CN-122022858-A - Dynamic correction method and system for non-marketing stock right heterogeneous data value characteristics

CN122022858ACN 122022858 ACN122022858 ACN 122022858ACN-122022858-A

Abstract

The invention belongs to the technical field of financial science and technology and data processing, and particularly discloses a dynamic correction method and a system for non-marketing stock right heterogeneous data value characteristics, wherein the method comprises the steps of acquiring heterogeneous time sequence data of a target enterprise, generating an countermeasure network through double channels, and separating a basic trend data component and an instantaneous fluctuation data component; the method comprises the steps of generating and initializing value features from a feature definition library based on basic trend data components, constructing a value feature calculation instance set, quantizing instantaneous fluctuation data components into external environment quantization vectors, simulating dynamic iteration of the set to update dynamic weight scores, determining weights according to the dynamic weight scores and carrying out weighted fusion to generate dynamic estimation values, obtaining external transaction price calculation estimation value deviation, and optimizing interaction parameters among the value feature calculation instances or generating logic of a combination optimization feature definition library according to the value feature calculation value deviation. The invention realizes the self-adaptive calibration and the quantitative data processing of the non-marketing stock right heterogeneous data value evaluation model.

Inventors

BAO WENSHENG
Zhai Yishu
LI XIAOWEI
ZHU DAWEI
TENG FEI

Assignees

青岛大学资产经营有限公司

Dates

Publication Date: 20260512
Application Date: 20260206

Claims (10)

