Automatic enterprise financial analysis system and method based on AI

Abstract

The invention discloses an AI-based automatic enterprise financial analysis system and method, and relates to the technical field of financial analysis, and the method comprises the steps of S1, data integration, S2, time-space diagram modeling, S3, feature fusion, S4, risk prediction, and S5, early warning output. According to the method, the contribution degrees of different modes can be automatically adjusted according to real-time data characteristics, the weight of market signals is enhanced when the industry fluctuates violently, or the sensitivity of unstructured enterprise related information data is improved in a supply chain crisis event; by means of the design, subjective deviation caused by manual weight setting is avoided, the nonlinear synergistic effect between different modes can be captured, and the capacity of recognizing hidden risks is improved; meanwhile, the complex transaction relation between the enterprises is converted into quantifiable topological characteristics through space-time diagram structure modeling, so that the risk prediction model can analyze the interaction influence between the individual states of the enterprises and the industry ecological network at the same time, and the systematic risk conduction path is evaluated on the overall level.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of financial analysis, and particularly relates to an AI-based automatic enterprise financial analysis system and method. BACKGROUND

[0002] With the complication of global industrial chains and the dynamicization of market environments, the risks faced by enterprises are not only derived from their own financial conditions, but also closely related to the trading credit of upstream and downstream supply chains, the fluctuation of market enterprise-related information, and the interactive influence of industry ecological networks.

[0003] In the field of enterprise financial analysis and risk assessment, accurately identifying enterprise financial risks and systemic transmission paths is crucial for enterprise decision-making, industry regulation, and financial stability. Traditional financial analysis methods mainly rely on single structured financial data (such as balance sheets and profit statements) and use static index calculations or simple machine learning models for risk assessment, which cannot integrate multi-source heterogeneous data (such as unstructured enterprise-related information and supply chain transaction relationships) inside and outside the enterprise, resulting in insufficient ability to capture implicit risks. To address the above problems, the following solutions are proposed. SUMMARY

[0004] The purpose of the present application is to provide an AI-based automatic enterprise financial analysis system and method, which can automatically adjust the contribution of different modalities according to real-time data characteristics through a dynamic modal fusion weight mechanism, solving the problem of existing technologies that use static index calculations or simple machine learning models for risk assessment, which cannot integrate multi-source heterogeneous data inside and outside the enterprise.

[0005] To solve the above technical problems, the present application is realized by the following technical solutions:

[0006] The present application is an AI-based automatic enterprise financial analysis system and method, which includes:

[0007] Step S1, data integration: collect structured financial data, supply chain relationships, and unstructured enterprise-related information data inside and outside the enterprise, store them in a distributed database after classification and storage, and perform standardization alignment;

[0008] Step S2, spatiotemporal graph modeling: build a dynamic enterprise relationship graph, define node attributes and edge weights, and generate a continuous time series graph sequence by quarter;

[0009] Step S3, feature fusion: extract time series features through bidirectional LSTM, extract supply chain graph features through GraphSAGE, extract market features through CNN, and adaptively weight and fuse the three types of modal features based on dynamic modal fusion weights;

[0010] Step S4, risk prediction: design a double-layer attention mechanism to combine historical time series features and supply chain associated impacts, output the risk probability value of the next quarter of the enterprise;

[0011] Step S5, early warning output: trigger early warning according to industry dynamic threshold, automatically generate report.

