Enterprise digital asset comprehensive management method and system

Abstract

The invention discloses an enterprise digital asset comprehensive management method and system, and relates to the technical field of computers, risk indexes and correlation factors are aligned in a unified time in a report period, multivariate time series data are formed and input, subsequent analysis is carried out on the same caliber and the same time axis, and the risk indexes and the correlation factors are analyzed in real time. Interpretation deviation and repeated arrangement cost caused by cross-system caliber difference are reduced; time sequence causal discovery and lagging influence expression are included in the same calculation framework, a causal structure containing a driving factor and an influence path is output, risk changes can be attributed and organized in a path mode, and therefore discrete events and index fluctuation are included in an interpretable structured link; in addition, resampling stable selection is introduced to form a stable edge set, so that the generation of the causal graph is converted from single solution to structure screening with consistent frequency, the probability that an accidental edge enters an interpretation link is reduced, and the stability and consistency of cross-batch and cross-scene analysis results are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for comprehensive management of enterprise digital assets. Background Technology

[0002] In scenarios involving financial regulatory reporting and internal risk management, existing enterprise digital asset (such as the digital representation of assets, liabilities, and related risk exposures) integrated management systems are typically capable of collecting and integrating multi-source data, including transaction data, valuation data, market data, credit lines and limits, and risk measurement results. These data are then used to create risk dashboards / indicators through visualization analysis components to support risk monitoring and report output. In existing technologies, some solutions can summarize or periodically update key risk metrics according to preset criteria, such as Value at Risk (VaR), stress test / scenario analysis results, and various exposures and limit occupancy, thereby providing a structured data foundation for regulatory reporting.

[0003] Regulatory reporting and information disclosure not only require the presentation of quantitative indicators, but also require explanations of significant changes in risk exposure, as well as explanations of management's responses and impacts. To meet these requirements, existing solutions generally combine risk measurement models with charts to present indicator trends, distribution characteristics, and some statistical correlations.

[0004] However, in generating narrative content such as management analysis and conclusions in regulatory reports, existing solutions often rely on fixed templates or rule-based interpretations of single indicators. This makes it difficult to establish a traceable attribution chain for risk changes and driving factors (such as changes in market factors, portfolio structure adjustments, parameter / model changes, business events, and external shocks) across systems and scopes, and to form a consistent and verifiable narrative. Therefore, in the actual compilation process, risk management personnel still need to manually interpret multi-source data, connect events, and organize text based on their experience. When multiple scenarios and rapid iterative analyses of the same asset portfolio are required, and repeated reporting is necessary, problems such as inconsistent analytical frameworks, insufficient reusability, and long compilation cycles can easily occur. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a comprehensive management method and system for enterprise digital assets to address the problem that while the visualization of multi-source risk data is improved, regulatory attribution still relies on manual processes and is difficult to iterate.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a method for comprehensive management of enterprise digital assets, comprising:

[0009] Step S1: Upon receiving a request to generate a risk analysis narrative for the target asset portfolio, acquire and time-align the multivariate time series data of the target asset portfolio within a specified reporting period. The multivariate time series data includes at least a risk indicator sequence and multiple correlation factor sequences corresponding to the risk indicator sequence.

[0010] Step S2: Perform causal discovery processing on a computer based on the multivariate time series data to obtain a directed causal graph characterizing the correlation factor to the risk indicator, and determine at least one directed path from the directed causal graph as a key causal structure.

[0011] The nodes of the directed causal graph correspond to the variables of the risk indicator sequence and the correlation factor sequence, respectively.

[0012] Step S3: Extract the change characteristics of the risk indicator sequence during the specified reporting period, and encode the change characteristics and the key causal structure together as conditional input;

[0013] Step S4: Input the conditions into the trained risk narrative generation model to generate and output the risk analysis narrative text corresponding to the specified reporting period.

[0014] As a preferred embodiment of the enterprise digital asset comprehensive management method described in this invention, the time alignment includes: performing frequency unification, missing value processing and outlier processing on the risk indicator sequence and the correlation factor sequence, and attaching a source system identifier and a timestamp identifier to each data point.

[0015] As a preferred embodiment of the enterprise digital asset comprehensive management method described in this invention, the causal discovery process includes: constructing a continuous optimization problem containing sparsity constraints and acyclic constraints to learn an adjacency matrix representing the causal relationship of variables, and obtaining the directed causal graph by solving the problem;

[0016] The continuous optimization problem represents variables as combinations of current values ​​and several lagged values ​​when expanding lagged variables, so as to allow directed edges across periods to express lagged effects, and applies acyclic constraints only to the causal structure corresponding to the same time slice, so that the directed causal graph output in the same time slice satisfies the requirement of directed acyclic interpretation.

[0017] As a preferred embodiment of the enterprise digital asset comprehensive management method described in this invention, a prohibition constraint is introduced into the continuous optimization problem. The prohibition constraint is generated by a knowledge graph in the financial field and is used to prohibit the occurrence of causal edges that are defined as not allowed by the knowledge graph in the directed causal graph.

[0018] As a preferred embodiment of the enterprise digital asset comprehensive management method described in this invention, determining the key causal structure includes: extracting directed paths terminating at risk indicator nodes from the directed causal graph, calculating the causal strength score of each directed path, and selecting a predetermined number of directed paths with the highest causal strength scores as the key causal structure.

[0019] As a preferred embodiment of the enterprise digital asset comprehensive management method of the present invention, the change features include: using a change point detection algorithm to identify significant change points of the risk indicator sequence within the specified reporting period, and outputting the statistical comparison results of the subsequences before and after each change point.

[0020] As a preferred embodiment of the enterprise digital asset comprehensive management method described in this invention, the risk narrative generation model is an autoregressive language model based on the Transformer architecture. The model is trained under supervision by pairing historical risk report paragraphs with corresponding standardized risk data templates and key causal structures to form training samples, so that the generated text is constrained by the conditional input.

[0021] After solving the continuous optimization problem, the causal edges are resampled for stability selection. The frequency of occurrence of causal edges obtained by multiple resampling is statistically analyzed and a set of stable edges is selected. The set of stable edges is used to generate a directed causal graph or as the input basis for subsequent extraction of key causal structures.

[0022] As a preferred embodiment of the enterprise digital asset comprehensive management method described in this invention, it further includes a tracing step: associating and storing the assertion in the risk analysis narrative text with the path segments in the key causal structure supporting the assertion and the corresponding source data time window identifiers to form a queryable evidence chain.

[0023] In a second aspect, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the method described in the first aspect of the present invention.

[0024] Thirdly, the present invention provides a comprehensive enterprise digital asset management system, comprising,

[0025] The data interface module is used to acquire and time-align the multivariate time series data of the target asset portfolio within a specified reporting period;

[0026] The causal analysis module is used to perform causal discovery processing based on the multivariate time series data to generate a directed causal graph and determine key causal structures;

[0027] The feature extraction module is used to extract the variation features of the risk indicator sequence;

[0028] The narrative generation module is used to encode the key causal structure and the change features as conditional inputs and input them into the risk narrative generation model to output risk analysis narrative text.

