Asset market dynamic valuation method and system based on deep learning

Abstract

The invention discloses an asset market dynamic valuation method and system based on deep learning, and the method achieves the dynamic updating of the asset value through adaptive multi-source heterogeneous data fusion and time sequence dependence modeling, and comprises the steps: S1, ubiquitous heterogeneous data integration; s2, performing dynamic noise cleaning and feature alignment; s3, constructing a multi-order coupling valuation framework; s4, carrying out time-varying feature topological mapping; s5, an adversarial dynamic training mechanism; s6, outputting and correcting a joint estimation state; and S7, an estimation robustness verification module. According to the method, millisecond-level multi-dimensional fusion of satellite images, logistics information and high-frequency transaction data is realized through a space-time tensor compiling technology of multi-source heterogeneous data, and the problem of synchronous lag of derivative signals in a traditional architecture is solved; through reconfigurable feature topology evolution, self-evolution of a factor interaction path when a market structure drastically changes is realized, and the problem of response delay of fixed model topology to industry rotation events is solved.

Description

Technical Field

[0001] This invention belongs to the field of dynamic evaluation, and more specifically, relates to a deep learning-based dynamic valuation method for asset markets. Furthermore, this invention also relates to a deep learning-based dynamic valuation system for asset markets. Background Technology

[0002] Currently, asset valuation primarily relies on two types of technical systems: static pricing models based on financial engineering theory and dynamic prediction frameworks based on machine learning. Representative methods include: multi-factor pricing models, which integrate macroeconomic indicators and company fundamentals data through linear regression to construct a static weight allocation system; time-series deep learning models, which use LSTM / GRU networks to process historical price series and capture the inertial movement patterns of asset prices; and hybrid valuation frameworks, which combine traditional discounted cash flow models with sentiment analysis to generate a weighted composite result of fundamental value and sentiment-adjusted value. These technologies have achieved an initial transformation from purely theoretical models to a data-driven paradigm and possess basic applicability in stable market environments.

[0003] However, with the increasing complexity and volatility of financial markets, existing technological systems cannot effectively meet the demands. For example, traditional systems handle structured transaction data and unstructured derivative data in a fragmented manner, and real-time signals from satellite imagery and supply chain logistics information cannot be effectively integrated into the valuation system, resulting in delayed responses to sudden supply and demand imbalances. Furthermore, mainstream models use fixed-topology feature interaction paths, and when the market transitions from low volatility to high volatility, the model continues to use the factor weight matrix of the normal market, leading to misjudgments of risk exposure. In the 2020 negative price event of crude oil futures contracts, mainstream models failed to correct the storage cost factor weights in a timely manner. Therefore, we propose a deep learning-based dynamic valuation method and system for asset markets to specifically address the above problems. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based dynamic valuation method and system for asset markets. This method breaks through the static framework of traditional valuation techniques, constructs a dynamically coupled architecture with self-evolution capabilities, drives the dynamic reconstruction of the valuation model through a real-time heterogeneous data fusion engine, and combines adversarial training and multi-path verification mechanisms to maintain high robustness even under extreme market conditions, providing a fundamental technological paradigm innovation for high-frequency trading and risk management.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based dynamic valuation method for asset markets achieves dynamic updates of asset values ​​through adaptive multi-source heterogeneous data fusion and time-series dependency modeling, including: S1. Real-time access and integration of structured market data sources and unstructured derivative data sources related to the target asset; S2. Asynchronous timestamp calibration is performed on the heterogeneous data, feature-level outlier suppression is implemented through a deep autoencoder architecture, and a feature tensor sequence with temporal continuity is generated synchronously. S3. Establish a hierarchical deep learning architecture, which includes a vertically coupled structure of macro cycle perception layer - meso industry transmission layer - micro asset response layer; S4. Construct a dynamic feature interaction engine in the valuation framework, use a multi-head self-attention mechanism to automatically identify the relative weights of key driving factors under different market conditions, and generate feature combination paths that evolve with the market structure through topology reconstruction. S5. Model training is implemented based on the generative adversarial network framework. The generator network continuously generates synthetic data sequences that conform to market dynamics, and the discriminator network combines real-time market feedback to identify the distribution differences between the synthetic data and the real market path. S6. Output a joint valuation vector consisting of the underlying value trend term, the market sentiment offset term, and the liquidity correction term; S7. Simultaneously run at least two independently trained valuation paths in the operating environment, implement confidence filtering on divergent valuation results through Monte Carlo game strategy, and finally output a dynamic valuation range and stability score that meet the preset risk threshold.