[0012] Further, the step S1, data integration specifically includes the following steps:

[0013] Step S11: define the data source, specifically:

[0014] Internal structured data: extract financial statements (balance sheet, cash flow statement, profit statement), subject balance sheet, accounts receivable and payable from enterprise ERP system;

[0015] External structured data: crawl public market data (stock price, industry index), public financial indicators of upstream and downstream enterprises in the supply chain;

[0016] Unstructured data: obtain news related to the enterprise (including keyword sentiment analysis of the enterprise name) through NLP interface, CSRC punishment announcement, and social media related to the enterprise;

[0017] Step S12: establish a distributed database, and store and align according to the following labels:

[0018] Financial indicators: select multiple core indicators and align them according to quarterly timestamps;

[0019] Supply chain map: take the enterprise as the node, and the edge attributes include transaction amount, account period, and historical default times;

[0020] Market signal: industry beta coefficient, 30-day volatility, and short selling ratio;

[0021] Enterprise-related information label: sentiment polarity (-1 to 1), and keyword coverage rate (TF-IDF weighting);

[0022] Integrate structured and unstructured data sources inside and outside the enterprise to solve the problem of single data and information silos in financial analysis;

[0023] Through classified storage, a standardized data system is established to provide a multi-modal input basis for subsequent spatio-temporal graph modeling, ensuring data traceability and timeliness.

[0024] Further, the step S2, spatio-temporal graph modeling specifically includes the following steps:

[0025] Step S21: define node attributes, the attribute vector of each enterprise node v i contains: ​

[0026]

[0027] In the formula, As a business v i The node attribute vector at time t, t is the time identifier (such as 2023Q4), representing the current quarter; ROE is the return on equity, and the public opinion score is the numerical score obtained through NLP sentiment analysis;

[0028] The dimension is 32 dimensions, and the PCA dimension reduction is performed;

[0029] Step S22: Constructing edge relationship, when there is a supply chain transaction between enterprises v i And v j , a directed edge e ij Is constructed, and the weight w ij The calculation formula is:

[0030]

[0031] In the formula, w ij The edge weight between enterprises v i And v j , the annual transaction amount is the annual transaction amount of enterprises v i And v j , the historical number of delayed payment is the historical number of delayed payment from enterprises v i To v j , and the total transaction times are the total transaction times of both parties;

[0032] Step S23: Dynamic time slicing, generating time sequence snapshot G (t) ={V (t) ,E (t) ,X (t)} per quarter, retaining data of 8 consecutive quarters (window period) to form a graph sequence {G (t-7) ,…,G (t)}, wherein G (t) Is the time sequence snapshot at time t, V (t) Is the node set at time t, E (t) Is the edge set at time t, and X (t) Is the node attribute matrix at time t;

[0033] The enterprise entity, supply chain relationship and market environment are abstracted as a dynamic graph structure, and the spatio-temporal correlation between enterprises is explicitly described.

[0034] The historical state evolution is retained through the time sequence snapshot, and a calculable topological framework is provided for risk transmission path analysis.

[0035] Further, the step S3, the feature fusion specifically includes the following steps:

[0036] Step S31: Extracting time sequence features, using bidirectional LSTM to process the financial indicator sequence of each enterprise, specifically:

[0037]

[0038] In the formula, is the forward LSTM vector v i The encoding result of the financial indicators of the past 8 quarters, is the reverse sequence encoding result of the same data by reverse LSTM, is the vector of fused bidirectional time sequence features, is the vector splicing operation;

[0039] Step S32: Extracting graph structure features, using GraphSAGE algorithm to aggregate neighbor information, specifically:

[0040]

[0041] In the formula, is the neighbor aggregation feature of enterprise node v i N(i) is the neighbor node set of enterprise v i W agg is the trainable weight matrix of neighbor feature aggregation, is the final graph structure feature of enterprise node v i W comb is the trainable weight matrix of the combination of self feature and neighbor feature;

[0042] Step S33: Modal fusion, calculating the modal weight α time ,α graph ,α market according to the formula, specifically:

[0043]

[0044] In the formula, is the fusion weight of the mth modal at time t, W a is the trainable weight matrix, is the feature vector of the mth modal at time t, g (t) is the global time sequence state, σ is the LeakyReLU activation function, and M is the total number of modes;

[0045] Then generate the fused features:

[0046]

[0047] In the formula, is the vector of market data features extracted by CNN, αtime ,α graph ,α market are the fusion weights of time series, graph structure and market modality, respectively;

[0048] The dynamic financial indicators and supply chain structure features are extracted by using a time series model (LSTM) and a graph neural network (GraphSAGE), respectively;

[0049] Cross-modal feature adaptive weighting is realized through dynamic modality fusion weights to enhance the expression ability of complex risk signals.