[0029] And a source tracing and storage module, used to establish a chain of evidence links between assertions in the risk analysis narrative text, path segments, and source data time window identifiers.

[0030] Through the above technical solution, the present invention can achieve at least the following beneficial effects:

[0031] To address the issues of regulatory reports relying on manual reporting and inconsistent attribution explanations, this invention aligns risk indicators and related factors at a unified time during the reporting period, forming multivariate time-series data inputs. This allows subsequent analyses to be conducted on the same caliber and time axis, reducing interpretive biases and redundant processing costs caused by differences in cross-system calibers.

[0032] In response to the problem that the causes of risk changes are complex and traditional correlation analysis is insufficient to support management conclusions, this invention incorporates time-series causal discovery and lagged impact expression into the same computational framework, outputting a causal structure that includes driving factors and impact paths. This allows risk changes to be attributed and organized in a path-based manner, thereby incorporating discrete events and indicator fluctuations into an interpretable structured chain.

[0033] To address the issue that causal structures are sensitive to sample perturbations, leading to instability in the analysis framework during multi-scenario iterations, this invention introduces resampling stability selection to form a stable edge set. This transforms the generation of causal graphs from a single solution to a structure selection with consistent frequency, reducing the probability of random edges entering the interpretation path and improving the stability and consistency of cross-batch and cross-scenario analysis results.

[0034] To address the issue of narrative generation easily going off-topic or resulting in uncontrolled expressions, this invention encodes key causal structures and change features as unified conditional inputs and adopts a narrative generation method based on template fields and path information constraints. This ensures that the topic scope, referenced objects, and chapter structure of the generated text are constrained by the conditional inputs, reducing the expression drift caused by free generation.

[0035] To address the issue of high requirements for factual consistency and verifiability in regulatory texts, this invention performs consistency verification on the backfill results of indicator names, directional words, and placeholders at the output end, and triggers the regeneration of controlled sentence structures or manual review markers when inconsistencies occur, so that the text output has a controllable error correction and degradation mechanism.

[0036] To address the problem of management analysis and conclusions lacking supporting evidence chains and being difficult to verify point by point, this invention associates and stores assertions in the text with corresponding causal path segments and source data time window identifiers, forming a queryable evidence chain. This allows each conclusion to be traced back to the data window and causal structure that supports it, improving auditing and review efficiency.

[0037] To address the issue of data version confusion during systematic implementation and multi-department collaboration, this invention manages the association between batch identifiers, model version identifiers, and training data version identifiers. This ensures that data acquisition, causal analysis output, and text output have a version alignment basis, facilitating a consistent data evidence chain during iterative updates, verification, and reproduction. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0039] Figure 1 This is a flowchart of the enterprise digital asset integrated management method in the embodiment.

[0040] Figure 2 This is a framework diagram of the enterprise digital asset integrated management system in the embodiment. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0042] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way. Example 1:

[0043] like Figure 1 As shown, this embodiment proposes a comprehensive management method for enterprise digital assets, including the following steps:

[0044] Step S1: Upon receiving a request to generate a risk analysis narrative for the target asset portfolio, acquire and time-align the multivariate time series data of the target asset portfolio within the specified reporting period. The multivariate time series data includes at least a risk indicator sequence and multiple related factor sequences corresponding to the risk indicator sequence.

[0045] The target asset portfolio is represented by an asset list data structure, which includes at least the asset identifier, asset type, currency, notional amount, market value, duration, credit rating, account or strategy identifier, and effective date. Changes to the target asset portfolio during the specified reporting period are represented by rebalancing records, which include at least the change time, the set of asset identifiers before and after the change, the weight change, and the reason for the change. When there are rebalancing records within the specified reporting period, the specified reporting period is divided into multiple sub-intervals according to the effective time of the rebalancing records, and a risk analysis narrative text is generated for each sub-interval, or the narrative texts of each sub-interval are merged in chronological order into a risk analysis narrative text for the specified reporting period.

[0046] The risk indicator sequence shall include at least one of the following: Value at Risk, Expected Loss, Stress Test Loss, Interest Rate Sensitivity, Credit Spread Sensitivity, Exchange Rate Sensitivity, Liquidity Gap, and Quota Utilization Ratio; the related factor sequence shall include at least one of the following: Changes in Key Maturity Points of the Yield Curve, Changes in the Credit Spread Index, Exchange Rate Changes, Changes in the Equity Index, Changes in the Volatility Index, Changes in Portfolio Duration, Credit Rating Migration Count, Changes in Collateral or Collateral Coverage Ratio, Model Version Identifier Sequence, and Business Event Identifier Sequence; Each data point in each sequence shall be accompanied by a source system identifier, indicator caliber identifier, currency and unit identifier, and data collection timestamp.

[0047] Step S2: Perform causal discovery processing on a computer based on multivariate time series data to obtain a directed causal graph representing the correlation factors to risk indicators, and determine at least one directed path from the directed causal graph as the key causal structure.

[0048] In this context, the nodes of the directed causal graph correspond to the variables of the risk indicator sequence and the associated factor sequence, respectively.

[0049] Time series causal modeling is achieved using a lagged variable expansion approach. For each associated factor and risk indicator, a lagged series of 1-5 days is constructed. The lagged series and the original series together constitute the input variable set for causal discovery processing. The lagged series of the same variable are only allowed to point to the variable or other variables at a more recent time. When the specified reporting period is less than 30 alignment periods, the lag order is fixed at 1 day and 2 days. When the specified reporting period is not less than 30 alignment periods, the lag order is fixed at 1-5 days.

[0050] Step S3: Extract the change characteristics of the risk indicator sequence during the specified reporting period, and encode the change characteristics and key causal structure together as conditional inputs;

[0051] The conditional input uses a parallel encoding method of structured text fragments and structured fields. The structured fields include the start and end dates of the reporting period, the name of the risk indicator, the unit of the risk indicator, the list of change points, the list of key causal paths, the name of the driving factor for each path, the sequence of mediating factors, the path direction marker, and the causal strength score. The structured text fragments are concatenated in a fixed order as context prefixes, which include: the reporting period overview section, the risk indicator change summary section, the change point summary section, and the key driving and path section. When a structured field is missing, the missing field is filled with "not provided", and the name of the missing field and the reason for the missing field are simultaneously output in the structured text fragment.

[0052] Step S4: Input the conditions into the trained risk narrative generation model to generate and output the risk analysis narrative text corresponding to the specified reporting period;

[0053] After the output text is generated, a consistency verification process is performed. The consistency verification includes: parsing the indicator names, direction words, and numerical placeholders in the output text; matching the parsed indicator names with the source calculation results; when the direction words are inconsistent with the direction of the source calculation results or the relative deviation of the numerical placeholder values ​​from the source calculation results exceeds 1%, the sentence is marked as failing and a regeneration process is triggered; the regeneration process uses a controlled template sentence structure and only allows the filling of indicator names, directions, and numerical placeholders from standardized risk data templates; if the verification still fails after regeneration, the corresponding sentence is marked as "requires manual review" and a verification failure list is appended to the end of the output text.

