Asset valuation dynamic evaluation system based on large model

Abstract

The invention relates to the technical field of asset evaluation, in particular to a large-model-based asset valuation dynamic evaluation system, which comprises a data fusion and preprocessing module, a feature mining module, a probabilistic valuation core module, a dynamic assignment and decision module and a closed-loop self-optimization module. According to the scheme, policy texts and public opinion data are subjected to deep semantic analysis through a natural language processing technology, macroeconomic expectation, market emotion and other recessive features are extracted, and dynamic mapping of second-level market information data and monthly macroscopic indexes is achieved in combination with a space-time alignment algorithm. By means of the multi-dimensional data fusion capability, the system can capture value driving factors which are difficult to recognize through a traditional method. According to the scheme, the dimension number of asset value description is greatly increased, multi-level influence factors such as macroeconomy, industry dynamics and market emotion are effectively covered, and a more comprehensive and more stereoscopic input feature set is provided for a subsequent valuation model.

Description

Technical Field

[0001] The present invention relates to the technical field of asset evaluation, and in particular to a dynamic asset valuation evaluation system based on a large model. Background Art

[0002] In today's digital age, data has become a vital asset for businesses and organizations, encompassing user information, transaction records, market analysis, and more, playing an irreplaceable role in strategic decision-making. The rapid development of cutting-edge technologies such as big data, cloud computing, and artificial intelligence has further highlighted the value of data assets, making them a key component of a company's core competitiveness. Simultaneously, demand for asset valuation continues to grow across various industries. Whether in traditional financial investment or the emerging digital asset sector, accurate asset valuation is a crucial foundation for decision-making.

[0003] However, current mainstream asset valuation methods suffer from numerous limitations. Traditional approaches often rely on manual analysis and expert scoring, which is not only labor-intensive and time-consuming, resulting in extremely low efficiency, but also highly subjective, making it difficult to ensure the accuracy and objectivity of the valuation results. Traditional methods are particularly inadequate when faced with massive, complex, and rapidly changing data. They are unable to process and analyze it promptly and accurately, making it difficult to capture dynamic changes in asset value. Furthermore, these methods often lack effective utilization of unstructured data, resulting in an incomplete assessment dimension and an inability to deeply explore the potential value of assets. Therefore, we propose a dynamic asset valuation system based on large models. Summary of the Invention

[0004] In view of the deficiencies of the existing technology, the present invention provides a dynamic asset valuation assessment system based on a large model, thereby solving the technical problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] The dynamic asset valuation system based on a large model includes:

[0007] The data fusion and preprocessing module is used to deeply integrate structured data and unstructured data to build a unified asset feature representation space. The structured data includes second-level market information and quarterly financial statements, and the unstructured data includes policy texts and social media public opinion.

[0008] The feature mining module builds an interpretable framework for asset value-driven analysis through causal reasoning and risk transmission modeling.

[0009] The core module of probabilistic valuation, based on the conditional diffusion model, builds a probability distribution generation mechanism for asset value and realizes dynamic model calibration through real-time deviation detection;

[0010] The dynamic scoring and decision-making module converts probabilistic valuation results into actionable decision signals and generates intelligent hedging strategies based on the risk transmission network;

[0011] The closed-loop self-optimization module enables the system to adapt to the non-stationary characteristics of the financial market through continuous learning and memory management.

[0012] In a possible implementation, the data fusion and preprocessing module includes:

[0013] The unstructured text parsing unit uses FinBERT, a pre-trained model in the financial field, to implement semantic parsing of unstructured text and output semantic vectors. Based on the BERT architecture, FinBERT is domain-adapted and trained on a financial corpus to output a 768-dimensional semantic vector, where dimensions 1-128 correspond to macroeconomic factors, dimensions 129-384 represent industry-specific features, and dimensions 385-768 represent market sentiment. The unstructured text parsing unit uses FinBERT's classification head to implement probabilistic detection of key events. If the probability of an event is greater than a preset threshold, it is determined to be a strong semantic event, triggering an event priority processing link.

