CN118152971A

CN118152971A - Method and system for pairing target objects and models based on kernel functions by multiple factors

Info

Publication number: CN118152971A
Application number: CN202311820607.XA
Authority: CN
Inventors: 艾永华
Original assignee: Shenzhen Fangfeidao Technology Co ltd
Current assignee: Shenzhen Fangfeidao Technology Co ltd
Priority date: 2023-12-26
Filing date: 2023-12-26
Publication date: 2024-06-07

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a method and a system for pairing target objects and models based on a kernel function by multiple factors so as to improve the pairing reliability. The method comprises the following steps: assigning an initial target object to any model, mapping corresponding factor group data in a sample of the target object to a higher dimension based on a kernel function, and constructing a hyperplane on a feature space of the higher dimension to classify the hyperplane to obtain corresponding model parameters; and evaluating the scores of the fitting degree and the correlation of each target object in each model through a regression comparator, and then adjusting the unique models corresponding to each target object according to the scores.

Description

Method and system for pairing target objects and models based on kernel functions by multiple factors

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a method and a system for pairing target objects and models based on a kernel function by multiple factors.

Background

Funds are one of the main financial objects of the invested individuals and institutions.

Investment institutions mostly purchase in experience and intuition in selecting funds. With the development of artificial intelligence and other technologies, how to screen specific models to match funds and simulate the net value of the models can reduce the risk of investment.

In the conventional method, a certain commercial factor series or an own factor series is directly used for carrying out orthogonal fusion on factors, and then statistical regression is carried out, so that the disadvantage is that models of the foundation to be specified are always different in size, and the common feature is that: the orthogonality factors obtained purely mathematically over different calculation time periods may vary widely, and evaluation of the long-term performance of the fund based on these widely varying factors may cause significant errors. Meanwhile, the model obtained by the traditional method is easy to be over-fitted, the model performs very well in a sample, but the performance error outside the sample is larger.

Thus, although the interest in the investment benefits of funds is a non-technical problem, how to develop a mechanism to match funds based on models and screen out the factors that play the main role is a technical problem that the funds forecast itself and other fields that can be extended to face together.

Disclosure of Invention

The invention aims to disclose a method and a system for pairing target objects and models based on a kernel function by multiple factors so as to improve the pairing reliability.

In order to achieve the above object, the present invention discloses a method for pairing a target object and a model based on a kernel function, which is characterized by comprising:

determining an initial number, a final number and factor distribution in each factor group; wherein the final number is less than or equal to the initial number;

Determining a plurality of models corresponding to the initial factor groups one by one;

Assigning an initial target object to any model, mapping corresponding factor group data in a sample of the target object to a higher dimension based on a kernel function, and constructing a hyperplane on a feature space of the higher dimension to classify the hyperplane to obtain corresponding model parameters;

Evaluating the scores of the fitting degree and the correlation of each target object in each model through a regression comparator, determining a unique model corresponding to each target object according to the highest score, judging whether the number of the models corresponding to at least one target object according to the highest score is smaller than or equal to the final number, and if so, determining the pairing relation among the models, the factor groups and the target objects according to the highest score; otherwise, the model corresponding to at least one target object according to the highest score is partially eliminated until the number of reserved models is equal to the final number, and then the target object corresponding to the eliminated model is re-paired to the model corresponding to the highest score in the reserved models.

Optionally, the target object is a fund. Meanwhile, the model corresponding to at least one target object according to the highest score is partially eliminated according to the average income and the quantity of the foundation.

Preferably, in determining the factor distribution within each factor group, the corresponding factor group is generated according to the following constraints:

A. the upper and lower limits of the number of factors in each factor group;

B. an upper limit on the number of times a factor occurs in a different set of factors.

C. And selecting 2-3 factors from the factor group of the same type as basic factors, and generating a new factor group by adding other factors in the factor pool in a pseudo-random manner.

Optionally, the kernel function includes: gaussian kernel, linear kernel and quadratic polynomial kernel. Correspondingly, the model comprises at least one SVR model; in the SVR model, after a kernel function is determined, training data including a factor value and a net value of the foundation are input, and the net value of the foundation and a corresponding factor are subjected to fitting calculation to obtain corresponding model parameters.

Preferably, the method of the present invention also generates a revenue curve for the corresponding fund comprising the following four parts, based on the finally determined model-factor set-target object:

The first part is a historical net value curve of the fund corresponding to the training set interval in the sample;

the second part is a historical net value curve of the fund corresponding to the test set interval in the sample;

The third part is a future curve outside the net value of the foundation history based on rolling fitting;

The fourth part is a history curve outside the net value of the foundation history based on rolling fit.

Further, after determining the final model-factor set-target object, it further comprises:

And calculating the fitting matching degree of each factor group in different time windows, and carrying out weighted matching on the results of the different time windows to obtain a foundation profit curve which is smooth in all the time windows.