[0050] Further, the step S4 of risk prediction specifically comprises the following steps:

[0051] Step S41: A double-layer attention mechanism is designed, specifically as follows:

[0052] Time attention layer: The contribution degree of each historical quarter to the current prediction is calculated:

[0053]

[0054] In the formula, β (k) is the time attention weight of the kth historical quarter to the prediction, q is a trainable query vector, T is a transpose operator, W t is a time attention weight matrix, is the fusion feature vector of the enterprise v i at time t-k;

[0055] Cross-enterprise attention layer: The influence weight of the supply chain associated enterprise is calculated:

[0056]

[0057] In the formula, γ ij is the influence weight of the supply chain associated enterprise v j on the target enterprise v i , and W r is a trainable weight matrix for modeling the cross-enterprise relationship;

[0058] Step S42: Prediction output, the weighted features are input into a fully connected layer:

[0059]

[0060] In the formula, W p is a fully connected layer weight matrix, b is a bias term, and τ is a Sigmoid function;

[0061] The designed hierarchical attention mechanism (time + cross-enterprise) captures key risk driving factors and solves the problem of insufficient modeling of long-range dependencies in traditional models;

[0062] Fusion of spatio-temporal features and associated enterprise influence, generate quantitative risk score, improve the prediction accuracy of financial risk transmission chain.

[0063] Further, the step S5, early warning output specifically includes the following steps:

[0064] Step S51, threshold determination: dynamic threshold adjustment, according to the industry risk median μ and standard deviation

[0065] Step S52, report automatic generation: key factor attribution, extract attention weight And β (k) , generate explanation text and chart;

[0066] Based on the dynamic threshold, the risk grading early warning is realized, and the false alarm / miss alarm caused by the fixed threshold is avoided;

[0067] Through the attention weight, the explainability report is fed back, the risk attribution analysis and visual decision support are output, and the enterprise audit and supervision demand is met.

[0068] The application is an AI-based automatic enterprise financial analysis system, characterized in that the analysis system comprises a data acquisition module, a space-time graph modeling module, a feature fusion module, a risk prediction module and an early warning output module;

[0069] The data acquisition module, the space-time graph modeling module, the feature fusion module, the risk prediction module and the early warning output module are connected in sequence.

[0070] Further, the data acquisition module is used for collecting structured and unstructured data from multiple channels, completing data cleaning, alignment and classified storage;

[0071] The space-time graph modeling module is used for constructing a dynamic enterprise relationship graph, defining node attributes and edge weights, and generating a time sequence graph sequence for representing the dynamic correlation relationship between enterprises;

[0072] The feature fusion module is used for extracting financial time sequence features, supply chain graph structure features and market enterprise related information features respectively, and realizing adaptive weighted fusion of cross-modal features through dynamic modal fusion weight;

[0073] The risk prediction module captures key risk signals based on a hierarchical attention mechanism, and outputs future quarter financial risk probability values through a full connection network;

[0074] The early warning output module is used for triggering risk early warning according to the dynamic threshold, and automatically generating a standardized report containing attribution analysis, visual chart and coping suggestions.

[0075] The present application has the following advantages:

[0076] 1、 The present application can automatically adjust the contribution of different modalities according to real-time data characteristics through a dynamic modal fusion weight mechanism, enhance the weight of market signals in a volatile industry, or improve the sensitivity of unstructured enterprise-related information data in a supply chain crisis event. This design not only avoids subjective bias in manual weight setting, but also captures the nonlinear synergistic effect between different modalities, improving the ability to identify implicit risks. At the same time, the space-time graph structure modeling converts the complex transaction relationship between enterprises into quantifiable topological features, enabling the risk prediction model to analyze the interaction between individual enterprise states and industry ecological networks, thereby assessing the systemic risk transmission path at the global level.