[0014] The spatiotemporal alignment unit adopts a composite alignment method combining dynamic time warping (DTW) and cubic spline interpolation to solve the problem of time dimension differences in multi-source data. The spatiotemporal alignment unit first aligns the time points of low-frequency data with the high-frequency data series through the DTW algorithm to address the time dimension differences between high-frequency market data and low-frequency macro data, determines the mapping position of macro indicators on the high-frequency time axis, and then uses cubic spline interpolation to fill in the missing points at the minute level. The spatiotemporal alignment unit introduces a double-layer volatility verification mechanism. When the ratio of the interpolated volatility to the original volatility is greater than the preset threshold, it is determined that the interpolation result has an overfitting risk and automatically switches to linear interpolation and re-verifies.

[0015] In a possible implementation, the feature mining module includes:

[0016] A causal graph construction unit identifies core influencing factors based on an acyclic causal model. The unit uses the NOTEARS algorithm to learn the causal structure between asset prices and market factors. To address the high noise characteristics of financial data, an L1 regularization term is introduced into the optimization objective function to enhance sparsity. After weight normalization, if the causal weight of a feature on the target asset is greater than a preset threshold, it is determined to be a core driving factor. If the causal weight of a feature on the target asset is less than a preset threshold, the feature is determined to be an auxiliary factor and stored in the system database for use in secondary analysis scenarios.

[0017] The risk contagion modeling unit quantifies the systemic risk transmission between assets through conditional covariance analysis. The risk contagion modeling unit characterizes the risk impact of asset pairs through risk contagion coefficients, and systematically constructs a dynamic risk transmission network by accumulating the risk contagion coefficients of all asset pairs. Different asset categories follow differentiated transmission logic. When calculating the risk contagion coefficient, if the conditional covariance analysis results show that the return correlation between assets contains common factors driven by public opinion, the direct risk transmission effect between assets is separated by controlling the text semantic influence weight.

[0018] In one possible implementation, the probabilistic valuation core module includes:

[0019] The conditional diffusion prediction unit, whose denoising network adopts a U-Net structure, takes as input a combination of noisy prices and multimodal conditional variables. The unit assesses tail risk by predicting the kurtosis of the distribution. When the kurtosis is greater than a preset threshold, it indicates that the distribution has a significant fat tail, and the system automatically triggers the high-frequency data enhancement mode. If the kurtosis is less than or equal to the preset threshold, the current data collection frequency and data type are maintained, and the high-frequency data enhancement mode is not triggered.

[0020] The value mutation detection unit uses JS divergence to quantify the difference between the predicted distribution and the actual price distribution. When the divergence is greater than the preset threshold, it is determined that a value mutation has occurred in the market, and the system performs causal graph reconstruction, diffusion model fine-tuning and manual collaborative triggering operations; when the value mutation detection unit performs causal graph reconstruction, if the reconstructed divergence is still higher than the preset threshold, a visual report containing mutation factor analysis will be pushed to the investment team. The report includes a mutation factor heat map, a model confidence waterfall chart and a historical similar event library.

[0021] In one possible implementation, the dynamic scoring and decision-making module includes:

[0022] The dynamic threshold scoring unit combines the predicted mean and historical price percentiles, using double-segment linear interpolation to balance the sensitivity of high and low ranges. When the market volatility is greater than the preset threshold, the scoring range is expanded in proportion to the volatility, and the final score is mapped to a five-level decision signal. When the market volatility is less than or equal to the preset threshold, the scoring range remains unchanged and is not expanded.

[0023] The hedging strategy generation unit constructs a hedge portfolio based on the risk contagion coefficient. The optimization goal is to minimize the risk value of the hedged portfolio. The historical simulation method is used to calculate the risk value of the hedged portfolio. If the risk value is less than the preset threshold, the strategy is automatically pushed to the trading system for execution; if the risk value is greater than or equal to the preset threshold, the weight distribution of the hedge portfolio is re-evaluated and the hedging strategy is adjusted; if the conditions are not met, the system will return the latest market data to the conditional diffusion prediction unit, regenerate the valuation distribution to adjust the hedging parameters.

[0024] In one possible implementation, the closed-loop self-optimization module includes:

[0025] The elastic weight consolidation unit prevents the model from forgetting historical knowledge when learning new data. When the divergence between the new data distribution and the original distribution is greater than a preset threshold, the mechanism is automatically activated to constrain the update range of key parameters to ensure that the model retains historical key features. When the divergence between the new data distribution and the original distribution is less than or equal to the preset threshold, the constraint mechanism is not activated, allowing the model parameters to be updated freely.