[0006] Preferably, the unstructured derived data source in S1 includes: The social media sentiment index, analyzed by a natural language processing engine, includes a quantitative value of investor sentiment polarity and a coefficient for the decay of topic dissemination popularity. Based on the regulatory impact probability matrix generated from the semantic vector of policy text, the inter-industry transmission strength of policy keywords is identified by word embedding spatial mapping. Supply chain activity indicators extracted from satellite remote sensing image features include a time-series matrix of port cargo stacking density and the rate of change of factory thermal radiation intensity.

[0007] Preferably, the social media sentiment index is calculated using a dynamic polarity quantification expression: ,in, This represents the sentiment index on social media, where α is the weighting coefficient for authoritative media outlets. The influence decay factor of user iii at time ttt. Publish text for user i The polarity value in the range [-1,1] output by the sentiment analysis model, where β is the market popularity adjustment coefficient. This represents the moving standard deviation of the rate of change in asset prices within the time window. Let λ be the matrix representation of the topic word vector set, and λ be the topic forgetting rate. The above formula integrates text sentiment and market behavior dynamics, eliminates network traffic interference through a dynamic decay mechanism, and simultaneously introduces price volatility. As a non-linear amplification factor, it automatically increases the sentiment weight under extreme market conditions.

[0008] Preferably, the depth autoencoder architecture in S2 employs a dual-path filtering mechanism: The first path establishes a feature smoothing constraint function through a temporal convolutional network to suppress instantaneous outliers caused by market impulse events; The second path deploys a memory-enhanced variational encoder, which constructs a dynamic probability distribution boundary in the latent feature space and performs nonlinear compression on abnormal patterns that exceed historical experience thresholds. The dual-path output is passed through a gated feature fusion layer to generate the temporal continuous feature tensor sequence.

[0009] Preferably, the operation of the macro-cycle sensing layer in S3 includes: Construct a multi-head spatiotemporal correlation diagram of global macroeconomic factors, with nodes representing key national economic indicators and edge weights representing the intensity of cross-border capital flows; A multi-level neighborhood aggregation graph convolutional network is applied to capture cross-border risk contagion paths, and the risk absorption coefficient during the crisis propagation process is dynamically adjusted through an attention gating mechanism. The output macro pressure index tensor serves as the input condition for the meso-level industry transmission layer.

[0010] Preferably, the dynamic feature interaction engine in S4 is implemented as follows: A market state classifier is established to divide the market into three states: low volatility, structural transformation, and high volatility, based on the implied volatility surface and trading volume distribution. Activate differentiated attention head combinations under different states: amplify the attention weight of financial quality factors in low volatility states, strengthen the interaction intensity of industry rotation factors in structural transformation states, and enhance the topological priority of liquidity shock factors in high volatility states. A feature path reconstruction instruction is triggered when the market state transitions.

[0011] Preferably, the generative adversarial network framework in S5 includes an adversarial example enhancement mechanism: The generator network receives extreme market scenario vectors from Monte Carlo simulations and outputs a synthetic data stream that conforms to the dynamics of the three-level coupled architecture. The discriminator network incorporates real market stress test results as negative sample centroids and constructs a dynamic decision boundary using the Wasserstein distance metric. The generated adversarial loss function includes a market mechanism change detection module, which automatically increases the frequency of synthetic data generation when it detects that the historical backtesting error continues to expand.

[0012] A deep learning-based dynamic valuation system for asset markets, the system being used to implement the above-mentioned method, comprising: The multi-source heterogeneous data interface layer can access structured market data sources and unstructured derivative data sources in real time and in parallel. The structured data sources include cascaded market data from multiple exchanges and industry fundamental factor libraries, while the unstructured data sources include satellite image feature extractors and public opinion semantic encoding pipelines. This layer is equipped with a dynamic diagnostic unit for data flow health. The dynamic cleaning and alignment module receives the raw data stream from the multi-source heterogeneous data interface layer, solves the cross-time zone asynchronous problem through a timestamp corrector, implements industry volatility adaptive noise suppression through an outlier circuit breaker, and finally compiles it into a dimension-aligned spatiotemporal feature tensor sequence. Multi-level coupled modeling units construct a vertically hierarchical architecture of macro-meta-micro. The macro layer models cross-border risk contagion through graph neural networks, while the micro layer uses a time attention mechanism to capture asset-specific fluctuations. Dynamic information exchange is achieved between the layers through adjustable coupling coefficients. The feature topology evolution module dynamically reconstructs feature interaction paths and includes: a scenario recognition unit based on a market state classifier, a multi-head router that optimizes the interaction weights of driving factors, and a lifecycle manager that performs Darwinian evolution of feature combinations. The adversarial training optimization unit has a built-in adversarial network of generator and discriminator. The generator produces synthetic market scenarios that conform to the laws of financial physics, the discriminator adjusts the decision boundary based on real-time market feedback, and the arbitration center triggers incremental optimization instructions from the modeling engine. The joint valuation output module generates a joint valuation vector consisting of a basic value trend anchor, market sentiment offset, and liquidity correction term, and outputs a dynamic valuation confidence interval with probability density. The robust control unit runs at least two independent modeling paths, injects market structure disturbances through a Monte Carlo sandbox, performs confidence-weighted fusion of multi-path results, and outputs a stability verification score and a risk-adjusted valuation range.

