CN-121981833-A - Method and system for dynamically estimating asset market based on deep learning

CN121981833ACN 121981833 ACN121981833 ACN 121981833ACN-121981833-A

Abstract

The invention discloses an asset market dynamic estimation method and system based on deep learning, wherein the method realizes dynamic update of asset value through self-adaptive multi-source heterogeneous data fusion and time sequence dependent modeling, and comprises the steps of S1, ubiquitous heterogeneous data integration, S2, dynamic noise cleaning and feature alignment, S3, multi-order coupling estimation framework construction, S4, time-varying feature topological mapping, S5, an opposite dynamic training mechanism, S6, joint estimation state output and correction, and S7, an estimation robustness verification module. The invention realizes millisecond multi-dimensional fusion of satellite images, logistics information and high-frequency transaction data through a space-time tensor compiling technology of multi-source heterogeneous data, eliminates the problem of synchronous hysteresis of derived signals in the traditional architecture, realizes self-evolution of factor interaction paths when market structures suddenly change through reconfigurable characteristic topological evolution, and solves the response hysteresis of fixed model topology to industrial rotation events.

Inventors

MA LIHUA
XU FENG
GE PEIWEN
ZHANG QI
XU ZHIWEI
XU KAI

Assignees

上海申威资产评估有限公司

Dates

Publication Date: 20260505
Application Date: 20251222

Claims (9)