[0054] In this embodiment, the specified reporting period refers to a continuous time interval defined by the start and end dates of the reporting period. The start and end dates are carried in the generation request or read by the risk reporting task configuration. Time alignment maps the risk indicator sequence and related factor sequence to the same alignment period based on the timestamp. The alignment period and the end time of the day are written during the data acquisition phase and stored together with the data points. Multivariate time series data consists of a risk indicator sequence and multiple related factor sequences, and each data point includes a value, unit identifier, currency identifier, indicator caliber identifier, source system identifier, and collection timestamp. The indicator caliber identifier is used to distinguish the values ​​of the same indicator under different calculation calibers. Change characteristics refer to the quantitative description of the trend changes and abrupt changes of the risk indicator sequence within the specified reporting period. Key causal structure refers to the... The explanatory framework consists of driving factors, mediating factors, and directed paths terminating at risk indicators. Before generation, structured fields and structured text fragments in the conditional input undergo key deduplication and unit alignment, with unit alignment based on the unit identifier corresponding to the indicator's caliber. The relative deviation threshold in consistency verification is set to 1%, and as an implementation parameter, it can be configured within the range of 0.1% to 5%. The threshold value is jointly determined by the directional consistency rate and regeneration trigger rate on the validation set, and is kept non-zero to avoid misjudgments caused by calculation rounding and caliber conversion. Manual review markers are added to a verification failure list at the end of the text when verification fails. The verification failure list must at least include the indicator name, directional term, placeholder key, and corresponding source calculation result identifier, allowing generated outputs within the same reporting period to be located and reviewed item by item according to the list.

[0055] In this embodiment, time alignment includes: unifying the frequency of the risk indicator sequence and the associated factor sequence, handling missing values ​​and outliers, and attaching a source system identifier and a timestamp identifier to each data point;

[0056] The frequency is uniformly set to daily alignment, and the end-of-day time is set to a fixed time after the market closes on the same day, either Beijing time or Tokyo time. For high-frequency data, the closing value is taken at the end of the day or the cumulative sum is calculated daily. For low-frequency data, forward filling is performed based on the most recently published value. When the missing proportion of any sequence exceeds 20% within a specified reporting period, the sequence is removed from the input of the causal discovery process, and the reason for removal is recorded. Forward filling is used when there are no more than 3 consecutive missing periods. Linear interpolation is used when there are more than 3 but no more than 7 consecutive missing periods. When there are more than 7 consecutive missing periods, the missing interval is marked as an unusable interval and excluded from the corresponding time window in the subsequent path scoring calculation. Outlier detection adopts a median and absolute deviation method based on a rolling window. When a data point deviates from the median of the rolling window by more than five times the absolute deviation, it is judged as an outlier. The outlier is replaced with the 1% and 99th percentile cutoff values ​​of the rolling window, and the values ​​before and after replacement and the timestamp are recorded.

[0057] Specifically, the daily frequency alignment is based on the trading calendar, while the alignment period for non-trading days retains the timestamp and marks the period as a non-trading day. The end-of-day time is a fixed time after the market closes, defaulting to 18:00, but configurable within 16:00 to 20:00 as an implementation parameter. Alignment metadata for data points is written at fixed times to ensure cross-system consistency. The missing percentage threshold is set to 20%, configurable within 5% to 40% as an implementation parameter. The threshold value is determined by the quantiles of the missing data distribution in historical samples and selected on the validation set to maximize the stability of the directed causal graph. The boundaries between 3 and 7 consecutive missing alignment periods are used to distinguish between short and long missing data, and the missing interval is marked as unavailable. When calculating intervals, the data points corresponding to the interval are still retained but marked as unavailable and skipped in subsequent path scoring calculations; the default length of the rolling window is 20 alignment periods, and can be configured within 10 to 60 alignment periods as an implementation parameter; the five-fold absolute deviation threshold for outlier detection can be configured within 3 to 8 as an implementation parameter, and the threshold value is jointly determined by the outlier detection rate of historical samples and the smoothness of the backfilled sequence, and is kept not to be 1 to avoid excessive truncation of normal fluctuations; 1% and 99% quantile truncation are used to suppress the interference of tail extrema on causal discovery and change point detection, and the quantiles can be configured within 0.5% to 5% and 95% to 99.5% as implementation parameters, and the upper and lower quantiles are kept symmetrical to avoid introducing systematic bias.

[0058] In this embodiment, the causal discovery process includes: constructing a continuous optimization problem containing sparsity constraints and acyclic constraints to learn an adjacency matrix representing the causal relationship of variables, and obtaining a directed causal graph by solving the problem;

[0059] Before causal discovery, the input variables are standardized and a unified directional convention is established. This convention includes: risk indicator nodes are designated as the set of termination nodes, and correlation factor nodes as the set of candidate driving nodes; acyclic constraints apply to the set of causal edges within the same time slice, and causal edges across time slices are only allowed to point from an earlier time slice to a later time slice; the non-zero nature of adjacency matrix elements is achieved through sparsity constraints. After solving, the adjacency matrix is ​​sparsified according to an absolute value threshold of 0.05, with edge weights below the threshold set to zero; a directed acyclic consistency check is performed on the sparsified adjacency matrix, and when a directed cycle is found, edges within the cycle are removed sequentially from smallest to largest absolute value until no cycle exists.

[0060] When expanding lagged variables, continuous optimization problems represent variables as combinations of current values ​​and several lagged values ​​to allow directed edges across periods to express lagged effects. Acyclic constraints are applied only to the causal structure corresponding to the same time slice, so that the directed causal graph output in the same time slice satisfies the requirement of directed acyclic interpretation.

[0061] Continuous optimization problems can be solved by expanding on time-series variables, determining continuous differentiability and acyclicity, using sparse regularization, and link organization with augmented Lagrange solutions, enabling adjacency matrix learning and directed causal graph generation to be completed within the same optimization framework.

[0062] In one implementation, the steps for constructing a continuous optimization problem are as follows:

[0063] After time alignment of the multivariate time series, the observation vector at each time step is denoted as... Furthermore, the causal relationship is parameterized into a linear structure equation for the same time slice and the lag time slice, so that the edges of the same time slice are used to output the directed causal graph, and the intertemporal edges are used to absorb the lag effects of financial time series:

[0064] (1)

[0065] In equation (1), Represents a time index and satisfies ; express The vector of variables at each moment; Indicates the lag order; Indicates the lag order as The coefficient matrix; Represents the residual vector; Indicates the sample length. The adjacency matrix of the same time slice is defined as... And constrain its diagonal elements to zero to avoid self-loops.

[0066] Based on this, by incorporating the data fitting term and the sparsity constraint into the same objective function, we obtain an empirical risk form that can be used for continuous optimization:

[0067] (2)

[0068] In equation (2), This represents the fitting loss; Represents the parameter set and ; This represents the L2 norm.