[0077] 2、 In the time dimension, the combination of bidirectional LSTM and time attention mechanism can capture both long-term trends of financial indicators and short-term mutation signals. In the spatial dimension, the graph neural network aggregates the features of upstream and downstream nodes in the supply chain to automatically discover the path of risk transmission across enterprises. At the same time, the dynamic time slicing technology preserves the evolution process of historical graph structures, enabling the model to learn the phased characteristics of risk accumulation.

[0078] 3、 The time attention layer quantifies the contribution of each historical period to the current risk, automatically identifying key turning points. The cross-enterprise attention layer reveals the location of high-risk nodes in the supply chain network and their impact strength. This design not only improves the model's focus on key risk signals, but also makes the prediction results business interpretable through visualized weight distribution. When the system warns an enterprise, it can clearly indicate that the risk mainly comes from the associated influence of a specific supplier or the persistent deviation of a certain financial indicator, reducing the application threshold of artificial intelligence models in financial supervision scenarios and enabling risk control personnel to quickly locate the problem source and develop targeted strategies.

[0079] Of course, implementing any product of the present application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0080] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0081] Figure 1 The flowchart of the present application is shown in the figure.

[0082] Figure 2A framework schematic diagram of an AI-based automatic enterprise financial analysis system of the present application. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0084] Please refer to Figure 1 The present application is an AI-based automatic enterprise financial analysis method, which comprises the following steps:

[0085] Step S1, data integration:

[0086] Step S11, defining data sources, specifically:

[0087] Internal structured data: extracting financial statements, subject balance tables, accounts receivable and payable from enterprise ERP systems;

[0088] External structured data: crawling public market data and public financial indicators of upstream and downstream enterprises in the supply chain;

[0089] Unstructured data: obtaining news related to enterprises, CSRC punishment announcements and social media related discussions about enterprises through NLP interfaces;

[0090] Step S12, establishing a distributed database, and storing and aligning according to the following labels:

[0091] Financial indicators: selecting multiple core indicators and aligning according to quarterly timestamps;

[0092] Supply chain map: taking enterprises as nodes, and the edge attributes including transaction amount, account period and historical default times;

[0093] Market signals: industry beta coefficient, 30-day volatility and short selling ratio;

[0094] Enterprise-related information labels: sentiment polarity and keyword coverage rate.

[0095] Step S2, spatiotemporal graph modeling:

[0096] Step S21, defining node attributes, and the attribute vector of each enterprise node v i contains:

[0097]

[0098] In the formula,​ For enterprise v i Node attribute vector at time t, t is time identifier (such as 2023Q4), representing the current quarter; ROE is the return on equity, and the public opinion score is the numerical score obtained through NLP sentiment analysis;

[0099] The dimension is 32, and the PCA dimension reduction is performed;

[0100] Step S22: Constructing edge relationship, when there is a supply chain transaction between enterprise v i and v j , a directed edge e ij is constructed, and the weight w ij is calculated according to the following formula:

[0101]

[0102] In the formula, w ij is the edge weight between enterprise v i and v j , the annual transaction total is the annual transaction amount of enterprise v i and v j , the historical number of delayed payment is the historical number of delayed payment from enterprise v i to v j , and the total transaction number is the total historical transaction number between the two parties.

[0103] Step S23: Dynamic time slicing, generating time sequence snapshot G (t) ={V (t) ,E (t) ,X (t)} by quarter, retaining data of 8 consecutive quarters to form graph sequence {G (t-7) ,…,G (t)}, wherein G (t) is the time sequence snapshot at time t, V (t) is the node set at time t, E (t) is the edge set at time t, and X (t) is the node attribute matrix at time t.