[0026] The memory replay unit stores samples with prediction errors exceeding 2 times the standard deviation in the memory library. When the library capacity exceeds the preset threshold, the distributed retraining process is triggered, including data sampling, model retraining and parameter synchronization. When storing samples in the memory library, if the library capacity does not exceed the preset threshold, the distributed retraining process is not triggered and sample data continues to be accumulated.

[0027] Beneficial effects compared with existing technologies:

[0028] 1. This solution uses natural language processing technology to conduct in-depth semantic analysis of policy texts and public opinion data, extracting implicit features such as macroeconomic expectations and market sentiment. Combined with a spatiotemporal alignment algorithm, it dynamically maps second-by-second market data with monthly macroeconomic indicators. This multi-dimensional data fusion capability enables the system to capture value drivers that are difficult to identify using traditional methods. This solution significantly increases the number of dimensions used to characterize asset value, effectively covering multiple influencing factors such as the macroeconomy, industry dynamics, and market sentiment, providing a more comprehensive and multi-dimensional input feature set for subsequent valuation models.

[0029] 2. In this solution, an interpretable asset value-driven framework is constructed through causal reasoning and risk transmission networks. The causal reasoning module eliminates false correlation factors, accurately locates core driving factors, and quantifies their direct impact on asset value. The risk transmission network constructs a dynamic risk contagion map based on historical correlations between assets and real-time market data, which can identify the sources of contagion and sensitive points of industry credit risk in advance. This interpretable driving framework not only enhances the credibility of valuation results, but also provides users with a specific path for risk management and control, achieving a breakthrough from single-asset factor analysis to multi-asset risk linkage;

[0030] 3. This solution utilizes a dynamic diffusion model and closed-loop self-optimization mechanism to achieve real-time evolution of the valuation system. The dynamic diffusion model integrates textual semantics and causal weights to generate a price distribution that incorporates policy expectations and cost drivers, enabling rapid response to emergencies. The closed-loop self-optimization module utilizes adversarial sample training and elastic weight consolidation techniques to continuously optimize model parameters. In practice, the system can quickly complete causal graph reconstruction and model fine-tuning for extreme scenarios such as financial report failures. The parameter update cycle is shortened from quarterly to real-time dynamic adjustment, forming a positive cycle of "data accumulation-model evolution-decision optimization," significantly improving the timeliness and risk resilience of asset valuations. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention with reference to the accompanying drawings.

[0032] Figure 1 This is a schematic diagram of the asset valuation dynamic assessment system framework of the present invention;

[0033] Figure 2 This is a flow chart of the asset valuation dynamic assessment system of the present invention. DETAILED DESCRIPTION

[0034] Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention can be implemented in various forms, and therefore the present invention is not limited to the embodiments described below. In addition, in order to more clearly describe the present invention, components that are not related to the present invention will be omitted from the drawings.

[0035] The technical solution in the embodiments of the present application is to solve the problems of the above-mentioned background technology, and the overall idea is as follows:

[0036] Example:

[0037] This embodiment introduces a dynamic asset valuation assessment system based on a large model. The system includes a data fusion and preprocessing module, a feature mining module, a probabilistic valuation module, a dynamic scoring and decision-making module, and a closed-loop self-optimization module.

[0038] 1. Data fusion and preprocessing module

[0039] This module builds a unified asset feature representation space by deeply integrating structured data (specifically, second-level market information and quarterly financial statements) with unstructured data (specifically, policy texts and social media sentiment). It includes an unstructured text parsing unit and a spatiotemporal alignment unit.

[0040] The system first converts unstructured text into computable semantic vectors through the unstructured text parsing unit, and then uses the spatiotemporal alignment unit to solve the problem of time dimension differences in multi-source data, providing standard data for subsequent causal reasoning and valuation modeling.