1. The method for dynamically estimating the asset market based on deep learning realizes the dynamic update of the asset value through the self-adaptive multi-source heterogeneous data fusion and time sequence dependent modeling, and is characterized by comprising the following steps: S1, accessing and integrating a structured market data source and an unstructured derivative data source related to a target asset in real time; S2, performing asynchronous time stamp calibration on the heterogeneous data, and performing feature level outlier suppression through a depth automatic encoder architecture to synchronously generate a feature tensor sequence with time domain continuity; s3, establishing a hierarchical deep learning architecture which comprises a vertical coupling structure of a macroscopic period sensing layer, a mesoscopic industry conducting layer and a microscopic asset response layer; S4, constructing a dynamic feature interaction engine in the estimation framework, automatically identifying the relative weights of key driving factors under different market states by utilizing a multi-head self-attention mechanism, and generating a feature combination path evolving along with a market structure through topology reconstruction; S5, model training is implemented based on the generated countermeasure network frame, the generator network continuously generates a synthesized data sequence conforming to market dynamics, and the discriminator network combines real-time market feedback to identify the distribution difference of the synthesized data and the real market path; s6, outputting a joint estimation vector formed by a basic value trend item, a market emotion deviation item and a fluidity correction item; and S7, synchronously running at least two independently trained estimated paths in an operation environment, performing confidence filtering on the divergence estimated result through a Monte Carlo game strategy, and finally outputting a dynamic estimated interval and a stability score which accord with a preset risk threshold.
2. The deep learning based asset market dynamic valuation method of claim 1, wherein the unstructured derivative data source in S1 comprises: The social media public opinion emotion tendency index is analyzed through a natural language processing engine and comprises an investor emotion polarity quantized value and a topic transmission heat attenuation coefficient; Based on the supervision impact probability matrix generated by the policy text semantic vector, identifying inter-industry conduction intensity of the policy keywords by using word embedding space mapping; the supply chain activity index extracted by satellite remote sensing image features specifically comprises a port cargo stacking density time sequence matrix and a factory heat radiation intensity change rate.
3. The method for dynamically estimating an asset market based on deep learning according to claim 2, wherein the calculation of the social media public opinion emotion tendency index adopts a dynamic polarity quantization expression as follows: , wherein, Represents social media opinion emotion tendency index, alpha is authoritative media weight coefficient, For user iii to influence the decay factor at ttt, Publishing text for user i The polarity value of the [ -1,1] interval output by the emotion analysis model, beta is the market heat regulating coefficient, Is the mobile standard deviation of the price change rate of the asset in the time window, The method is characterized in that the matrix representation of a topic word vector set is that lambda is topic forgetting rate, text emotion and market behavior dynamics characteristics are fused, network flow interference is eliminated through a dynamic attenuation mechanism, and meanwhile, price fluctuation rate is introduced As a nonlinear amplification factor, the emotion weight under extreme conditions is automatically improved.
4. The method for dynamically estimating an asset market based on deep learning of claim 1, wherein said deep auto encoder architecture in S2 employs a dual path filtering mechanism: The method comprises the steps that a characteristic smooth constraint function is established through a time convolution network in a first path, and instantaneous outliers caused by market pulse events are restrained; a second path is provided with a memory enhancement type variational encoder, a dynamic probability distribution boundary is constructed in a potential feature space, and nonlinear compression is implemented on an abnormal mode exceeding a historical experience threshold; and generating the time domain continuity characteristic tensor sequence by the dual-path output through a gating characteristic fusion layer.
5. The method of claim 1, wherein the operation of the macrocycle aware layer in S3 comprises: Constructing a multi-head space-time correlation diagram of global macro economic factors, wherein nodes represent national key economic indexes, and side weights represent the flow intensity of the transnational capital; Capturing a cross-border risk infection path by using a multi-order neighborhood aggregation graph rolling network, and dynamically adjusting a risk absorption coefficient in the crisis transmission process through an attention gating mechanism; and outputting the macroscopic pressure index tensor as an input condition of the conductive layer in the mesoscopic industry.
6. The method for dynamically estimating an asset market based on deep learning according to claim 1, wherein the dynamic feature interaction engine in S4 is implemented as follows: Establishing a market state classifier, and dividing the market into three states of low fluctuation/structural transition/high fluctuation based on implicit fluctuation rate curved surfaces and transaction amount distribution; activating a differential attention head combination under different states, wherein the attention weight of a financial quality factor is amplified in a low-fluctuation state, the interaction strength of industrial rotation factors is enhanced in a structural transition state, and the topological priority of a fluidity impact factor is improved in a high-fluctuation state; And triggering a characteristic path reconstruction instruction when the market state transitions.
7. The deep learning based asset market dynamic valuation method of claim 1, wherein the generating of the challenge network framework in S5 comprises a challenge sample reinforcement mechanism: the generator network receives the extreme market scenario vector simulated by Monte Carlo and outputs a synthetic data stream conforming to the dynamics rule of the three-level coupling architecture; the discriminator network introduces a real market pressure test result as a negative sample centroid, and constructs a dynamic decision boundary through Wasserstein distance measurement; the generation of the countermeasures loss function comprises a market mechanism change detection module, and the generation frequency of the synthesized data is automatically improved when the continuous expansion of the historical return errors is monitored.
8. An asset market dynamic valuation system based on deep learning for implementing the method of any of claims 1-7, comprising: The multi-source heterogeneous data interface layer is connected with a structured market data source and a non-structured derivative data source in parallel in real time, the structured data source comprises a multi-exchange cascade quotation and industry basic surface factor library, the non-structured data source comprises a satellite image feature extractor and a public opinion semantic coding pipeline, and the layer is provided with a data stream health degree dynamic diagnosis unit; the dynamic cleaning alignment module receives an original data stream of the multi-source heterogeneous data interface layer, solves the problem of cross-time zone asynchronism through the timestamp corrector, implements noise suppression of industry fluctuation rate self-adaption through the outlier fuse, and finally compiles a space-time characteristic tensor sequence with aligned dimensions; the multi-order coupling modeling unit is used for constructing a macroscopic-mesoscopic-microscopic vertical layered architecture, a macroscopic layer is used for modeling cross-country risk infection through a graph neural network, a microscopic layer is used for capturing asset-specific fluctuation through a time attention mechanism, and dynamic information exchange is realized between layers through adjustable coupling coefficients; The feature topology evolution module dynamically reconstructs a feature interaction path and comprises a scene recognition unit based on a market state classifier, a multi-head router for driving factor interaction weight optimization and a life cycle manager for executing feature combination Darwin evolution; The countermeasure training optimizing unit is internally provided with a generator and a discriminator countermeasure network, the generator generates a synthesized market scene conforming to the financial physical rule, the discriminator adjusts a decision boundary according to real-time market feedback, and the arbitration center triggers an increment optimizing instruction of the modeling engine; The joint estimation output module is used for generating a joint estimation vector formed by basic value trend anchors, market emotion offsets and fluidity correction items and outputting a dynamic estimation confidence interval with probability density; And the robustness control unit runs at least two sets of independent modeling paths, injects the modeling paths into the disturbance of the market structure through the Monte Carlo sandbox, carries out the confidence weighting fusion of the multipath result, and outputs a stability verification score and a risk adjustment estimated value interval.
9. The deep learning based asset market dynamic valuation system of claim 8, wherein the system achieves dynamic valuation by a collaborative mechanism of: the dynamic cleaning alignment module outputs a space-time characteristic tensor sequence to the multi-order coupling modeling unit to perform coupling modeling with the characteristic topology evolution module; The countermeasure training optimization unit periodically injects countermeasure training sample driving parameter optimization into the multi-order coupling modeling unit; The robustness control unit monitors the output divergence degree of the joint estimation output module in real time and triggers the multi-order coupling modeling unit to reconstruct online; The characteristic topology evolution module perceives the market state transition signal and pushes the signal to the countermeasure training optimization unit to generate a targeted countermeasure sample.