1. S1, acquiring heterogeneous time sequence data of a target enterprise, and inputting the heterogeneous time sequence data into a double-channel generation countermeasure network comprising a steady-state data extraction unit and an instantaneous pulse perception unit which are mutually restricted so as to generate a separated basic trend data component and an instantaneous fluctuation data component; S2, generating a plurality of value feature calculation examples based on a plurality of value features in a feature definition library and by utilizing basic trend data components, initializing dynamic weight scores and interaction parameters for each value feature calculation example to construct an initial value feature calculation example set; S3, carrying out quantization processing on the instantaneous fluctuation data component to generate an external environment quantization vector, applying the external environment quantization vector to a value feature calculation instance set, driving the value feature calculation instance set to iterate to a numerical convergence state according to a set of dynamic iteration convergence models defining mutual exclusion inhibition and cooperative enhancement interaction, and outputting updated dynamic weight scores of all value feature calculation instances after iteration; S4, determining the contribution weight of each value characteristic calculation example according to the updated dynamic weight score, and carrying out weighted fusion by combining the characteristic values of the contribution weights to generate a dynamic estimation result of the target enterprise; S5, obtaining an external real transaction price to calculate an estimated value deviation, and executing parameter optimization operation for adjusting interaction parameters or executing feature extraction rule construction optimization operation for modifying feature definition according to whether the estimated value deviation exceeds a preset deviation threshold.
2. The method of claim 1, wherein generating the separated base trend data component and the transient fluctuation data component comprises: extracting a first data component representing a steady-state trend from heterogeneous time sequence data by using a steady-state data extraction unit; Capturing a second data component characterizing the transient pulse from the heterogeneous timing data using the transient pulse sensing unit; In the countermeasure training, the countermeasure disturbance is applied to the extraction process of the steady-state data extraction unit through the second data component, the separation effect of the second data component is verified by utilizing the output of the steady-state data extraction unit, and after iterative optimization, the first data component is converged to be a basic trend data component, and the second data component is converged to be an instantaneous fluctuation data component.
3. The method for dynamically modifying value characteristics of non-equity stock weight data according to claim 1, wherein constructing the initial value characteristic calculation instance set comprises: reading a plurality of value features from a feature definition library, each value feature corresponding to a feature extraction logic; According to the feature extraction logic and the basic trend data component, calculating and generating an initial feature value corresponding to each value feature; And allocating an independent memory structure for each value feature calculation example with an initial feature value, storing initial dynamic weight scores and initial interaction parameters comprising mutual exclusion suppression coefficients and cooperative enhancement coefficients in the memory structure, and forming an initial value feature calculation example set by all value feature calculation examples which finish initialization.
4. A method for dynamically modifying a value characteristic of non-equity stock weight data according to claim 3, wherein driving the value characteristic calculation instance set to iterate to a value convergence state comprises: Calculating external fluctuation pressure born by each value feature calculation example according to the external environment quantization vector and the feature value of each value feature calculation example in the value feature calculation example set; Calculating feature similarity between any two value feature calculation examples based on interaction parameters of each value feature calculation example, and calculating weight inhibition quantity between any two value feature calculation examples according to mutual exclusion inhibition coefficients when the feature similarity is higher than a preset mutual exclusion inhibition threshold; Inquiring a data association rule base defining complementarity between features, and calculating weight gain quantity between two value feature calculation examples according to the cooperative enhancement coefficient when the two value feature calculation examples accord with the complementarity rule; Substituting external fluctuation pressure, weight inhibition quantity and weight gain quantity as input parameters into a coupled differential equation set configured with a characteristic attenuation term and a characteristic growth term to solve; And iteratively updating the dynamic weight score of each value feature calculation example until the dynamic weight score change rate of all value feature calculation examples is lower than a preset convergence threshold value, and judging that the value convergence state is reached.
5. The method for dynamically modifying a value characteristic of non-equity stock weight data according to claim 1, wherein generating a dynamic valuation result for a target enterprise comprises: normalizing the updated dynamic weight scores of all the value feature calculation examples in the value feature calculation example set to generate a group of contribution weights for representing relative importance; Converting the feature value corresponding to each value feature calculation example into a standardized value factor through a preset estimated value mapping function; And carrying out weighted fusion on the contribution weight of each value characteristic calculation example and the corresponding standardized value factor so as to calculate and obtain the comprehensive value index of the target enterprise, and generating a dynamic estimation result based on the comprehensive value index.
6. The method for dynamically modifying a value characteristic of non-equity data in accordance with claim 5, further comprising, after generating the dynamic valuation result for the target enterprise: Comparing the dynamic estimation result with the historical estimation result to generate an estimation fluctuation sequence; Performing anomaly detection on the estimated value fluctuation sequence, and if abnormal fluctuation exceeding a preset fluctuation threshold is detected, generating a risk early warning signal; the transient fluctuation data component that caused this abnormal fluctuation is traced back and marked as a high weight stress event.
7. The method for dynamic modification of non-market share-right heterogeneous data value features according to claim 1, wherein performing a parameter optimization operation for adjusting interaction parameters comprises: Calculating an estimated value deviation based on the dynamic estimated value result and an external real transaction price, and constructing a loss function according to the estimated value deviation; calculating gradients of the loss function on mutual exclusion inhibition coefficients and collaborative enhancement coefficients contained in each value characteristic calculation example in the value characteristic calculation example set respectively; And updating the mutual exclusion inhibition coefficient and the cooperative enhancement coefficient by adopting a gradient descent method according to the gradient so as to enable the value of the loss function to tend to be reduced.
8. The method for dynamic modification of non-market share-right heterogeneous data value features according to claim 1, wherein performing feature extraction rule construction optimization operations for modifying feature definitions comprises: selecting at least one target feature definition from a feature definition library according to the amplitude and duration of the estimated variation; Performing a parameter perturbation operation on the target feature definition to modify its extracted logic parameters, or performing a linear weighted combination operation to construct a new feature definition comprising weighted sum computation logic on the output values of the two associated feature definitions; And adding the new feature definitions generated after the parameter disturbance operation or the linear weighted combination operation to a feature definition library.
9. The method for dynamically modifying a value characteristic of non-equity data in the market according to claim 8, further comprising, after the feature extraction rule construction optimization operation: When the step of constructing the initial value feature calculation example set is executed next time, loading and initializing new feature definitions from the updated feature definition library to generate corresponding value feature calculation examples and combining the value feature calculation examples into the value feature calculation example set; In the subsequent dynamic iteration process, the new value feature calculation example participates in mutual exclusion inhibition and collaborative enhancement relation calculation, the dynamic weight score of the new value feature calculation example is regulated by a dynamic selection mechanism, and if the new value feature calculation example is invalid, the dynamic weight score of the new value feature calculation example is inhibited, so that the new value feature calculation example is automatically eliminated in iteration.
10. A system for dynamically modifying a value characteristic of non-marketable equity heterogeneous data, comprising: the double-channel generation countermeasure module acquires heterogeneous time sequence data of a target enterprise, and inputs the heterogeneous time sequence data into a double-channel generation countermeasure network comprising a steady-state data extraction unit and an instantaneous pulse sensing unit which are mutually restricted so as to generate a separated basic trend data component and an instantaneous fluctuation data component; The feature set construction module is used for generating a plurality of value feature calculation examples based on a plurality of value features in a feature definition library and by utilizing basic trend data components, initializing dynamic weight scores and interaction parameters for each value feature calculation example to construct an initial value feature calculation example set; The set dynamic iteration module is used for carrying out quantization processing on instantaneous fluctuation data components to generate external environment quantization vectors, applying the external environment quantization vectors to the value characteristic calculation instance set, driving the value characteristic calculation instance set to iterate to a numerical value convergence state according to a set of dynamic iteration convergence models defining mutual exclusion inhibition and cooperative enhancement interaction, and outputting updated dynamic weight scores of all value characteristic calculation instances after iteration; the dynamic estimation calculation module is used for determining the contribution weight of each value characteristic calculation example according to the updated dynamic weight score, and carrying out weighted fusion by combining the characteristic values of the contribution weight to generate a dynamic estimation result of a target enterprise; And the double-layer optimization driving module acquires an external real transaction price to calculate an estimated value deviation, and executes parameter optimization operation for adjusting interaction parameters or executes feature extraction rule construction optimization operation for modifying feature definition according to whether the estimated value deviation exceeds a preset deviation threshold.