[0041] 1. Unstructured text parsing unit

[0042] This unit uses the financial field pre-training model FinBERT to achieve semantic parsing of unstructured text. Based on the BERT architecture, FinBERT performs domain adaptation training on the financial field corpus (including 1 billion financial words) and outputs a 768-dimensional semantic vector v t , the specific steps are as follows:

[0043] Receives raw text data, specifically, news headlines such as "Central Bank announces 50 basis point interest rate cut" and policy document paragraphs such as "Ministry of Finance issues management measures for local government special bonds." It supports formats such as TXT and PDF, and automatically identifies text source types, including policy and public opinion.

[0044] The WordPiece word segmentation algorithm is used to segment the text into a sequence of tokens, adding [CLS] (classification marker) and [SEP] (separation marker). Taking the input text "CSRC intends to relax refinancing restrictions on listed companies" as an example, the word segmentation result is ["[CLS]", "CSRC", "intend", "relax", "listed company", "refinancing", "restriction", "[SEP]"].

[0045] The token sequence is input into the FinBERT model and the semantic vector v is generated by the 12-layer Transformer encoder. t ∈R 768 Each dimension of the vector corresponds to a semantic feature in the financial sector. Specifically, dimensions 1-128 correspond to macroeconomic factors (such as the semantic representation of GDP growth and interest rate policy), dimensions 129-384 are industry-specific features (such as the "carbon quota price" dimension in the new energy industry), and dimensions 385-768 represent market sentiment (including a quantitative indicator of the "fear index" on social media). Specifically, when inputting the news "A pharmaceutical company's innovative drug has been approved," the weight of the "patent protection period" dimension can be increased from 0.2 to 0.6, directly triggering a long-term positive adjustment to the valuation model.

[0046] The system pre-defined 12 key events (such as policy adjustments, major mergers and acquisitions, and credit risks) are probabilistically detected through FinBERT’s classification head. Specifically, the model outputs a probability value between [0,1] for each event category, and FinBERT outputs the event probability distribution P(e j |v t ), if the probability of an event P(e j)>0.85, it is determined to be a strong semantic event. The formal expression is:

[0047] Text features v t Trigger events are prioritized and pushed directly to the event impact analysis subunit, forming a parallel processing architecture with the regular analysis of the causal graph construction unit. This design enables the system to respond immediately to sudden financial events. For example, when an event with a strong semantic meaning of "M&A" is detected, the valuation process of related assets will prioritize the impact of the M&A expectations, rather than waiting for the regular data fusion process.

[0048] For text that does not trigger a strong semantic event, its semantic vector v t The cross-modal processing of structured data (such as asset prices, trading volumes, and price-to-earnings ratios) is performed through the feature splicing-attention fusion mechanism. Specifically, the numerical features are first mapped to a 768-dimensional space through a linear transformation and then compared with v t After splicing, input two layers of fully connected network and output fusion feature vector v fusion This process automatically assigns weights to textual and numerical features through an attention mechanism. Specifically, when analyzing technology stocks, the weight of textual features of social media sentiment may be higher than that of financial indicators, while when evaluating bonds, the weight of macroeconomic data prevails.

[0049] 2. Spatiotemporal alignment unit

[0050] This unit aims to solve the time dimension difference between high-frequency market data (seconds) and low-frequency macro data (months). The system adopts a composite alignment method of dynamic time warping (DTW) combined with cubic spline interpolation. First, the time points of low-frequency data are aligned with the high-frequency data sequence through the DTW algorithm to determine the mapping position of macro indicators on the high-frequency time axis. Then, cubic spline interpolation is used to fill in the missing points at the minute level. The interpolation function is: x(t) = a(tt k ) 3 +b(tt k ) 2 +c(tt k )+d,t k ≤t<t k+1 , where t k ,t k+1 The timestamps of adjacent data points are represented by a, b, c, and d, which are coefficients obtained by solving the function values ​​of the four adjacent data points and the continuity condition of the first-order derivative. This ensures the continuity of the first-order and second-order derivatives of the interpolation curve, thereby maintaining a smooth transition between the macro trend and high-frequency fluctuations.

[0051] To avoid interpolation distortion, the system introduces a two-layer volatility verification mechanism:

[0052] Original volatility σ raw calculate, Among them, T i is the return rate on day i, The 30-day average return is based on the original daily data and measures the asset volatility through the standard deviation of the return over the past 30 days;

[0053] Interpolated volatility σ interp Calculate σ interp Calculation method and σ raw The standard deviation is calculated based on interpolated minute-level data within the same time window.