[0104] Step S3, feature fusion:

[0105] Step S31: Extracting time sequence features, using bidirectional LSTM to process the financial indicator sequence of each enterprise, specifically:

[0106]

[0107]

[0108] In the formula, is the forward LSTM of v iThe coding results of the financial indicators in the past 8 quarters, The coding results of the reverse sequence of the same data for the reverse LSTM, The vector fused with the bidirectional time sequence features, The vector splicing operation;

[0109] Step S32: Extract the graph structure features, and aggregate the neighbor information by using the GraphSAGE algorithm. Specifically,

[0110]

[0111] In the formula, is the neighbor aggregation feature of the enterprise node v i , N(i) is the neighbor node set of the enterprise v i , W agg is a trainable weight matrix for neighbor feature aggregation, is the final graph structure feature of the enterprise node v i , W comb is a trainable weight matrix for combining the self feature and the neighbor feature;

[0112] Step S33: Modal fusion, calculate the modal weights α time , α graph , and α market according to the formula, and the specific formula is:

[0113]

[0114] In the formula, is the fusion weight of the mth modal at time t, W a is a trainable weight matrix, is the feature vector of the mth modal at time t, g (t) is the global time sequence state, σ is the LeakyReLU activation function, and M is the total number of modes;

[0115] The fusion features are regenerated as follows:

[0116]

[0117] In the formula, is the vector extracted by CNN from the market data features, α time , α graph , and α market are the fusion weights of the time sequence, graph structure, and market modal respectively.

[0118] Step S4: Risk prediction,

[0119] Step S41: Design a double-layer attention mechanism, which is specifically as follows:

[0120] Temporal attention layer: calculate the contribution of each historical quarter to the current prediction:

[0121]

[0122] where β (k) is the temporal attention weight of the kth historical quarter to the prediction, q is the trainable query vector, T is the transpose operator, W t is the temporal attention weight matrix, is the enterprise v i at time t-k;

[0123] Cross-enterprise attention layer: calculate the influence weight of the supply chain associated enterprise:

[0124]

[0125] where γ ij is the influence weight of the supply chain associated enterprise v j to the target enterprise v i , W r is the trainable weight matrix for modeling cross-enterprise relationships.

[0126] Step S42: prediction output, input the weighted features into the fully connected layer:

[0127]

[0128] where W p is the fully connected layer weight matrix, b is the bias term, and τ is the Sigmoid function.

[0129] Step S5, early warning output:

[0130] Step S51, threshold determination: dynamic threshold adjustment, according to the industry risk median μ and standard deviation

[0131] Step S52, automatic report generation: key factor attribution, extract attention weights and β (k) , generate explanation text and charts.

[0132] Please refer to Figure 2 , the present application is an AI-based automated enterprise financial analysis system, comprising a data acquisition module, a spatio-temporal graph modeling module, a feature fusion module, a risk prediction module and an early warning output module;

[0133] The data acquisition module, the spatio-temporal graph modeling module, the feature fusion module, the risk prediction module and the early warning output module are connected in sequence.

[0134] The data acquisition module is used for collecting structured and unstructured data from multiple channels, completing data cleaning, alignment and classified storage.

[0135] The spatio-temporal graph modeling module is used for constructing a dynamic enterprise relationship graph, defining node attributes and edge weights, and generating a time series graph sequence for representing dynamic correlation relationships between enterprises.

[0136] The feature fusion module is used for extracting financial time series features, supply chain graph structure features and market enterprise related information features, respectively, and realizing adaptive weighted fusion of cross-modal features through dynamic modal fusion weights.

[0137] The risk prediction module captures key risk signals based on a hierarchical attention mechanism and outputs future quarter financial risk probability values through a fully connected network.

[0138] The early warning output module is used for triggering risk early warning according to a dynamic threshold and automatically generating a standardized report containing attribution analysis, visual charts and coping suggestions.

[0139] One specific application of the embodiment is:

[0140] Enterprise background: XX Automobile Group is one of the top ten automobile manufacturers in the world, with more than 200 suppliers and distributors, and its business involves vehicle manufacturing, parts production and sales services. Based on the operation data of the enterprise from Q1 2020 to Q4 2022, the financial risk of the enterprise in Q1 2023 is predicted.