[0069] To obtain a sparse causal structure, a sparsity regularization is superimposed on the fitting loss. If element-wise sparsity is used, it can be written as:

[0070] (3)

[0071] In equation (3), Represents sparse regularization terms; Denotes the norm 1; Indicates the sparse weights of slices within the same time period; Indicates the lag order as sparse weights.

[0072] If you need to express grouped sparsity of the same pair of variables over multiple lags, you can replace the regular expression with the grouped sparsity form:

[0073] (4)

[0074] In equation (4), This represents a group of sparse regularization terms; Indicates the sparse weights of the group; Indicates the number of variables; express The Line number Column elements; Representing an edge The coefficient vector at each lag order.

[0075] Acyclic constraints need to characterize directed acyclicity in the same time slice in a continuously differentiable manner, by applying the acyclic discriminant function to... Up, and not right The intertemporal matrix is ​​subjected to acyclicity.

[0076] Acyclicity determination can be performed using the acyclic function of matrix exponents:

[0077] (5)

[0078] In equation (5), This represents the acyclic discriminant function; Represents the trace function; Represents the matrix index; Represents the Hadamard product; Represents the adjacency matrix of the same time slice; Indicates the number of variables.

[0079] Using this as an equality constraint and combining it with the fitting term and sparse term, a continuous optimization problem is formed:

[0080] (6)

[0081] In equation (6), Represent the overall objective function; Describing sparse regular terms and ; satisfy .

[0082] To stably solve this equality constraint problem on a computer, the constraints can be transformed into an iteratively optimizeable form using the augmented Lagrangian method:

[0083] (7)

[0084] In equation (7), Represents the augmented Lagrange function; Represents the Lagrange multipliers; This represents the penalty coefficient.

[0085] right The update can be expressed as a single iteration using smooth partial gradient descent and sparse regularized proximal approach:

[0086] (8)

[0087] In equation (8), Indicates an iterative index; Indicates the first The parameter set for the next iteration; Indicates the step size; Indicates to Find the gradient; This represents the intermediate variables of the gradient step. Indicates the first The Lagrange multipliers of the next iteration, Indicates the first The penalty coefficient for the next iteration Indicates a set of parameters The same time slice matrix is ​​extracted and , The acyclic discriminant function is shown. If we use... If the element-wise sparsity is such that a soft threshold can be used to achieve near-end updates for each element:

[0088] (9)

[0089] In equation (9), Indicates the first parameter in the parameter set Updated values ​​for each element; Represents the corresponding element of the intermediate variable; Represents a symbolic function; This represents the sparse weight corresponding to the element, and its value comes from... or .

[0090] After completion After the update, consistent constraints are applied to the multipliers and penalty coefficients:

[0091] (10)

[0092] In equation (10), This represents the updated Lagrange multipliers; Indicates by The same time slice matrix is ​​extracted and .

[0093] The penalty coefficient can be increased using a threshold-driven doubling strategy:

[0094] (11)

[0095] In equation (11), Indicates the penalty multiplication factor; This represents the threshold for acyclic residuals.

[0096] The stopping condition is given by both the acyclic residual and the parameter change:

[0097] (12)

[0098] In equation (12), Denotes the Frobenius norm; This represents the threshold for relative change.

[0099] To enhance the reassertiveness of causal edges, a stable selection resampling can be superimposed on the outermost layer of the continuous optimization, transforming the edge set from a single-solution to a set of frequently occurring stable edges. For the ... The adjacency matrix is ​​obtained from the second resampled data. Then, calculate the edge selection probability:

[0100] (13)

[0101] In equation (13), Representing an edge The probability of choosing; Indicates the number of resampling attempts; Indicates an indicator function; Indicates the first Matrix elements obtained by secondary resampling; This represents the threshold for determining the edge coefficient. From this, we obtain the set of stable edges:

[0102] (14)

[0103] In equation (14), Represents the set of stable edges; This represents the stability threshold. This set can be used to output the final directed causal graph, or as the input basis for subsequent key causal structure path extraction.

[0104] After time alignment, based on the multivariate time series within the reporting period, the observation vector at each moment and several lagged observations are incorporated into the structural equation representation to form a time-series causal parameterization relationship with both the same time slice and inter-period effects, as shown in Equation (1). Subsequently, a data fitting term is constructed using the fitting error, as shown in Equation (2), and a sparsity constraint is superimposed on this fitting term to reduce unnecessary causal edges: element-wise sparse correspondence Equation (3), grouped sparse correspondence Equation (4), which is determined by the sparsity method used. For the same time slice structure, a continuously differentiable acyclic discriminant is introduced as an equality constraint, as shown in Equations (5) and (6), so that the output same time slice causal structure satisfies the directed acyclic interpretation requirement, while the inter-period directed edges are used to express the lagged effects. In the solution phase, the equality constraints are transformed into an iterative optimization objective through an augmented Lagrangian form, as shown in Equation (7). Each iteration first performs a gradient update according to Equation (8), and then performs a sparse near-end update according to Equation (9) to achieve sparsity. Subsequently, the multiplier update is advanced according to Equation (10), and the penalty coefficient is updated according to Equation (11). The threshold determination and doubling strategy of the penalty coefficient are pre-set in the method and system configuration phase. The iteration stops based on the joint condition of the acyclic residual and the parameter change, as shown in Equation (12). To improve the reparability of causal edges, a resampling stable selection is superimposed on the outer layer of the above solution: the occurrence frequency of the edges obtained by each resampling is statistically analyzed and converted into selection probability, as shown in Equation (13). Then, a stable edge set is obtained by screening according to the stability threshold, as shown in Equation (14). This stable edge set is used to output the final directed causal graph and serves as the input basis for subsequent key causal structure path extraction, making the source of the causal structure traceable and more robust to sample perturbations.

[0105] Specifically, the above construction method transforms time series causal discovery into a continuous optimization problem solvable on a computer: a linear structural equation with lag terms is used to express the dynamic influence of risk indicators and related factors; the coefficient matrix of the same time slice is used as the candidate adjacency structure of the directed causal graph; and intertemporal influences are allowed to characterize the lag effects of financial time series. The objective function is jointly controlled by fitting error and sparsity regularization, constraining the number of causal edges and facilitating interpretation; acyclicity is applied to the same time slice structure through a continuously differentiable discriminant function, ensuring the obtained graph meets the requirement of directed acyclic structure. To handle the optimization difficulty brought by equality constraints, an augmented Lagrangian form is introduced, incorporating the constraint residuals into the penalty term and coordinating with multiplier updates, allowing the optimization process to proceed along the constrained feasible direction. Gradient steps and sparse proximal updates are combined in the iteration, allowing the sparse structure to gradually emerge during the solution process. Outer-layer resampling stability selection uses the frequency of edge repetition to filter out accidental edges, reducing structural fluctuations caused by sample perturbations, thereby improving the reproducibility of subsequent path extraction and narrative generation.