[0054] When σ interp >2σ raw When the interpolation result is judged to be at risk of overfitting, the threshold is set based on the statistical characteristics of financial data. That is, the volatility magnification of normal interpolation usually does not exceed 2 times. If it exceeds, the linear interpolation switch mechanism is activated. Specifically, when PMI data deviates from the stock price trend, linear interpolation can preserve the authenticity of high-frequency data and avoid overfitting the macro trend. Automatically switch to linear interpolation The data is then recalibrated until the volatility ratio meets the threshold. This mechanism ensures that the aligned data reflects macro trends without distorting the true volatility characteristics of high-frequency trading.

[0055] 2. Feature Mining Module

[0056] This module builds an interpretable framework for asset value-driven risk analysis through causal reasoning and risk transmission modeling. It includes a causal graph construction unit and a risk contagion modeling unit. The causal graph construction unit identifies core influencing factors based on an acyclic causal model, while the risk contagion modeling unit quantifies systemic risk transmission between assets through conditional covariance analysis. Together, they form a comprehensive feature mining system, extending from single-asset factor analysis to multi-asset risk networks.

[0057] 1. Causal diagram construction unit

[0058] The system uses the NOTEARS algorithm to learn the causal structure of asset prices and market factors. In view of the high noise characteristics of financial data, the L1 regularization term is introduced in the optimization objective function to enhance sparsity: Here, X′ is the normalized feature matrix, W is the causal adjacency matrix, and the constraint function h(W) = 0 ensures that the causal graph is acyclic through matrix trace operations. Specifically, when analyzing a bank stock, the algorithm can identify the direct causal effects of factors such as "net interest margin changes" and "macroeconomic prosperity index" on the stock price, eliminating any spurious correlations between "stock market trading volume" and the stock price.

[0059] Weight normalization, normalize W by row, Make the sum of the weights of each row equal to 1. After the causal weight is normalized, if the causal weight of a feature j on the target asset y is These factors are directly injected into the input layer of the conditional diffusion prediction unit, forming joint conditional variables with the text semantic features. Specifically, when the causal weight of the "10-year Treasury bond yield" is 0.45, its dynamic changes will directly affect the valuation model of bond assets, while auxiliary factors (such as turnover rate) are stored in the system database for secondary analysis scenarios such as outlier detection.

[0060] 2. Risk Contagion Modeling Unit

[0061] This unit quantifies the risk transmission effect between assets through the risk contagion coefficient B i,j The formula to describe the risk impact of asset i on j is: Among them, the conditional covariance Cov(r i ,r j |v t ) Eliminate the indirect effect of text semantics on yield through regression model, r i ,r j is the rate of return of asset i,j, is the return variance of asset j, calculated in the same way as σ raw Same, v t is the text feature vector. Specifically, when the market has a consensus opinion of "favorable industry policies" (v t The weight of the “favorable policy” dimension is high), the correlation of returns between assets may include common factors driven by public opinion. By controlling v t The direct risk transmission effect between assets can be separated.

[0062] By accumulating the risk contagion coefficients of all asset pairs, the system constructs a dynamic risk transmission network. Different asset classes follow differentiated transmission logics: equity assets focus on "industry chain linkage + sentiment resonance" (public opinion factor weighting 40%), fixed income assets focus on "credit rating linkage" (text semantic indirect influence weighting 20%), and the derivatives market focuses on "volatility transmission" (strongly correlated with Delta and Vega). The out-degree of the node ∑ j B i,j Reflects the asset's ability to export risk to the market. i B i,j This reflects its sensitivity to external risks. Specifically, in a credit event in a certain industry in 2023, the system identified the top three leading enterprises in the out-degree ranking as risk contagion sources and the top five small and medium-sized enterprises in the in-degree ranking as risk-sensitive points through the risk network, thus issuing an industry-wide risk warning in advance.

[0063] 3. Probabilistic Valuation Core Module

[0064] This module uses the conditional diffusion model as its core to construct a probability distribution generation mechanism for asset value. It also dynamically calibrates the model through real-time deviation detection, forming a closed-loop valuation process of "prediction-verification-correction." It includes a conditional diffusion prediction unit and a value mutation detection unit.

