support vector machine optimization model by a finishing formula:
Figure FDA0002844564570000035
wherein Q ═ AATIs a matrix of m x m order,
Figure FDA0002844564570000036
is an m block diagonal matrix.
9. According to the rightThe generalized multi-tasking asymmetric proximal support vector machine-based stock price prediction method as claimed in claim 1, wherein in step 6, let λt＝αt-βt，
Figure FDA0002844564570000037
Solving the optimization problem formula (6) to obtain the optimal lambda, and further calculating by the formula (5) to obtain omega0，υtAnd bt(ii) a To calculate the insensitive pipeline width ε in the tth subtask decision function, consider the data points at the upper boundary of the pipeline
Figure FDA0002844564570000041
And is provided with
Figure FDA0002844564570000042
The set is represented as
Figure FDA0002844564570000043
The following conditions are satisfied:
Figure FDA0002844564570000044
searching any sample point on the boundary of the pipeline, and solving to obtain the width epsilon of the insensitive pipeline; solving a generalized mathematical optimization model based on a multitask asymmetric adjacent support vector regression machine to obtain a decision function of each subtask as follows:
Figure FDA0002844564570000045
wherein K0(-) corresponds to a mapping function in a generic model
Figure FDA0002844564570000046
Kt(-) corresponds to the mapping function φ (-) in the private model, th sub-The upper and lower boundaries of the epsilon-insensitive pipeline for a task may be denoted as ft(x) + ε and ft(x)-ε。
10. A computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor when executing the computer program implements the steps of the prediction method according to any one of claims 1 to 9.
11. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the prediction method according to any one of claims 1 to 9.
CN202011504526.5A 2020-12-18 2020-12-18 Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine Pending CN112561702A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011504526.5A CN112561702A (en) 2020-12-18 2020-12-18 Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011504526.5A CN112561702A (en) 2020-12-18 2020-12-18 Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine

Publications (1)

Publication Number Publication Date
CN112561702A true CN112561702A (en) 2021-03-26

Family

ID=75063632

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011504526.5A Pending CN112561702A (en) 2020-12-18 2020-12-18 Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine

Country Status (1)

Country Link
CN (1) CN112561702A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114036208A (en) * 2021-11-09 2022-02-11 建信金融科技有限责任公司 Model training and sensitivity analysis method, device, equipment and medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114036208A (en) * 2021-11-09 2022-02-11 建信金融科技有限责任公司 Model training and sensitivity analysis method, device, equipment and medium

Similar Documents

Publication Publication Date Title
Bardwell et al. Most recent changepoint detection in panel data
Xu et al. A novel UMIDAS-SVQR model with mixed frequency investor sentiment for predicting stock market volatility
Akbari et al. Drivers of economic and financial integration: A machine learning approach
Lee et al. Shap value-based feature importance analysis for short-term load forecasting
Sen et al. A comparative study of hierarchical risk parity portfolio and eigen portfolio on the NIFTY 50 stocks
Popa et al. Composite financial performance index prediction–a neural networks approach
Tung et al. Binary classification and data analysis for modeling calendar anomalies in financial markets
Kureljusic et al. Revenue forecasting for European capital market-oriented firms: A comparative prediction study between financial analysts and machine learning models
US11995667B2 (en) Systems and methods for business analytics model scoring and selection
Khattak et al. A systematic survey of AI models in financial market forecasting for profitability analysis
CN112561702A (en) Generalized stock price prediction method based on multitask asymmetric adjacent support vector machine
Bahadur et al. Dimenslon estlmatlon of equlty markets
Ziesmer et al. Integrating fundamental model uncertainty in policy analysis: A Bayesian averaging approach combining CGE-models with metamodeling techniques
Berteloot et al. A novel credit rating migration modeling approach using macroeconomic indicators
Loonat et al. Learning zero-cost portfolio selection with pattern matching
CN109829745A (en) Business revenue data predication method, device, computer equipment and storage medium
Pomulev et al. Methodological aspects of credit portfolio management in financing innovative projects
Jimenez et al. A semi-supervised learning approach for variance reduction in life insurance
Mirbolouki et al. A parameter-free approach for estimating the quality and quantity of the right and left returns to scale in data envelopment analysis
Rozite et al. Towards a financial cycle for the US, 1973–2014
Tu et al. A novel grey relational clustering model under sequential three-way decision framework
CN109697528A (en) Business revenue data predication method, device, computer equipment and storage medium
Kanda et al. Forecasting South African inflation using non-linearmodels: a weighted loss-based evaluation
Wang et al. An integrative extraction approach for index-tracking portfolio construction and forecasting under a deep learning framework
Zhang et al. Portfolio Selection with Regularization

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20210326

RJ01 Rejection of invention patent application after publication