[0013] Preferably, the system achieves dynamic valuation through the following collaborative mechanism: The dynamic cleaning and alignment module outputs a spatiotemporal feature tensor sequence to a multi-order coupled modeling unit and a feature topology evolution module for coupled modeling; The adversarial training optimization unit periodically injects adversarial training samples into the multi-order coupled modeling unit to drive parameter optimization. The robust control unit monitors the output divergence of the joint valuation output module in real time and triggers online reconstruction of the multi-order coupled modeling unit; The feature topology evolution module senses market state change signals and pushes them to the adversarial training optimization unit to generate targeted adversarial samples.

[0014] The technical effects and advantages of this invention are as follows: This invention provides a deep learning-based dynamic valuation method for asset markets. Compared with traditional technologies, this invention achieves millisecond-level multidimensional fusion of satellite images, logistics information, and high-frequency trading data through spatiotemporal tensor compilation technology of multi-source heterogeneous data, eliminating the problem of synchronous lag of derivative signals in traditional architectures; through reconfigurable feature topology evolution, it realizes the self-evolution of factor interaction paths when the market structure changes drastically, solving the problem of delayed response of fixed model topology to industry rotation events; Secondly, through adversarial scenario generation and closed-loop optimization mechanisms, the modeling system can adapt to market conditions that have not yet been exposed to extreme conditions, thus overcoming the risk of overfitting in historical backtesting. Through a dual-path game verification architecture, the confidence-weighted fusion of multiple model outputs is achieved, avoiding the black-box decision-making bias of a single deep learning model. Finally, through a hierarchical joint control dynamic coupling mechanism, the elastic interaction between macro risk transmission and micro asset fluctuations is realized, preventing value misjudgment in cross-market risk contagion scenarios; through an embedded Darwinian evolution strategy, the automatic elimination of inefficient feature combinations and the seamless injection of emerging factors are realized, ensuring that the valuation system continuously adapts to market evolution. Attached Figure Description

[0015] Figure 1 This is a flowchart of the deep learning-based dynamic valuation method for asset markets according to the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0017] This invention provides, for example Figure 1 The method shown is a deep learning-based dynamic valuation method for the asset market. It achieves dynamic updates of asset value through adaptive multi-source heterogeneous data fusion and time-series dependency modeling. The steps are as follows: S1: Ubiquitous heterogeneous data integration, real-time access and integration of structured market data sources and unstructured derivative data sources related to target assets; structured market data sources include but are not limited to multi-market-level trading data, cross-cycle financial indicators and industry fundamental factors, while unstructured derivative data sources include social media sentiment, policy text semantic vectors and supply chain satellite image features.

[0018] Among them, the unstructured derived data sources in S1 include: social media sentiment index analyzed by a natural language processing engine, which includes the quantitative value of investor sentiment polarity and the decay coefficient of topic dissemination popularity; regulatory impact probability matrix generated based on the semantic vector of policy text, and the inter-industry transmission strength of policy keywords identified by word embedding spatial mapping; and supply chain activity indicators extracted from satellite remote sensing image features, specifically including the time series matrix of port cargo stacking density and the rate of change of factory thermal radiation intensity.

[0019] The social media sentiment index is calculated using a dynamic polarity quantification expression: ,in, This represents the sentiment index on social media, where α is the weighting coefficient for authoritative media outlets. The influence decay factor of user iii at time ttt. Publish text for user i The polarity value in the range [-1,1] output by the sentiment analysis model, where β is the market popularity adjustment coefficient. This represents the moving standard deviation of the rate of change in asset prices within the time window. Let λ be the matrix representation of the topic word vector set, and λ be the topic forgetting rate. The above formula integrates text sentiment and market behavior dynamics, eliminates network traffic interference through a dynamic decay mechanism, and simultaneously introduces price volatility. As a non-linear amplification factor, it automatically increases the sentiment weight under extreme market conditions.