[0065] 1. Conditional Diffusion Prediction Unit

[0066] The denoising network of the diffusion model adopts the U-Net structure, whose input contains the noise price x t and multimodal conditional variables c = CONCAT(v t ,W y ). Among them, v t Carrying textual semantic information such as policies and public opinions, W y Contains the core causal weight of asset value, and the two are integrated into the denoising process through channel splicing and attention mechanism. Specifically, when predicting a new energy stock, v t The semantic vector of “new energy subsidy policy adjustment” in W y The causal weights of “lithium price fluctuations” in the model jointly guide the model to generate a price distribution that includes policy expectations and cost drivers.

[0067] The model predicts the kurtosis of the distribution Assess tail risk. When k > 5, indicating a significant fat tail in the distribution (e.g., extreme ups and downs), the system automatically triggers high-frequency data enhancement mode: data collection frequency increases from minutes to seconds, while order book depth data is also collected to capture changes in market microstructure. This mechanism provides early warning, enabling the system's valuation model to adjust volatility parameters in a timely manner.

[0068] 2. Value mutation detection unit

[0069] Jensen-Shannon divergence (JS divergence) is used to quantify the difference between the predicted distribution P and the actual price distribution Q: When JS>0.2, it is determined that a sudden change in market value has occurred (such as an unexpected financial report crash), and the system performs the following operations:

[0070] Causal graph reconstruction: Call the NOTEARS algorithm to relearn the causal structure of the latest 1 hour of data and identify sudden factors (such as abnormal key indicators in financial reports);

[0071] Diffusion model fine-tuning: injecting newly identified causal factors into conditional variables, fine-tuning the diffusion model in real time, and updating the valuation distribution for the next 24 hours;

[0072] Manual collaborative triggering: If the JS divergence is still higher than the threshold after reconstruction, a visual report containing mutation factor analysis will be pushed to the investment team. The report includes a mutation factor heat map (Sankey diagram showing the transmission path), a model confidence waterfall chart (comparing predicted changes at different time scales) and a historical similar event library (matching cases with similarity >80%) to assist manual decision-making.

[0073] 4. Probabilistic Valuation Core Module

[0074] This module converts probabilistic valuation results into actionable decision signals, adapts to market changes through a dynamic scoring mechanism, and generates intelligent hedging strategies based on a risk transmission network, achieving full automation from valuation to trading. It includes a dynamic threshold scoring unit and a hedging strategy generation unit.

[0075] 1. Dynamic threshold scoring unit

[0076] Basic scoring formula combined with predicted mean μ P Compared with the historical price percentile, double-segment linear interpolation is used to balance the sensitivity of the high and low ranges: Among them, V med is the median.

[0077] When the market volatility σ>20%, the scoring range is expanded in proportion to the volatility: Specifically, when σ=30%, the upper limit of the score is increased from 100 to 110 to avoid over-conservatism in asset ratings under high volatility. The final score is mapped to a five-level decision signal: Buy, it is recommended to increase holdings; when Hold, it is recommended to hold; when Wait and see, it is recommended to follow closely; when Reduce holdings, it is recommended to reduce holdings; when Sell, it is recommended to clear the inventory.

[0078] 2. Hedge strategy generation unit

[0079] For assets with a rating of ≤ Wait and See, the system constructs a hedge portfolio based on the risk contagion coefficient, with the optimization goal being: Among them, the covariance matrix ∑ is obtained through the risk contagion coefficient B i,j and asset volatility σ i Calculation, ∑ i,j =B i,j σ i σ j , β is the risk contagion coefficient vector of the target asset, and w is the hedge portfolio weight vector. This model is solved using the Lagrange multiplier method, ensuring that the portfolio's risk characteristics are highly negatively correlated with the target asset.

[0080] The historical simulation method is used to calculate the risk value of the hedged portfolio VaR: 0.95 =Quantile(-ΔP, 0.95), where ΔP is the portfolio's profit and loss sequence over the past 250 trading days. If VaR is less than 5%, the strategy is automatically pushed to the trading system for execution. If the conditions are not met, the system feeds the latest 5-minute market data back to the conditional diffusion prediction unit, regenerating the valuation distribution to adjust the hedging parameters.