The invention also discloses a system for pairing target objects and models based on kernel functions by multiple factors, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the method when executing the computer program.

The invention has the following beneficial effects:

1. The four-tuple of model-factor group-target object-kernel function was initially constructed. Any element in the quadruple is changed, and the corresponding model fitting result is also changed, so that a large number of pairing relations for screening by the regression comparator are provided. Moreover, the final result will determine which target object (i.e. fund) belongs to which model, and also determine which factor group is the best match, thus ensuring the usability of the pairing result.

2. In the scoring process through the regression comparator, any target object is matched with other models trained by similar objects, the matching result is scored, and the best matching model is determined according to the scoring result, so that the reliability of pairing is ensured; meanwhile, the severity of initial pairing of each target object and the model before training is reduced, and the final result has both completeness and precision.

3. The final reserved model generality can be controlled through the final number of the reasonably set factor groups.

4. In the processing process, the score of the regression comparator is included, and the score index is integrated with the fitting degree of the model; therefore, the finally screened model can be subjected to simulation based on the screened factor group and the kernel function in the direction of the paired target object, and the problems of over fitting and the like do not exist.

The invention will be described in further detail with reference to the accompanying drawings.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

FIG. 1 is a flow chart of a method for pairing target objects and models based on kernel functions according to an embodiment of the invention.

Fig. 2 to 6 are schematic diagrams for comparing the kernel function of the present embodiment with the hyperplane and the conventional orthogonality factor.

Detailed Description

Embodiments of the invention are described in detail below with reference to the attached drawings, but the invention can be implemented in a number of different ways, which are defined and covered by the claims.

Example 1

The embodiment discloses a method for pairing target objects and models based on a kernel function by multiple factors, which comprises the following steps as shown in fig. 1:

step S1, determining the initial number and the final number of factor groups and factor distribution in each factor group; wherein the final number is less than or equal to the initial number.

In this step, when the target object is a fund, the corresponding factor may be a characteristic parameter in BARRA models commonly used in the industry, or may be a consistent expectation, a price, a practical attribute of a fund manipulator, or a deep learning-based factor. In the present invention, one factor group includes at least two factors.

And S2, determining a plurality of models corresponding to the initial factor groups one by one.

In this step, preferably, when determining the factor distribution within each factor group, the corresponding factor group is generated according to the following constraints:

A. the upper and lower limits of the number of factors within each factor group.

Also for example: the model is for 500-digit incremental funds, and the associated constraints may set in factors whether or not it is necessary to introduce a setting for the initial factor of the exponential 300 calculation synthesis.

The same model using different kernel functions is also considered as different models in association with subsequent steps. Thus, in this embodiment, the multiple models in this step may be different classes, the same class (a class typically includes multiple different models, for example, a model based on a convolutional neural network may be derived based on different logic architectures inside, or even the same model).

And S3, distributing an initial target object to any model, mapping corresponding factor group data in a sample of the target object to a higher dimension based on a kernel function, and constructing a hyperplane on a feature space of the higher dimension to classify the hyperplane to obtain corresponding model parameters.

In the process of distributing the initial target objects to the model in the step, the initial target objects can be paired according to experience of an applied scene (corresponding constraint conditions can be set), so that the subsequent difference degree based on score readjustment is reduced, and meanwhile, the average matching degree of the whole target objects can be improved. As an alternative, a random allocation may also be used such that one model may pair at least two target objects. A simple or special way of allocation is to allocate all the target objects (funds) separately for each model. Or empirical data after convergence over a period of practice may exclude many combinations. Taking funding as an example, the results indicate that: contradictory combinations of fund features (including policies of the fund, including the reconciliation reports on the style of the fund manager) and factor features (features of factors within the factor group) can be directly excluded; for example: the fund policy is a broad base index enhancement, then the fund need not be assigned to a model that is dominated by some industry index factor.

Wherein, in the pairing relation of the step, one target object can pair a plurality of initial models; preferably, the upper pairing limit of each target object may also refer to constraint condition C of the factor set.

In this step, as a comparison: if some data are distributed on a two-dimensional plane as shown in fig. 2, if points in fig. 2 are given a classification and prediction manner by linear regression, the result may be as shown in fig. 3, and the upper and lower sides of the middle line are used to represent different classifications: the middle line can then be regarded as a model of the linear classification.

In practice, however, the points in fig. 2 may be classified more reasonably as shown in fig. 4, and the circles in the graph are more stable in terms of classification performance and prediction, and the circles are updated to a high-dimensional space by a certain transformation rule on the points in fig. 2 based on the kernel function of the step, for example, the following graph becomes a three-dimensional distribution as shown in fig. 5. Thereby, it is possible to obtain: the point in fig. 2 can be easily divided into two parts from the Z-axis dimension.