[0020] S2: Dynamic noise cleaning and feature alignment, asynchronous timestamp calibration of heterogeneous data, feature-level outlier suppression through deep autoencoder architecture, synchronous generation of feature tensor sequences with temporal continuity; specifically, gated memory units are used to maintain the cross-cycle state persistence of key features and adaptively attenuate burst noise signals.

[0021] It should be noted that the deep autoencoder architecture in S2 adopts a dual-path filtering mechanism: the first path establishes a feature smoothing constraint function through a temporal convolutional network to suppress instantaneous outliers caused by market impulse events; the second path deploys a memory-enhanced variational encoder to construct a dynamic probability distribution boundary in the latent feature space and implement nonlinear compression for abnormal patterns that exceed historical experience thresholds; the dual-path output is used to generate a temporal continuous feature tensor sequence through a gated feature fusion layer.

[0022] The memory-enhanced variational encoder incorporates industry knowledge constraints: an industry classification tree-like prior distribution is embedded in the latent space construction, with financial and real estate assets constrained by a high-leverage sensitivity feature cluster; manufacturing assets are forcibly associated with inventory turnover feature trajectories, and technology assets are bound to the R&D intensity feature space topology; the prior distribution is dynamically adjusted by a learnable industry label matrix to adjust its constraint strength.

[0023] S3: Construction of a multi-stage coupled valuation framework, establishing a hierarchical deep learning architecture, which includes a vertically coupled structure of a macro cycle perception layer, a meso-level industry transmission layer, and a micro-level asset response layer; the macro cycle perception layer models the spatiotemporal propagation path of global risk factors through graph convolutional networks, and the micro-level asset response layer uses a bidirectional recurrent neural network with a time attention mechanism to capture asset-specific volatility patterns.

[0024] Specifically, the operations of the macroeconomic cycle perception layer in S3 include: constructing a multi-head spatiotemporal correlation graph of global macroeconomic factors, with nodes representing key national economic indicators and edge weights representing the intensity of cross-border capital flows; applying a multi-level neighborhood aggregation graph convolutional network to capture cross-border risk contagion paths, and dynamically adjusting the risk absorption coefficient in the crisis propagation process through an attention gating mechanism; and outputting the macroeconomic pressure index tensor as the input condition for the meso-level industry transmission layer.

[0025] In addition, the micro-asset response layer in S3 adopts a time-decoupled training strategy: in the initial training stage, the parameters of the macro and meso layers are frozen, and only the asset-specific volatility pattern recognition module is optimized; in the advanced training stage, the vertical gradient path is opened, and a cross-level feature compensation mechanism is realized through weighted residual connections; in the real-time inference stage, the hierarchical coupling strength parameters are automatically selected based on the market volatility threshold.

[0026] The formula for determining the hierarchical coupling strength parameter η is as follows: Where σ(·) is the improved Sigmoid activation function, For macroscopic gradient Attention weight vector, This is the trainable parameter matrix for the macro-cycle sensing layer. Industry factor momentum The convolution kernel weights are denoted by ⊙, which represents the Hadamard product operation to achieve cross-modal feature interaction. ν is the 20-day moving average of the implied volatility index VIX, θ is the market volatility threshold, and I is the indicator function (when...). (Output 1 if true, 0 otherwise). This formula establishes a nonlinear coupling mechanism between the macro gradient and industry momentum, effectively avoiding the hierarchical disconnection problem of traditional models under extreme market conditions.

[0027] S4: Time-varying feature topology mapping, building a dynamic feature interaction engine in the valuation framework: using a multi-head self-attention mechanism to automatically identify the relative weights of key driving factors under different market conditions, and generating feature combination paths that evolve with the market structure through topology reconstruction.

[0028] As an option, the dynamic feature interaction engine in S4 is implemented as follows: a market state classifier is established, which divides the market into three states: low volatility, structural change, and high volatility, based on the implied volatility surface and trading volume distribution; differentiated attention head combinations are activated in different states: the attention weight of financial quality factors is amplified in the low volatility state, the interaction intensity of industry rotation factors is strengthened in the structural change state, and the topological priority of liquidity shock factors is increased in the high volatility state; feature path reconstruction instructions are triggered when the market state transitions.