[0141] 1. Data acquisition

[0142] (1) Internal data acquisition:

[0143] Financial indicators: Extract the balance sheet (total assets, current liabilities, etc.), cash flow statement (net cash flow from operating activities), and income statement (gross profit margin, net profit margin) for the past 8 quarters from the SAP system, and standardize them into 12 core indicators.

[0144] Supply chain data: Extract transaction records with 35 core suppliers, including quarterly transaction amount and accounts payable period (e.g. Supplier A: average payment period 90 days, historical delayed payment 3 times).

[0145] (2) External data integration:

[0146] Market data: Obtain automobile industry index (average volatility rate of the past 8 quarters is 18%) and enterprise stock 30-day Beta value (1.2) through Wind API.

[0147] Enterprise-related information data: 12,000 news articles containing the keyword "XX car" were crawled during Q4 of 2022. According to the sentiment analysis of the BERT model, negative enterprise-related information accounted for 15% (mainly involving "chip shortage leading to reduced production capacity").

[0148] (3) Data alignment and cleaning:

[0149] All data are aligned quarterly, and missing values ​​are filled using time series interpolation (e.g., missing dealer data for Q3 2021 is replaced by the average of the preceding and following quarters).

[0150] Unstructured text data is converted into structured tags (e.g., news article "XX car battery supplier catches fire" → sentiment score -0.8, keyword "supply chain risk" weight 0.7);

[0151] 2. Example of Spatiotemporal Graph Modeling

[0152] (1) Node and edge definition:

[0153] Node attributes:

[0154] XX Automobile Group node: ROE = 8.200, Accounts Receivable Turnover Ratio = 6.5, Cash Flow to Current Liabilities Ratio = 0.9, Enterprise Related Information Score = -0.6 (32-dimensional vector after PCA dimensionality reduction);

[0155] Key supplier nodes: Attributes include their financial health (e.g., Supplier B: current ratio 1.3, enterprise-related information score 0.2);

[0156] Edge weight calculation:

[0157] Edge weight calculation with supplier C:

[0158] w ij =log(1+85 million / 10 6 )×(1-2 / 15)=2.24×0.87=1.95;

[0159] (2) Generation of dynamic graph sequences:

[0160] Construct 8 time series snapshots from Q1 2020 to Q4 2022. Each snapshot contains:

[0161] Nodes: XX Automobile Group, 35 suppliers, 20 dealers;

[0162] Edges: Supply chain transaction relationships, equity relationships (e.g., distributor D holds 5% of the shares);

[0163] 3. Feature Fusion Examples

[0164] (1) Temporal feature extraction:

[0165] The ROE sequence of XX Automobile Group (Q8 values: [7.1%, 7.8%, ..., 8.2%]) was processed using a bidirectional LSTM, outputting a 128-dimensional time-series feature vector.

[0166] (2) Graph structure feature extraction:

[0167] Perform GraphSAGE aggregation on the supplier network:

[0168] Supplier A's neighbors include XX Auto and 3 Tier 2 suppliers, which, after aggregation, generate a 64-dimensional graph feature.

[0169] (3) Dynamic modal fusion

[0170] Calculate the modal weights:

[0171] α time =0.52,α graph =0.33,α market =0.15;

[0172] Feature fusion generation:

[0173]

[0174] 4. Risk Prediction and Attribution Analysis

[0175] (1) Attention weight calculation:

[0176] Temporal attention: The feature weight was highest in Q4 2021 (β) (5) =0.26, corresponding to the "peak period of chip shortage after the pandemic";

[0177] Cross-enterprise attention: The influence weight γ of supplier E (core battery supplier) on XX Automobile ij =0.41 (because this supplier accounts for 35% of the purchase amount);

[0178] (2) Risk score output:

[0179] Predicted risk probability for Q1 2023:

[0180] RiskScore = 0.67 (threshold = 0.58);