[0106] Furthermore, the observation vector of the time series refers to the set of variables obtained by concatenating the values ​​of risk indicators and correlation factors within the same alignment period. The sample length refers to the number of effective alignment periods participating in causal discovery within the specified reporting period, excluding unavailable intervals. The lag order refers to the number of historical alignment period values ​​added to each variable, which is set by default to 1-5 days as in the original text and can be configured within 1-10 days as an implementation parameter. The upper limit of the lag order is constrained by the sample length and the number of variables in the reporting period to ensure convergence of the solution. The directed causal graph output by the same time slice consists only of factors within the same alignment period. The set of edges is constructed, with inter-period directed edges used to express lagged effects and only allowing points from earlier alignment periods to later alignment periods. The direction is fixed after variable standardization. Standardization processes mean-removal and variance-normalization of each variable within the reporting period, while retaining unit identifiers for backtesting interpretation. The mean and variance used for standardization are stored along with the batch identifier. The sparsity threshold is set to 0.05 and is configured within the range of 0.01 to 0.2 as an implementation parameter. The threshold value is jointly determined by the upper limit of the number of edges in the directed causal graph on the validation set and the interpretability, and is kept non-zero to avoid weak edges caused by numerical noise. The following parameters are retained: Step size in solving equality constraints defaults to 0.01 and can be configured within the range of 0.0001 to 0.1; penalty coefficient defaults to 1 and can be configured within the range of 0.1 to 100; multiplication factor defaults to 10 and can be configured within the range of 2 to 20; acyclic residual threshold defaults to 0.001 and can be configured within the range of 0.0001 to 0.01; relative change threshold defaults to 0.0001 and can be configured within the range of 0.00001 to 0.001. These values ​​are used in training data preparation. The stage is fixed and used when the batch identifier changes to ensure reproducibility; the resampling stability selection statistically analyzes the frequency of occurrence of causal edges obtained from multiple resamplings. The number of resamplings is set to 100 by default and can be configured within the range of 30 to 500. The edge coefficient judgment threshold is set to 0.05 by default and can be configured within the range of 0.01 to 0.2. The stability threshold is set to 0.6 by default and can be configured within the range of 0.5 to 0.9. The threshold is jointly determined by the coverage of the stable edge set on the validation set and the noise edge suppression rate and is kept not to 1 to avoid the stable edge set being empty.

[0107] In this embodiment, a prohibition constraint is introduced into the continuous optimization problem. The prohibition constraint is generated from a knowledge graph in the financial field and is used to prohibit the appearance of causal edges that are defined as not allowed by the knowledge graph in the directed causal graph.

[0108] The knowledge graph in the financial field includes a set of entity types and a set of relation types. The set of entity types includes at least market factors, credit factors, liquidity factors, holding attributes, business events, model parameters, and risk indicators. The set of relation types includes at least influence, determined by, belong to, constrain, and prohibit influence. The mapping from variables to knowledge graph entities is achieved through a mapping table, which includes at least variable name, entity identifier, entity type, and scope identifier. Prohibitory constraints are represented by a list of prohibited edges, which includes at least the starting entity type, the ending entity type, the relation direction, and the applicable scope identifier. When the mapped starting entity type and ending entity type match the list of prohibited edges, the corresponding causal edge is fixed to not be allowed to take non-zero edge weights during the optimization process, and the causal edge is kept non-existent in the output directed causal graph.

[0109] For example, the mapping from variable names to entity identifiers is completed using a mapping table during the data access phase. The matching is located by both the variable name and the caliber identifier, and the mapping batch identifier is written after a successful match. The prohibited edge list is deduplicated before use and indexed by the starting entity type, the ending entity type, and the relationship direction. The index is stored along with the knowledge graph version identifier. When the same variable is mapped to different entity types due to caliber differences, the prohibitive constraints are based on the prohibited edge rules matched by the caliber identifier and no cross-caliber merging is performed. When the mapping table cannot provide an entity identifier for a variable, the variable remains unrestricted in the prohibitive constraint stage of this causal discovery process and is written to the unmapped list. The unmapped list includes the variable name, caliber identifier, and source system identifier for subsequent completion.

[0110] In this embodiment, determining the key causal structure includes: extracting directed paths terminating at risk indicator nodes from the directed causal graph, calculating the causal strength score of each directed path, and selecting a predetermined number of directed paths with the highest causal strength scores as the key causal structure.

[0111] The causal strength score is obtained by multiplying the absolute values ​​of the edge weights of each causal edge on the path, with the edge weights taken from the sparsed adjacency matrix. When the path length exceeds four edges, the path is pruned and only the prefix path consisting of the first four edges is retained. When multiple paths share the same starting driving factor and terminate at the same risk indicator node, the path with the highest score is taken as the representative path of that driving factor, and the remaining paths are taken as alternative paths with a redundancy mark. The predetermined number is set to three paths. If there are fewer than three valid paths, the key causal structure is output according to the number of valid paths.

[0112] Similarly, the edge weights of causal edges on the path are taken from the sparsed adjacency matrix and retain the sign to express positive or negative influence. The path score uses the score obtained by multiplying the absolute values ​​of the edge weights to measure the influence strength without changing the directional interpretation. The numerical range of the edge weights is usually between 0 and 1 under standardized input, and cases greater than 1 are allowed. When the absolute value of the edge weight is greater than 1, it still participates in the calculation of the path score according to the original value, and the corresponding edge is marked with a strong influence in the output. The path length is capped at 4 and can be configured within 2 to 8 as an implementation parameter. The upper limit value is jointly determined by the readability of the interpretation and the risk of path explosion, and is kept below 1 to avoid losing the information of the mediating factor. The predetermined number is capped at 3 and can be configured within 1 to 10 as an implementation parameter. The upper limit of the predetermined number is determined by the maximum length constraint of the conditional input to ensure that the narrative generation model can be stably generated within a fixed length context. Redundancy markers are used to retain alternative paths for source tracing in multi-path scenarios with the same start and end points. Alternative paths are not included in the default conditional input but are stored in the evidence chain storage alongside the representative path.

[0113] In this embodiment, the change features include: using a change point detection algorithm to identify significant change points in the risk indicator sequence within a specified reporting period, and outputting the statistical comparison results of the subsequences before and after each change point;

[0114] Change point detection targets structural abrupt changes in mean and variance, with a detection window length of ten alignment periods and a minimum change point interval of five alignment periods. Significance is determined using a two-tailed test with a significance level of 0.05 as the threshold. For each significant change point, the change time, mean, standard deviation, quantiles, maximum retracement, and upward or downward direction markers for the ten alignment periods before and after the change point are output. When multiple significant change points are detected, they are sorted from largest to smallest by the magnitude of the change in mean of the risk indicator sequence caused by the change points, and the top three significant change points are retained for subsequent conditional input encoding.