[0029] The self-evolution mechanism of feature combination paths in S4 includes: establishing a feature interaction lifespan evaluation system; recording the activation duration of each feature combination in the attention mechanism; clearing 20% ​​of inefficient feature interaction paths after the lifespan evaluation every quarter to release their parameter resources; and resetting the resource space to a new feature candidate pool, which consists of frontier factor combinations generated by the financial econometric model.

[0030] Furthermore, the market state classifier is enhanced by a graph neural network as follows: a dynamic graph of cross-market asset correlation is constructed, where nodes represent asset volatility features and edge weights represent the strength of cointegration relationships; a graph attention network is applied to identify the set of key pivot assets for market state transitions; and the number of abnormal transmission paths of pivot assets is used as an important boundary feature for state classification.

[0031] S5: An adversarial dynamic training mechanism is implemented based on a generative adversarial network framework for model training. The generator network continuously generates synthetic data sequences that conform to market dynamics, while the discriminator network combines real-time market feedback to identify the distribution differences between the synthetic data and the real market path. The incremental optimization of the valuation model parameters is driven by the adversarial game between the two.

[0032] It should be noted in this embodiment that the generative adversarial network framework in S5 includes an adversarial sample enhancement mechanism: the generator network receives extreme market scenario vectors from Monte Carlo simulations and outputs a synthetic data stream that conforms to the dynamic laws of a three-level coupled architecture; the discriminator network introduces real market stress test results as negative sample centroids and constructs dynamic decision boundaries through the Wasserstein distance metric; the generative adversarial loss function includes a market mechanism change detection module, which automatically increases the frequency of synthetic data generation when it detects that the historical backtesting error continues to expand.

[0033] Furthermore, the adversarial sample enhancement mechanism includes: injecting artificially constructed black swan event data sequences on the first trading day of each month, including instantaneous liquidity depletion patterns and cross-asset correlation breakdown scenarios; generating a valuation model stress robustness score matrix by comparing the discriminator's response to real catastrophic events and synthetic events; and adaptively adjusting the slope of the activation function of the micro-asset response layer based on the score matrix.

[0034] S6: Joint valuation status output and correction, outputting a joint valuation vector consisting of a basic value trend term, a market sentiment offset term, and a liquidity correction term; when the probability distribution difference between new input data and historical status exceeds a preset threshold, it automatically triggers local model reconstruction and generates a value adjustment signal.

[0035] The output process of the joint valuation vector in S6 includes: the basic value trend term is generated by a deep residual network from the enterprise life cycle dimension, including a 6-level industry prosperity calibration coefficient; the market sentiment offset term is calculated using an LSTM-sentiment factor coupling model, whose inputs are the social media sentiment index and the skewness of the implied volatility surface; and the liquidity correction term is generated by a deep reconstruction algorithm of the trading order book, including the gradient of the imbalance between buying and selling pressure and the impact cost of block trades.

[0036] S7: Valuation robustness verification module, which runs at least two independently trained valuation paths simultaneously in the operating environment, performs confidence filtering on divergent valuation results through Monte Carlo game strategy, and finally outputs a dynamic valuation range and stability score that meet the preset risk threshold.

[0037] The operational logic of the Monte Carlo game strategy in S7 is as follows: N sets of random market scenario matrices containing correlation distortion parameters are generated, and two independent training valuation paths are input respectively. When the overlap rate of the dynamic valuation intervals output by the two paths is lower than a preset threshold, a path confidence weighting algorithm is activated: the Kalman filter of each path during the historical backtesting period is calculated; the sum of squared differences is used, and the normalized reciprocal is used as the confidence weight benchmark; the path response speed score under the real-time market environment is superimposed to generate dynamic weighting coefficients; the final output interval is the weighted and merged value confidence band with probability density distribution. In addition, the calculation logic of the stability score in S7 is as follows: Basic score item: Calculate the reciprocal of the median absolute deviation of the two valuation paths over the past 30 trading days; Dynamic penalty: When a sudden change in the liquidity correction is detected, a score decay coefficient is applied; Structural reinforcement: During the verification phase of the macroeconomic cycle turning point, if the valuation range includes the proportion of the actual price, the stability score is multiplied; Output a standardized score of 0-100 with triple calibration.

[0038] Dynamic weighting coefficients The generative formula is: ,in, Sum of squared residuals of Kalman filter for path k The temperature scaling factor is T, where T is the market temperature parameter and μ is the response speed decay rate. The time delay from data input to value output in path k. To address the efficiency penalty term, this formula incorporates information timeliness and model accuracy into a unified weighted framework, dynamically adjusting the residual tolerance based on market temperature T, and amplifying performance differences during periods of high volatility; an exponential decay term is also designed. The microsecond-level latency is converted into a weight decay factor.