Description

Dynamic correction method and system for non-marketing stock right heterogeneous data value characteristics Technical Field The invention belongs to the technical field of financial science and technology and data processing, and relates to a dynamic correction method and a system for non-marketing stock right heterogeneous data value characteristics. Background The non-marketing share right valuation is a key link in financial activities such as private share right investment, parallel purchase reorganization, financial report and the like, and is characterized in that the intrinsic value of an enterprise without public transaction is reasonably evaluated. The data of non-marketing enterprises has typical isomerism, sparsity and nonstandard characteristics, and has wide data sources including regular financial reports, irregular management bulletins, industry research reports, real-time network public opinion and the like, and the data has great differences in time frequency, credibility and data structure. Existing non-marketable share-right estimation techniques typically employ a linear computational model based on static assumptions or a comparative analysis method based on transverse benchmarks, etc., which are highly dependent on the expertise and subjective judgment of an analyst. With the development of data science and technology, some methods begin to attempt to introduce a machine learning model, and a regression model is constructed to evaluate by performing feature engineering on multi-source data, extracting a series of indexes such as a revenue growth rate, a profit margin, a marketing rate and the like. These methods typically simply clean and align the data from different sources and then input it into a model with fixed or periodically updated weights for calculation. However, prior art solutions have significant drawbacks in dealing with complex scenarios that are not equity. On the one hand, the processing mode of heterogeneous data is rough, and it is often difficult to effectively distinguish the long-term business data with high credibility from the short-term market information filled with noise, so that the model is easy to be interfered by irrelevant information or has slow response to key changes. On the other hand, the importance of each value feature in the estimation model, namely the weight, is usually set in a static state or a slowly varying state, so that the instantaneous change of the internal and external environments of an enterprise cannot be reflected dynamically, and when an emergency occurs in the market or industry logic changes, the solidified model structure is difficult to make numerical calculation matched with real-time data distribution. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a method for dynamically correcting the value characteristics of non-marketing stock right heterogeneous data, which comprises the following steps of S1, acquiring heterogeneous time sequence data of a target enterprise, and inputting the heterogeneous time sequence data into a double-channel generation countermeasure network comprising a steady-state data extraction unit and an instantaneous pulse sensing unit which are mutually restricted so as to generate a separated basic trend data component and an instantaneous fluctuation data component. S2, generating a plurality of value feature calculation examples based on a plurality of value features in a feature definition library and by utilizing basic trend data components, initializing dynamic weight scores and interaction parameters for each value feature calculation example to construct an initial value feature calculation example set. S3, carrying out quantization processing on the instantaneous fluctuation data component to generate an external environment quantization vector, applying the external environment quantization vector to the value feature calculation instance set, driving the value feature calculation instance set to iterate to a numerical convergence state according to a set of dynamic iteration convergence models defining mutual exclusion inhibition and cooperative enhancement interaction, and outputting updated dynamic weight scores of all value feature calculation instances after iteration. And S4, determining the contribution weight of each value characteristic calculation example according to the updated dynamic weight score, and carrying out weighted fusion by combining the characteristic values of the contribution weights to generate a dynamic estimation result of the target enterprise. S5, obtaining an external real transaction price to calculate an estimated value deviation, and executing parameter optimization operation for adjusting interaction parameters or executing feature extraction rule construction optimization operation for modifying feature definition according to whether the estimated value deviation exceeds a preset deviation threshold. The