Mathematically, a number of planes can be found that divide the points into two parts, where all points are selected to have the shortest distance to a plane, called a hyperplane, and the square or other distance. Such as the plane shown in fig. 6. This hyperplane is represented as a circle in the x-y plane (where x may be a factor set feature of a multiple factor fusion and y may be the net value of the fund). The mapping function that forms this hyperplane is the kernel function. There are mathematical requirements for kernel functions in which gaussian kernels can map planes to infinity.

In this step, optional kernel functions include, but are not limited to: gaussian kernel, linear kernel, and quadratic polynomial kernel, etc. The corresponding model may employ a SVR (Support Vector Regression ) model; in the SVR model, after a kernel function is determined, training data including a factor value (X) and a net value (Y) of the foundation are input, and fitting calculation is carried out on the net value of the foundation and the corresponding factor (X-Y) to obtain corresponding model parameters.

S4, evaluating scores of fitting degree and correlation of each target object in each model through a regression comparator, determining unique models corresponding to each target object according to the highest scores, judging whether the number of the models corresponding to at least one target object according to the highest scores is smaller than or equal to the final number, and if so, determining pairing relations among the models, the factor groups and the target objects according to the highest scores; otherwise, the model corresponding to at least one target object according to the highest score is partially eliminated until the number of reserved models is equal to the final number, and then the target object corresponding to the eliminated model is re-paired to the model corresponding to the highest score in the reserved models.

Optionally, in this step, the regression comparator may further integrate correlation (or called similarity, rather than minimum variance matching, which is commonly used in the industry) based on the regression result and the actual data, and the specific calculation method of the correlation may use the pearson correlation coefficient calculation method, the convergence speed, and so on to perform the phase synthesis scoring. If necessary, it is also possible to distinguish between the sample performance in the test set and the sample performance in the training set. The fitting degree refers to the matching degree of the prediction result of the model and the actual occurrence, and the common fitting degree test method comprises the following steps: residual sum of squares test, chi-square test and linear regression test, etc.

Taking a target object as a fund as an example, in the elimination process, elimination can be carried out according to the average income and the number of the fund; models with the lowest average benefit and low number of fund pairings are typically prioritised.

Further, the method of the embodiment also generates a benefit curve of the corresponding fund comprising the following four parts according to the finally determined model-factor group-target object:

The first part is the historical net value curve of the fund corresponding to the training set interval in the sample. Wherein, whether the combination of the factors correctly reflects the real performance of the interval foundation can be judged by regression coefficients, p values for judging the correlation of the single factors and the whole, and the like.

The second part is the historical net value curve of the fund corresponding to the test set interval in the sample. The fitting degree analysis of the factor combination on the interval can judge whether the factor combination has the fitting phenomenon or not, and can accurately judge whether the foundation frequently has style drift or not.

The third part is a future curve outside the net value of the fund history based on rolling fit. Therefore, through the judgment of the style of the foundation, after the basic robustness of the style of the foundation is determined, the screened exposure factor combination is used for predicting the unpublished net value of the foundation, so that the net value prediction can be obtained before the foundation is published, and investment decision is facilitated.

The fourth part is a history curve outside the net value of the foundation history based on rolling fit. In some cases, because funds may be released for a short period of time, for example, half a year; the combination of factors exposed for half a year can be used to reversely generate a revenue curve before the creation of the fund, so that the basic judgment can be made on the long-term performance of the fund, and the reference basis can be provided for investment decision.

and calculating the fitting matching degree of each factor group in different time windows, and carrying out weighted matching on the results of the different time windows to obtain a foundation profit curve which is smooth in all the time windows. Preferably, the different models are smoothed using uniformly set time windows.

Example 2

The invention discloses a system for pairing target objects and models based on a kernel function with multiple factors, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the method corresponding to the embodiment when executing the computer program.

In summary, the method and system disclosed by the embodiment of the invention have at least the following beneficial effects:

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for pairing a target object and a model based on a kernel function by multiple factors, comprising:

2. The method of claim 1, wherein the target object is a fund.

3. The method of claim 2, wherein the mechanism for partial elimination of the model corresponding to the at least one target object according to the highest score is elimination according to the average benefit and quantity of the fund.

4. A method according to claim 3, characterized in that in determining the factor distribution within each factor group, the corresponding factor group is generated according to the following constraints:

A. the upper and lower limits of the number of factors in each factor group;

5. The method of any one of claims 1 to 4, wherein the kernel function comprises: gaussian kernel, linear kernel and quadratic polynomial kernel.

6. The method of claim 5, wherein the model comprises at least one SVR model; in the SVR model, after a kernel function is determined, training data including a factor value and a net value of the foundation are input, and the net value of the foundation and a corresponding factor are subjected to fitting calculation to obtain corresponding model parameters.

7. Method according to any of claims 2 to 4, characterized in that, from the model-factor set-target object finally determined, a revenue curve of the corresponding fund is generated comprising the following four parts:

8. The method of claim 7, further comprising, after determining the final model-factor set-target object:

9. A system for multi-factor kernel-based pairing of target objects and models, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method of any of the preceding claims 1 to 8 when executing the computer program.