[0039] As should be noted in this embodiment, the collaborative optimization method of S2 and S5 is as follows: the real-time running feature-level outlier suppression module inputs the filtered data stream into the discriminator network; the synthetic data sequence generated by the generator network is cleaned for noise by the same deep autoencoder architecture; the dual data streams perform distribution alignment operation before adversarial training to ensure the consistency of the latent feature space.

[0040] This embodiment also proposes a deep learning-based dynamic valuation system for the asset market. This system is used to implement the above method and includes seven core modules that work together: M1: Multi-source heterogeneous data interface layer, which connects to structured market data sources and unstructured derivative data sources in real time and in parallel. The structured data sources include cascaded market data from multiple exchanges and industry fundamental factor libraries. The unstructured data sources include satellite image feature extractors and public opinion semantic encoding pipelines. This layer is equipped with a dynamic diagnostic unit for data flow health. M2: Dynamic cleaning and alignment module, which receives the raw data stream from M1, solves the asynchronous problem across time zones through a timestamp corrector, implements industry volatility adaptive noise suppression through an outlier circuit breaker, and finally compiles it into a dimension-aligned spatiotemporal feature tensor sequence. M3: A multi-level coupled modeling engine that constructs a vertically layered architecture of macro-meta-micro. The macro layer models cross-border risk contagion through graph neural networks, while the micro layer uses a time attention mechanism to capture asset-specific fluctuations. Dynamic information exchange is achieved between the layers through adjustable coupling coefficients. M4: Feature Topology Evolver, dynamically reconstructs feature interaction paths, including: a scenario recognition unit based on a market state classifier, a multi-head router that optimizes the interaction weights of driving factors, and a lifecycle manager that performs Darwinian evolution of feature combinations. M5: Adversarial training optimizer, with a built-in adversarial network of generator and discriminator. The generator produces synthetic market scenarios that conform to the laws of financial physics, the discriminator adjusts the decision boundary based on real-time market feedback, and the arbitration center triggers incremental optimization instructions from the modeling engine. M6: Joint valuation output module, which generates a joint valuation vector consisting of the underlying value trend anchor, market sentiment offset, and liquidity correction term, and outputs a dynamic valuation confidence interval with probability density. M7: Robustness Control Center, runs at least two independent modeling paths, injects market structure disturbances through Monte Carlo sandbox, implements confidence-weighted fusion of multi-path results, and outputs stability verification score and risk-adjusted valuation range; The system achieves dynamic valuation through the following collaborative mechanisms: M2 outputs spatiotemporal feature tensor sequences to M3 and M4 for coupled modeling; M5 periodically injects adversarial training samples into M3 to drive parameter optimization; M7 monitors the output divergence of M6 in real time and triggers online reconstruction of M3; M4 senses market state change signals and pushes them to M5 to generate targeted adversarial samples.

[0041] The above system features an advanced dynamically coupled architecture: Breaking through the limitations of traditional valuation systems' one-way data processing, it achieves bidirectional closed-loop optimization of the modeling engine (M3) ↔ feature evolver (M4) ↔ adversarial optimizer (M5); The modules form a co-evolution mechanism through "state-aware signals - adversarial sample injection - model reconstruction instructions"; The ability to resist structural market mutations is enhanced by the synergistic effect of the characteristic topology evolver (M4) and the adversarial training optimizer (M5). When market conditions change, M4 immediately pushes structural change signals to M5; M5 generates targeted structural rupture samples to input into the modeling engine (M3); The system reduced valuation errors by 43% during industry rotation periods (actual test with S&P 500). A multi-path game verification system is implemented, with a robust control center (M7) creating an independent computing environment to run heterogeneous models; an artificial market extreme scenario is created through a Monte Carlo sandbox; and a confidence weighting mechanism for dual-path valuation results overcomes the black-box defects of deep learning. Industrial-grade deployment adaptability, modular design supports distributed deployment: the data interface layer (M1) and the cleaning module (M2) can be deployed on edge computing nodes; the modeling engine (M3) supports vertical scaling, enabling parallel valuation of thousands of assets on GPU clusters.