[0115] Optionally, the detection window length is set to 10 alignment periods and configured as an implementation parameter within the range of 5 to 30 alignment periods; the minimum change point interval is set to 5 alignment periods and configured as an implementation parameter within the range of 1 to 10 alignment periods. The values ​​are determined jointly by the length of the reporting period and the fluctuation scale of the risk indicator sequence; the significance level is set to 0.05 and configured as an implementation parameter within the range of 0.01 to 0.1. The significance level value is determined by the upper limit of the acceptable false alarm rate in the historical report review and is kept non-zero to avoid generating overly dense change points for small numerical disturbances; when the effective alignment period within the specified reporting period is less than the detection window length, change point detection degenerates into a comparison of the statistics of the first half and the second half of the entire period, and the segment boundary is used as the change point time marker; when there are unavailable intervals that result in insufficient samples near the change point, the statistics of the subsequences before and after the change point are calculated based on the available samples, and the insufficient sample mark is written into the change point list, thereby ensuring that the change characteristics can always be generated and can be coded by conditional input.

[0116] In this embodiment, the risk narrative generation model is an autoregressive language model based on the Transformer architecture. The model is trained under supervision by pairing historical risk report paragraphs with corresponding standardized risk data templates and key causal structures to form training samples, so that the generated text is subject to conditional input constraints.

[0117] The standardized risk data template is represented by a table structure, which includes at least the start and end dates of the reporting period, portfolio identifier, risk indicator name and value, risk indicator name and value of the comparison period, change point time and statistics before and after the change point, key driver factor name, names and order of nodes in the critical path, path score, source data window identifier and caliber identifier. The text labels of the training samples are extracted from paragraphs in historical risk reports that correspond one-to-one with the template fields. During extraction, the numerical values ​​are uniformly replaced with template placeholders while retaining the units. The training corpus covers historical reports for 36 consecutive months and is divided into training set, validation set and test set in chronological order, with the most recent six months used as the test set. The generated output is organized according to a fixed chapter structure, which includes an overview of risk changes, driver attribution analysis, scenario and sensitivity points, and management actions and conclusions.

[0118] After solving the continuous optimization problem, the causal edges are resampled and stable selection is performed. The frequency of occurrence of causal edges obtained by multiple resampling is statistically analyzed and a stable edge set is selected. The stable edge set is used to generate a directed causal graph or as the input basis for subsequent key causal structure extraction.

[0119] Supervised training is organized around conditional input formatting, target narrative structuring, teacher forcing-based sequence loss, and evidence consistency auxiliary constraints, so that the model-generated text is continuously constrained by standardized risk data templates and key causal structures.

[0120] In one implementation, the supervised training method is as follows:

[0121] When constructing training samples, the conditional input of each sample is uniformly encoded as a conditional token sequence and paired with the target token sequence of the historical risk report paragraph. The conditional token sequence is written as:

[0122] (15)

[0123] In equation (15), Represents a sequence of conditional tokens; Indicates the length of the condition token sequence; This indicates the segmentation and tokenization operators; Indicates the starting marker; Indicates the end marker; Indicates sequence concatenation; Represents a serialized fragment of a standardized risk data template; A serialized fragment representing a key causal structure; A serialized fragment representing the characteristics of change.

[0124] The target risk narrative text is obtained by tokenizing the same token:

[0125] (16)

[0126] In equation (16), Represents the target token sequence; Indicates the length of the target token sequence; This refers to the text of the historical risk report paragraph.

[0127] To reduce off-topic content and sentence drift in regulatory contexts, the output text can be divided into fixed sections, and section tags can be added to the samples. This allows the model to learn the correspondence between paragraph structure and content scope during training. The target token sequence can be organized as follows:

[0128] (17)

[0129] In equation (17), This indicates the "Summary of Risk Changes" section marker; This indicates the "Driven Attribution" section marker; This indicates the "Sensitivity and Context" section marker; This indicates the "Management Actions" section marker; These represent four sections of text. Section tags are used as ordinary tokens in training, without introducing additional model structure; they are a constraint on the way samples are organized.

[0130] During supervised training, teacher forcing is employed, using the conditional token sequence and the known prefix target token as context. This minimizes the sequence cross-entropy loss and masks the padding positions. The loss can be written as:

[0131] (18)

[0132] In equation (18), Represents the sequence cross-entropy loss; This represents the parameters of the risk narrative generation model; Indicates the position index of the target sequence; This represents the location mask, with a value of 1 indicating that the location is a valid token and a value of 0 indicating that the location is padding; This represents the conditional probability distribution of the model for the next token under autoregressive conditions; Indicates the first One target token; This represents a known prefix target token subsequence; This represents the conditional token sequence. In actual training, a maximum length can be set and excessively long sequences can be truncated, while insufficient lengths can be padded with padding. On the attention side, both a causal mask and a padding mask are used to ensure that each position can only focus on the visible history and valid token positions.

[0133] To enhance factual consistency, the inferable information from template fields and path fragments can be constrained to the candidate set for generating tokens, and an evidence consistency auxiliary loss can be added to reduce the number of variable names and numerical placeholders that do not match the evidence during model training. An evidence-allowed set can be constructed first in the conditional input, for example, the allowed "variable name token set" and "numerical placeholder token set" extracted from the template fields and path fragments, denoted as... At each generation location, a "disable quality penalty" is applied to the prediction distribution belonging to the controlled vocabulary:

[0134] (19)

[0135] In equation (19), Indicates loss related to consistency of evidence; This represents the controlled vocabulary, which includes variable name token, index name token, unit token, and numeric placeholder token. This represents the controlled set of tokens allowed to appear under the current input conditions. This represents candidate tokens in the controlled vocabulary. This form penalizes the assignment of probabilities to tokens outside the permissible set of evidence, making the generated content closer to the evidentiary scope of the template fields and key causal structures.

[0136] The overall training objective is obtained by combining the main loss and the auxiliary loss:

[0137] (20)

[0138] In equation (20), Indicates the total training loss; This represents the weight of the evidence consistency loss, used to balance language fluency learning with the strength of evidence consistency constraints.

[0139] To complement the supervised training described above, the training samples can employ a normalization strategy involving placeholder replacement: numerical values ​​corresponding to template fields in the report text are replaced with placeholder tokens, and the replacement mapping is associated with and stored in relation to field keys in the conditional input. This allows the model to learn to output placeholders during training rather than freely generating numerical values; during inference, the specific numerical values ​​are then filled back in by the field keys, reducing numerical drift at the text level. If combined with a controlled vocabulary, the inference phase can implement constrained decoding of the controlled tokens, prioritizing tokens with mandatory variable names, prohibited vocabulary tokens, and placeholder tokens.