[0081] 5. Closed-loop self-optimization module

[0082] This module enables the system to adapt to the non-stationary nature of the financial market through continuous learning and memory management, forming a positive cycle of "data accumulation-model evolution-decision optimization." It includes an elastic weight consolidation unit and a memory playback unit.

[0083] 1. Elastic weight consolidation unit

[0084] To prevent the model from forgetting historical knowledge when learning new data, the loss function is designed as: Among them, L new (θ) is the current data loss, is a snapshot of historical parameters, F i Fisher information is a measure of the importance of parameters to historical tasks. When the KL divergence D between the new data distribution and the original distribution KL When it is greater than 0.1, specifically, the market switches from a bullish to a bearish regime, and the EWC mechanism is automatically activated, by constraining the update range of key parameters. Ensure that the model retains key characteristics of historical bull and bear cycles.

[0085] 2.Memory playback unit

[0086] The system stores samples (x, y) with prediction errors exceeding 2 times the standard deviation or JS>0.2 in memory bank D mem , the sample contains metadata such as feature vector, actual price and predicted distribution. 5 When the number of items is equal to 0, the distributed retraining process is triggered:

[0087] Data sampling: From D mem 500,000 abnormal samples were randomly selected and mixed with regular samples at a ratio of 1:4;

[0088] Model retraining: We used the PyTorch distributed framework to train on multiple GPU nodes, reducing the learning rate to 0.1 times the initial value. We focused on optimizing the tail prediction capability of the diffusion model and the ability of the causal graph to identify sudden factors.

[0089] Parameter synchronization: The retrained model parameters are synchronized to each module through the message queue to ensure consistent updates across the entire system.

[0090] Finally, it should be noted that the above embodiments are merely examples for the purpose of illustrating the present invention and are not intended to limit the embodiments. Those skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. It is not necessary and impossible to provide an exhaustive list of all embodiments. However, obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A dynamic asset valuation system based on a large model, characterized by: include: The data fusion and preprocessing module is used to deeply integrate structured data and unstructured data to build a unified asset feature representation space. The structured data includes second-level market information and quarterly financial statements, and the unstructured data includes policy texts and social media public opinion. The feature mining module builds an interpretable framework for asset value-driven analysis through causal reasoning and risk transmission modeling. The core module of probabilistic valuation, based on the conditional diffusion model, builds a probability distribution generation mechanism for asset value and realizes dynamic model calibration through real-time deviation detection; The dynamic scoring and decision-making module converts probabilistic valuation results into actionable decision signals and generates intelligent hedging strategies based on the risk transmission network; The closed-loop self-optimization module enables the system to adapt to the non-stationary characteristics of the financial market through continuous learning and memory management.

2. The asset valuation dynamic evaluation system based on a large model according to claim 1, characterized in that: The data fusion and preprocessing module includes: The unstructured text parsing unit uses FinBERT, a pre-trained model in the financial field, to implement semantic parsing of unstructured text and output semantic vectors. Based on the BERT architecture, FinBERT is domain-adapted and trained on a financial corpus to output a 768-dimensional semantic vector, where dimensions 1-128 correspond to macroeconomic factors, dimensions 129-384 represent industry-specific features, and dimensions 385-768 represent market sentiment. The unstructured text parsing unit uses FinBERT's classification head to implement probabilistic detection of key events. If the probability of an event is greater than a preset threshold, it is determined to be a strong semantic event, triggering an event priority processing link. The spatiotemporal alignment unit adopts a composite alignment method combining dynamic time warping (DTW) and cubic spline interpolation to solve the problem of time dimension differences in multi-source data. The spatiotemporal alignment unit first aligns the time points of low-frequency data with the high-frequency data series through the DTW algorithm to address the time dimension differences between high-frequency market data and low-frequency macro data, determines the mapping position of macro indicators on the high-frequency time axis, and then uses cubic spline interpolation to fill in the missing points at the minute level. The spatiotemporal alignment unit introduces a double-layer volatility verification mechanism. When the ratio of the interpolated volatility to the original volatility is greater than the preset threshold, it is determined that the interpolation result has an overfitting risk and automatically switches to linear interpolation and re-verifies.