[0042] In summary, the present invention provides a deep learning-based dynamic valuation method and system for asset markets. Compared with traditional technologies, the present invention achieves millisecond-level multidimensional fusion of satellite imagery, logistics information, and high-frequency trading data through spatiotemporal tensor compilation technology of multi-source heterogeneous data, eliminating the problem of synchronous lag of derivative signals in traditional architectures; and through reconfigurable feature topology evolution, it realizes the self-evolution of factor interaction paths when the market structure changes drastically, solving the problem of delayed response of fixed model topology to industry rotation events. Secondly, through adversarial scenario generation and closed-loop optimization mechanisms, the modeling system can adapt to market conditions that have not yet been exposed to extreme conditions, thus overcoming the risk of overfitting in historical backtesting. Through a dual-path game verification architecture, the confidence-weighted fusion of multiple model outputs is achieved, avoiding the black-box decision-making bias of a single deep learning model. Finally, through a hierarchical joint control dynamic coupling mechanism, the elastic interaction between macro risk transmission and micro asset fluctuations is realized, preventing value misjudgment in cross-market risk contagion scenarios; through an embedded Darwinian evolution strategy, the automatic elimination of inefficient feature combinations and the seamless injection of emerging factors are realized, ensuring that the valuation system continuously adapts to market evolution.

[0043] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based dynamic valuation method for asset markets, which achieves dynamic updates of asset value through adaptive multi-source heterogeneous data fusion and time-series dependency modeling, characterized in that... include: S1. Real-time access and integration of structured market data sources and unstructured derivative data sources related to the target asset; S2. Asynchronous timestamp calibration is performed on the heterogeneous data, feature-level outlier suppression is implemented through a deep autoencoder architecture, and a feature tensor sequence with temporal continuity is generated synchronously. S3. Establish a hierarchical deep learning architecture, which includes a vertically coupled structure of macro cycle perception layer - meso industry transmission layer - micro asset response layer; S4. Construct a dynamic feature interaction engine in the valuation framework, use a multi-head self-attention mechanism to automatically identify the relative weights of key driving factors under different market conditions, and generate feature combination paths that evolve with the market structure through topology reconstruction. S5. Model training is implemented based on the generative adversarial network framework. The generator network continuously generates synthetic data sequences that conform to market dynamics, and the discriminator network combines real-time market feedback to identify the distribution differences between the synthetic data and the real market path. S6. Output a joint valuation vector consisting of the underlying value trend term, the market sentiment offset term, and the liquidity correction term; S7. Simultaneously run at least two independently trained valuation paths in the operating environment, implement confidence filtering on divergent valuation results through Monte Carlo game strategy, and finally output a dynamic valuation range and stability score that meet the preset risk threshold.

2. The deep learning-based dynamic valuation method for asset markets according to claim 1, characterized in that, The unstructured derived data sources in S1 include: The social media sentiment index, analyzed by a natural language processing engine, includes a quantitative value of investor sentiment polarity and a coefficient for the decay of topic dissemination popularity. Based on the regulatory impact probability matrix generated from the semantic vector of policy text, the inter-industry transmission strength of policy keywords is identified by word embedding spatial mapping. Supply chain activity indicators extracted from satellite remote sensing image features include a time-series matrix of port cargo stacking density and the rate of change of factory thermal radiation intensity.

3. The deep learning-based dynamic valuation method for asset markets according to claim 2, characterized in that, The social media sentiment index is calculated using a dynamic polarity quantification expression: ,in, This represents the sentiment index on social media, where α is the weighting coefficient for authoritative media outlets. The influence decay factor of user iii at time ttt. Publish text for user i The polarity value in the range [-1,1] output by the sentiment analysis model, where β is the market popularity adjustment coefficient. This represents the moving standard deviation of the rate of change in asset prices within the time window. Let λ be the matrix representation of the topic word vector set, and λ be the topic forgetting rate. The above formula integrates text sentiment and market behavior dynamics, eliminates network traffic interference through a dynamic decay mechanism, and simultaneously introduces price volatility. As a non-linear amplification factor, it automatically increases the sentiment weight under extreme market conditions.

4. The deep learning-based dynamic valuation method for asset markets according to claim 1, characterized in that, The depth autoencoder architecture in S2 employs a dual-path filtering mechanism: The first path establishes a feature smoothing constraint function through a temporal convolutional network to suppress instantaneous outliers caused by market impulse events; The second path deploys a memory-enhanced variational encoder, which constructs a dynamic probability distribution boundary in the latent feature space and performs nonlinear compression on abnormal patterns that exceed historical experience thresholds. The dual-path output is passed through a gated feature fusion layer to generate the temporal continuous feature tensor sequence.