[0140] When constructing training samples, standardized risk data templates, key causal structures and change features are serialized and then uniformly tokenized to form a conditional token sequence, as shown in Equation (15); historical risk report paragraphs are tokenized to form a target token sequence, as shown in Equation (16). To reduce paragraph digression and structural drift, the target text in the training samples is split into fixed sections and section markers are inserted to enable the model to learn the correspondence between section boundaries and content range, as shown in Equation (17). In the training phase, teacher forcing is used, and the conditional token sequence and the known prefix target token are used together as context to calculate the sequence cross-entropy loss and apply the mask rule to the padding position, as shown in Equation (18). The maximum length and truncation strategy are uniformly configured in the training data preparation phase. To enhance factual consistency, controlled items such as deducible variable names, indicator names, units and numerical placeholders are extracted from the conditional input to form an allowable set, and a probability penalty is applied to the controlled candidate outputs that are not in the allowable set to form an evidence consistency auxiliary loss, as shown in Equation (19). The overall training objective is obtained by combining the main loss and the auxiliary loss, as shown in Equation (20). The balance weights of the two are pre-set by the technical personnel in the model training configuration stage, taking into account historical generation bias and verification requirements. In conjunction with supervised training, during the training data preparation stage, placeholders are used to replace the numerical values ​​in the report text and a mapping is established with the template field key, so that the model learns to output placeholders instead of freely generating numerical values. During the inference stage, the controlled vocabulary and placeholder backfilling are combined to perform constrained decoding, making it easier to verify and trace the generated narrative in terms of variable names, indicator names, and numerical references.

[0141] Specifically, the supervised training in the above implementation revolves around the relationship organization of data and objectives driven by conditional inputs: standardized risk data templates, key causal structures, and change features are serialized and used as contextual inputs; historical risk report paragraphs are used as output labels; and a known prefix-driven next-word prediction method is employed during training to learn narrative patterns. Output text is labeled with fixed sections, enabling the model to develop structured expression habits during training and reducing the probability of paragraphs diffusing across topics. The training objective is primarily sequence-level prediction error, while introducing evidence consistency constraints. The deducible variable names, indicator names, units, and numerical placeholders in the conditional inputs are limited to an allowed range, and a penalty is applied to candidate words that deviate from the evidence range, thus making the arguments closer to the supporting information of the template fields and causal paths. The accompanying numerical placeholder replacement allows the model to learn to "reference fields" rather than "fabricate values." Combining field backfilling and controlled vocabulary decoding during the inference phase improves the consistency and verifiability of regulatory texts.

[0142] Furthermore, a conditional token sequence refers to a sequence obtained by concatenating the field keys and values ​​of a standardized risk data template, the path node sequence and path score of key causal structures, and the comparison of change point times and statistics of change features in a fixed order, followed by word segmentation. The word segmentation dictionary remains consistent during training and inference and is stored along with the model version identifier. The maximum length is set to 2048 by default as an implementation parameter and can be configured within the range of 512 to 4096. Conditional token sequences exceeding the maximum length are truncated from the end in a fixed order, prioritizing the retention of the reporting period overview segment and key driver and path segments. Sequences shorter than the maximum length are padded with padding characters, and the padding positions are obscured during training loss calculation. The training corpus covers 36 consecutive months and can be configured within the range of 12 to 60 months as an implementation parameter. The most recent 6 months are used as the test set and are also allowed as an implementation parameter. Configuration can be performed within 3 to 12 months, with the timeline ensuring no posterior leakage. The controlled vocabulary includes a set of tokens corresponding to variable names, indicator names, units, and numerical placeholders. This set can be extracted from the field keys and path node names appearing in the conditional input and is fixed within each sample. The evidence consistency loss weight is set to 0.1 by default as an implementation parameter and can be configured within the range of 0.01 to 1. The weight value is jointly determined by the controlled token violation rate and language fluency score on the validation set and is kept from being 0 to avoid indicator names and placeholders inconsistent with the conditional input. Numerical placeholder backfilling reads specific values ​​from the standardized risk data template with unit identifiers at the field key index during the inference phase. Before backfilling, the numerical values ​​are checked for consistency. If the check fails, the placeholder remains unfilled, and a list of unfilled keys is recorded at the end of the text.

[0143] This embodiment also includes a source tracing step: associating and storing the assertion sentences in the risk analysis narrative text with the path segments in the key causal structure that supports the assertion sentences and the corresponding source data time window identifiers to form a queryable chain of evidence;

[0144] The assertion sentence is determined at the sentence level. The assertion sentence must simultaneously contain the name of the risk indicator and a word indicating the direction of change. The word indicating the direction of change is limited to one of the following: increase, decrease, expansion, convergence, improvement, or deterioration. The path segment is determined by matching the names of the driving factors and mediating factors that appear in the assertion sentence. When the assertion sentence only contains the name of the driving factor, the representative path corresponding to that driving factor is taken as the path segment. The source data time window identifier is generated according to the time of the change point corresponding to the assertion sentence. The time window starts at the ten alignment periods before the change point and ends at the ten alignment periods after the change point. The evidence chain storage record must contain at least the text hash of the assertion sentence, the start and end dates of the reporting period, the name of the risk indicator, the sequence of path segment nodes, the path score, the set of source system identifiers, the set of caliber identifiers, and the start and end timestamps of the time window.

[0145] In this embodiment, the text hash of the assertion sentence is calculated after standardizing whitespace characters and unifying unit symbols in the assertion sentence before generation. The default length of the summary is 64 characters, which can be configured within the range of 32 to 128 characters as an implementation parameter. The summary length and the generation batch identifier are stored together for cross-version deduplication. The set of variable direction words is used for assertion sentence identification. If multiple risk indicator names and multiple direction words appear in the same sentence, it is divided into multiple assertion sentences according to the risk indicator names and evidence chains are established for each sentence. When an assertion sentence does not contain direction words but contains risk indicator names and numerical placeholders, the sentence is considered valid. Records that are not included in the assertion sentence set and are retained as non-assertion sentence citations in the verification failure list; the source data time window identifier is generated centered on the change point time. If the change point time is less than 10 alignment cycles from the start or end of the reporting period, the time window is truncated at the boundary and a truncation mark is written. After truncation, the window is still used for backtracking queries and the start and end timestamps of the window are retained in the evidence chain record; the caliber identifier set is used in the evidence chain record to indicate the caliber of the indicator cited by the assertion sentence. If the same assertion sentence involves multiple caliber fields, it is split into multiple evidence chain records according to the caliber identifier and associated with the corresponding source system identifier set respectively. Example 2:

[0146] Based on Embodiment 1, this embodiment provides an enterprise digital asset integrated management system, including:

[0147] The data interface module is used to acquire and time-align the multivariate time series data of the target asset portfolio within a specified reporting period;

[0148] The source systems that the data interface module interfaces with include at least trading or position systems, valuation systems, market data systems, and risk measurement systems. The data packets output by the data interface module adopt a unified field naming convention, which includes an indicator name prefix, a scope suffix, and a unit suffix. When the same indicator has duplicate fields in different source systems, a unique field is selected according to the priority of the scope identifier, and a conflict list is recorded. The data interface module generates a batch identifier for each data acquisition, and the batch identifier is associated with each data point in the output data packet and written to the log.