3. The asset valuation dynamic evaluation system based on a large model according to claim 1, characterized in that: The feature mining module includes: A causal graph construction unit identifies core influencing factors based on an acyclic causal model. The unit uses the NOTEARS algorithm to learn the causal structure between asset prices and market factors. To address the high noise characteristics of financial data, an L1 regularization term is introduced into the optimization objective function to enhance sparsity. After weight normalization, if the causal weight of a feature on the target asset is greater than a preset threshold, it is determined to be a core driving factor. If the causal weight of a feature on the target asset is less than a preset threshold, the feature is determined to be an auxiliary factor and stored in the system database for use in secondary analysis scenarios. The risk contagion modeling unit quantifies the systemic risk transmission between assets through conditional covariance analysis. The risk contagion modeling unit characterizes the risk impact of asset pairs through risk contagion coefficients, and systematically constructs a dynamic risk transmission network by accumulating the risk contagion coefficients of all asset pairs. Different asset categories follow differentiated transmission logic. When calculating the risk contagion coefficient, if the conditional covariance analysis results show that the return correlation between assets contains common factors driven by public opinion, the direct risk transmission effect between assets is separated by controlling the text semantic influence weight.

4. The asset valuation dynamic evaluation system based on a large model according to claim 1, characterized in that: The probabilistic valuation core module includes: The conditional diffusion prediction unit, whose denoising network adopts a U-Net structure, takes as input a combination of noisy prices and multimodal conditional variables. The unit assesses tail risk by predicting the kurtosis of the distribution. When the kurtosis is greater than a preset threshold, it indicates that the distribution has a significant fat tail, and the system automatically triggers the high-frequency data enhancement mode. If the kurtosis is less than or equal to the preset threshold, the current data collection frequency and data type are maintained, and the high-frequency data enhancement mode is not triggered. The value mutation detection unit uses JS divergence to quantify the difference between the predicted distribution and the actual price distribution. When the divergence is greater than the preset threshold, it is determined that a value mutation has occurred in the market, and the system performs causal graph reconstruction, diffusion model fine-tuning and manual collaborative triggering operations; when the value mutation detection unit performs causal graph reconstruction, if the reconstructed divergence is still higher than the preset threshold, a visual report containing mutation factor analysis will be pushed to the investment team. The report includes a mutation factor heat map, a model confidence waterfall chart and a historical similar event library.

5. The large model-based dynamic asset valuation system according to claim 1, characterized in that: The dynamic scoring and decision-making module includes: The dynamic threshold scoring unit combines the predicted mean and historical price percentiles, using double-segment linear interpolation to balance the sensitivity of high and low ranges. When the market volatility is greater than the preset threshold, the scoring range is expanded in proportion to the volatility, and the final score is mapped to a five-level decision signal. When the market volatility is less than or equal to the preset threshold, the scoring range remains unchanged and is not expanded. The hedging strategy generation unit constructs a hedge portfolio based on the risk contagion coefficient. The optimization goal is to minimize the risk value of the hedged portfolio. The historical simulation method is used to calculate the risk value of the hedged portfolio. If the risk value is less than the preset threshold, the strategy is automatically pushed to the trading system for execution; if the risk value is greater than or equal to the preset threshold, the weight distribution of the hedge portfolio is re-evaluated and the hedging strategy is adjusted; if the conditions are not met, the system will return the latest market data to the conditional diffusion prediction unit, regenerate the valuation distribution to adjust the hedging parameters.

6. The large model-based dynamic asset valuation system according to claim 1, characterized in that: The closed-loop self-optimization module includes: The elastic weight consolidation unit prevents the model from forgetting historical knowledge when learning new data. When the divergence between the new data distribution and the original distribution is greater than a preset threshold, the mechanism is automatically activated to constrain the update range of key parameters to ensure that the model retains historical key features. When the divergence between the new data distribution and the original distribution is less than or equal to the preset threshold, the constraint mechanism is not activated, allowing the model parameters to be updated freely. The memory replay unit stores samples with prediction errors exceeding 2 times the standard deviation in the memory library. When the library capacity exceeds the preset threshold, the distributed retraining process is triggered, including data sampling, model retraining and parameter synchronization. When storing samples in the memory library, if the library capacity does not exceed the preset threshold, the distributed retraining process is not triggered and sample data continues to be accumulated.

Images