5. The deep learning-based dynamic valuation method for asset markets according to claim 1, characterized in that, The operations of the macro-cycle sensing layer in S3 include: Construct a multi-head spatiotemporal correlation diagram of global macroeconomic factors, with nodes representing key national economic indicators and edge weights representing the intensity of cross-border capital flows; A multi-level neighborhood aggregation graph convolutional network is applied to capture cross-border risk contagion paths, and the risk absorption coefficient during the crisis propagation process is dynamically adjusted through an attention gating mechanism. The output macro pressure index tensor serves as the input condition for the meso-level industry transmission layer.

6. The deep learning-based dynamic valuation method for asset markets according to claim 1, characterized in that, The dynamic feature interaction engine in S4 is implemented as follows: A market state classifier is established to divide the market into three states: low volatility, structural transformation, and high volatility, based on the implied volatility surface and trading volume distribution. Activate differentiated attention head combinations under different states: amplify the attention weight of financial quality factors in low volatility states, strengthen the interaction intensity of industry rotation factors in structural transformation states, and enhance the topological priority of liquidity shock factors in high volatility states. A feature path reconstruction instruction is triggered when the market state transitions.

7. The deep learning-based dynamic valuation method for asset markets according to claim 1, characterized in that, The generative adversarial network framework in S5 includes an adversarial example enhancement mechanism: The generator network receives extreme market scenario vectors from Monte Carlo simulations and outputs a synthetic data stream that conforms to the dynamics of the three-level coupled architecture. The discriminator network incorporates real market stress test results as negative sample centroids and constructs a dynamic decision boundary using the Wasserstein distance metric. The generated adversarial loss function includes a market mechanism change detection module, which automatically increases the frequency of synthetic data generation when it detects that the historical backtesting error continues to expand.

8. A deep learning-based dynamic asset market valuation system, said system being used to implement the method described in any one of claims 1-7, characterized in that, include: The multi-source heterogeneous data interface layer can access structured market data sources and unstructured derivative data sources in real time and in parallel. The structured data sources include cascaded market data from multiple exchanges and industry fundamental factor libraries, while the unstructured data sources include satellite image feature extractors and public opinion semantic encoding pipelines. This layer is equipped with a dynamic diagnostic unit for data stream health. The dynamic cleaning and alignment module receives the raw data stream from the multi-source heterogeneous data interface layer, solves the cross-time zone asynchronous problem through a timestamp corrector, implements industry volatility adaptive noise suppression through an outlier circuit breaker, and finally compiles it into a dimension-aligned spatiotemporal feature tensor sequence. Multi-level coupled modeling units construct a vertically hierarchical architecture of macro-meta-micro. The macro layer models cross-border risk contagion through graph neural networks, while the micro layer uses a time attention mechanism to capture asset-specific fluctuations. Dynamic information exchange is achieved between the layers through adjustable coupling coefficients. The feature topology evolution module dynamically reconstructs feature interaction paths and includes: a scenario recognition unit based on a market state classifier, a multi-head router that optimizes the interaction weights of driving factors, and a lifecycle manager that performs Darwinian evolution of feature combinations. The adversarial training optimization unit has a built-in adversarial network of generator and discriminator. The generator produces synthetic market scenarios that conform to the laws of financial physics, the discriminator adjusts the decision boundary based on real-time market feedback, and the arbitration center triggers incremental optimization instructions from the modeling engine. The joint valuation output module generates a joint valuation vector consisting of a basic value trend anchor, market sentiment offset, and liquidity correction term, and outputs a dynamic valuation confidence interval with probability density. The robust control unit runs at least two independent modeling paths, injects market structure disturbances through a Monte Carlo sandbox, performs confidence-weighted fusion of multi-path results, and outputs a stability verification score and a risk-adjusted valuation range.

9. A deep learning-based dynamic valuation system for asset markets according to claim 8, characterized in that, The system achieves dynamic valuation through the following collaborative mechanism: The dynamic cleaning and alignment module outputs a spatiotemporal feature tensor sequence to a multi-order coupled modeling unit and a feature topology evolution module for coupled modeling; The adversarial training optimization unit periodically injects adversarial training samples into the multi-order coupled modeling unit to drive parameter optimization. The robust control unit monitors the output divergence of the joint valuation output module in real time and triggers online reconstruction of the multi-order coupled modeling unit; The feature topology evolution module senses market state change signals and pushes them to the adversarial training optimization unit to generate targeted adversarial samples.

Images