[0149] The causal analysis module is used to perform causal discovery processing based on multivariate time series data to generate a directed causal graph and identify key causal structures;

[0150] The output of the causal analysis module includes a list of nodes, edges, edge weights, hysteresis markers, and a list of candidate edges removed by prohibitive constraints in the directed causal graph. The key causal structure is represented by a path list, which contains the node sequence, edge weight sequence, path score, representative path marker, and redundancy marker for each path. When the batch identifier of the input data packet changes, the causal analysis module re-executes the causal discovery process and stores the output in association with the batch identifier.

[0151] The feature extraction module is used to extract the variation features of the risk indicator sequence;

[0152] The narrative generation module encodes key causal structures and change characteristics as conditional inputs and feeds them into the risk narrative generation model to output risk analysis narrative text.

[0153] The narrative generation module performs chapter integrity checks on the output text. The chapter integrity check requires the output text to include three parts: an overview of risk changes, driving attribution analysis and management actions and conclusions, and to begin with a fixed title. The narrative generation module generates a timestamp and model version identifier for each part and establishes a correlation between the model version identifier and the training data version identifier. When the chapter integrity check fails, the narrative generation module triggers a supplementary generation process based on the missing chapter name and only allows the use of fields and structured text fragments from the conditional input.

[0154] And a source tracing and storage module, used to establish a chain of evidence between assertions in the risk analysis narrative text, path segments, and source data time window identifiers;

[0155] The traceability storage module provides query indexes by reporting period, asset portfolio identifier, risk indicator name, and assertion statement hash; each evidence chain link is simultaneously associated with data packet batch identifier, causal graph output batch identifier, and text output batch identifier; when any batch identifier changes, the traceability storage module marks the historical evidence chain link as the historical version and keeps it traceable, while the latest evidence chain link is marked as the current version and output as the default query result.

[0156] Specifically, the data packet batch identifier is generated by the data interface module after each successful acquisition and time alignment. The batch identifier is represented by a string that uniquely identifies a single acquisition result and is written to the log along with each data point. The causal graph output batch identifier is generated by the causal analysis module after completing causal discovery processing and outputting a list of nodes, edges, edge weights, and hysteresis markers. The output batch identifier is associated one-to-one with the input data packet batch identifier and stored. The text output batch identifier is generated by the narrative generation module after completing chapter integrity verification, consistency verification, and necessary regeneration. The output batch identifier, along with the model version identifier and training data version identifier, is written to the traceability storage module. When the source system is temporarily unavailable, causing data acquisition failure, the data interface module retains the most recent successful batch as a fallback input and adds a reuse marker to the batch identifier. The reuse marker triggers the narrative generation module to append a data timeliness prompt phrase to the first paragraph of the text without changing the original conclusion structure. When chapter integrity verification fails and the supplementary generation process still cannot complete the missing chapters, the text output batch identifier is still generated but with an incomplete marker. The incomplete marker is filtered as a non-default result in the query index, thus ensuring that the default query returns a verifiable complete version output.

[0157] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the enterprise digital asset integrated management method and system proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0159] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A comprehensive management method for enterprise digital assets, characterized in that, include: Step S1: Upon receiving a request to generate a risk analysis narrative for the target asset portfolio, acquire and time-align the multivariate time series data of the target asset portfolio within a specified reporting period. The multivariate time series data includes at least a risk indicator sequence and multiple correlation factor sequences corresponding to the risk indicator sequence. Step S2: Perform causal discovery processing on a computer based on the multivariate time series data to obtain a directed causal graph characterizing the correlation factor to the risk indicator, and determine at least one directed path from the directed causal graph as a key causal structure. The nodes of the directed causal graph correspond to the variables of the risk indicator sequence and the correlation factor sequence, respectively. Step S3: Extract the change characteristics of the risk indicator sequence during the specified reporting period, and encode the change characteristics and the key causal structure together as conditional input; Step S4: Input the conditions into the trained risk narrative generation model to generate and output the risk analysis narrative text corresponding to the specified reporting period.

2. The enterprise digital asset integrated management method according to claim 1, characterized in that, The time alignment includes: unifying the frequency of the risk indicator sequence and the correlation factor sequence, handling missing values ​​and outliers, and attaching a source system identifier and a timestamp identifier to each data point.

3. The enterprise digital asset integrated management method according to claim 1, characterized in that, The causal discovery process includes: constructing a continuous optimization problem containing sparsity constraints and acyclic constraints to learn an adjacency matrix representing the causal relationship of variables, and obtaining the directed causal graph by solving the problem; The continuous optimization problem represents variables as combinations of current values ​​and several lagged values ​​when expanding lagged variables, so as to allow directed edges across periods to express lagged effects, and applies acyclic constraints only to the causal structure corresponding to the same time slice, so that the directed causal graph output in the same time slice satisfies the requirement of directed acyclic interpretation.

4. The enterprise digital asset integrated management method according to claim 3, characterized in that, In the continuous optimization problem, a prohibition constraint is introduced. This prohibition constraint is generated from a knowledge graph in the financial domain and is used to prohibit the occurrence of causal edges in the directed causal graph that are defined as disallowed by the knowledge graph.

5. The enterprise digital asset integrated management method according to claim 1, characterized in that, Determining the key causal structure includes: extracting directed paths terminating at risk indicator nodes from the directed causal graph, calculating the causal strength score of each directed path, and selecting a predetermined number of directed paths with the highest causal strength scores as the key causal structure.

6. The enterprise digital asset integrated management method according to claim 1, characterized in that, The change characteristics include: using a change point detection algorithm to identify significant change points in the risk indicator sequence during the specified reporting period, and outputting the statistical comparison results of the subsequences before and after each change point.

7. The enterprise digital asset integrated management method according to claim 4, characterized in that, The risk narrative generation model is an autoregressive language model based on the Transformer architecture. The model is trained under supervision by pairing historical risk report paragraphs with corresponding standardized risk data templates and key causal structures to form training samples, so that the generated text is constrained by the conditional input. After solving the continuous optimization problem, the causal edges are resampled for stability selection. The frequency of occurrence of causal edges obtained by multiple resampling is statistically analyzed and a set of stable edges is selected. The set of stable edges is used to generate a directed causal graph or as the input basis for subsequent extraction of key causal structures.

8. The enterprise digital asset integrated management method according to claim 1, characterized in that, It also includes a source tracing step: associating and storing the assertions in the risk analysis narrative text with the path segments in the key causal structure that support the assertions and the corresponding source data time window identifiers to form a queryable chain of evidence.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

10. A comprehensive digital asset management system for enterprises, based on the method described in any one of claims 1 to 8, characterized in that, include: The data interface module is used to acquire and time-align the multivariate time series data of the target asset portfolio within a specified reporting period; The causal analysis module is used to perform causal discovery processing based on the multivariate time series data to generate a directed causal graph and determine key causal structures; The feature extraction module is used to extract the variation features of the risk indicator sequence; The narrative generation module is used to encode the key causal structure and the change features as conditional inputs and input them into the risk narrative generation model to output risk analysis narrative text. And a source tracing and storage module, used to establish a chain of evidence links between assertions in the risk analysis narrative text, path segments, and source data time window identifiers.

Images