Description

Method and system for dynamically estimating asset market based on deep learning Technical Field The invention belongs to the field of dynamic evaluation, and particularly relates to an asset market dynamic evaluation method based on deep learning. Meanwhile, the invention also relates to an asset market dynamic valuation system based on deep learning. Background The current asset estimation field mainly depends on two technical systems, namely a static pricing model based on a financial engineering theory and a dynamic prediction framework based on machine learning, and the representative method comprises the steps of integrating macroscopic economic indexes and company basic surface data through linear regression to construct a static weight distribution system, adopting a time sequence deep learning model to process a historical price sequence by adopting an LSTM/GRU network to capture the inertia motion rule of an asset price, combining a traditional discount cash flow model and public opinion emotion analysis to generate a weighted synthesis result of a basic value and an emotion adjustment value, and realizing the primary transformation from a pure theory model to a data driving model by the technology, so that the technology has basic applicability under a stable market environment. However, along with the aggravation of the complexity and volatility of the financial market, the prior art system cannot effectively meet the requirements, such as the situation that a traditional system is in a splitting state for processing structured transaction data and unstructured derivative data, a satellite image and a real-time signal of supply chain logistics information cannot be effectively integrated into an estimation system, so that response to sudden unbalance of supply and demand is delayed and overtime, and a main stream model adopts a characteristic interaction path with fixed topology. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides an asset market dynamic estimation method and system based on deep learning, which breaks through a static framework of the traditional estimation technology, constructs a dynamic coupling framework with self-evolution capability, drives an estimation model to dynamically reconstruct through heterogeneous data real-time fusion engine, combines an countermeasure training and multipath verification mechanism, still keeps high-robustness output under extreme market conditions, and provides a bottom technical paradigm innovation for high-frequency transaction and risk management. In order to achieve the above purpose, the present invention provides the following technical solutions: an asset market dynamic valuation method based on deep learning, which realizes dynamic update of asset value through self-adaptive multi-source heterogeneous data fusion and time sequence dependent modeling, comprises the following steps: S1, accessing and integrating a structured market data source and an unstructured derivative data source related to a target asset in real time; S2, performing asynchronous time stamp calibration on the heterogeneous data, and performing feature level outlier suppression through a depth automatic encoder architecture to synchronously generate a feature tensor sequence with time domain continuity; s3, establishing a hierarchical deep learning architecture which comprises a vertical coupling structure of a macroscopic period sensing layer, a mesoscopic industry conducting layer and a microscopic asset response layer; S4, constructing a dynamic feature interaction engine in the estimation framework, automatically identifying the relative weights of key driving factors under different market states by utilizing a multi-head self-attention mechanism, and generating a feature combination path evolving along with a market structure through topology reconstruction; S5, model training is implemented based on the generated countermeasure network frame, the generator network continuously generates a synthesized data sequence conforming to market dynamics, and the discriminator network combines real-time market feedback to identify the distribution difference of the synthesized data and the real market path; s6, outputting a joint estimation vector formed by a basic value trend item, a market emotion deviation item and a fluidity correction item; and S7, synchronously running at least two independently trained estimated paths in an operation environment, performing confidence filtering on the divergence estimated result through a Monte Carlo game strategy, and finally outputting a dynamic estimated interval and a stability score which accord with a preset risk threshold. Preferably, the unstructured derivative data sources in S1 include: The social media public opinion emotion tendency index is analyzed through a natural language processing engine and comprises an investor emotion polarity quantized value and a topic transmissio