[0181] Warning triggered: Exceeding the "Dynamic Threshold for the Automobile Manufacturing Industry" (industry average 0.43 + 1.5 × 0.08 + 0.1 × 0.18);

[0182] 5. Report Output Examples

[0183] (1) Content of the early warning report:

[0184] Key attribution:

[0185] Supply chain risk: deteriorating financial health of supplier E (current ratio dropped to 1.1, 41% contribution);

[0186] Internal factor: 23% YoY decrease in operating cash flow in Q4 2022 (29% contribution);

[0187] Market signal: industry short interest ratio rose to 12% (18% contribution);

[0188] (2) Visualization charts:

[0189] Trend chart: shows 8-quarter risk score curve with warning threshold breakthrough points marked;

[0190] Supply chain heat map: highlights risk transmission path of supplier E and its secondary suppliers;

[0191] Comparison table: lists XX Auto's cash flow to liabilities ratio (0.9 vs. 1.4) against industry leader (e.g., Toyota);

[0192] In this embodiment, the system warned XX Auto Group of financial risks one quarter in advance, accurately identified supply chain bottlenecks and cash flow problems, and the company adjusted supplier payment terms and increased short-term loans accordingly. As a result, the actual risk occurrence probability in Q1 2023 was 0.61 (lower than the predicted value of 0.67), verifying the effectiveness of the method.

[0193] In the description of the present specification, the description referring to the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In the present specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in an appropriate manner.

[0194] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and do not limit the application to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of the present specification. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their entire scope and equivalents.

Claims

1. An AI-based automated enterprise financial analysis method, characterized in that, The analytical method includes the following steps: Step S1, Data Integration: Collect structured financial data, supply chain relationships, and unstructured enterprise-related information data from both inside and outside the enterprise, classify and store them in a distributed database, and perform standardization and alignment. Step S2, Spatiotemporal Graph Modeling: Construct a dynamic enterprise relationship graph, define node attributes and edge weights, and generate a continuous time series graph sequence by quarter; Step S3, Feature Fusion: Temporal features are extracted using bidirectional LSTM, supply chain graph features are extracted using GraphSAGE, and market features are extracted using CNN. The three types of modal features are then adaptively weighted and fused based on dynamic modality fusion weights. Step S4, Risk Prediction: Design a two-layer attention mechanism that combines historical time series characteristics with supply chain correlation to output the risk probability value of the enterprise in the next quarter; Step S5, Early Warning Output: Trigger an early warning based on industry dynamic thresholds and automatically generate a report.

2. The AI-based automated enterprise financial analysis method according to claim 1, characterized in that, Step S1, data integration, specifically includes the following steps: Step S11: Define the data source, specifically as follows: Internal structured data: Extract financial statements, trial balances, and accounts receivable and payable from the enterprise ERP system; External structured data: crawling public market data and publicly available financial indicators of upstream and downstream enterprises in the supply chain; Unstructured data: Obtaining news company-related information, CSRC penalty announcements, and company-related discussions on social media through NLP interfaces; Step S12: Establish a distributed database and perform partitioned storage alignment according to the following tags: Financial metrics: Selected key metrics, aligned by quarterly timestamps; Supply chain map: with enterprises as nodes, edge attributes include transaction amount, payment period, and historical number of defaults; Market signals: industry beta coefficient, 30-day volatility, short selling ratio; Company-related information tags: emotional polarity, keyword coverage.

3. The AI-based automated enterprise financial analysis method according to claim 1, characterized in that, Step S2, spatiotemporal graph modeling, specifically includes the following steps: Step S21: Define node attributes, for each enterprise node v i attribute vector Include: In the formula, For enterprise v i In the node attribute vector at time t, t is the time identifier, representing the current quarter; ROE is the return on net assets; and the public opinion score is a numerical score obtained through NLP sentiment analysis. The dimension is 32, after dimensionality reduction by PCA; Step S22: Construct edge relationships, when enterprise v i With v j If supply chain transactions exist, then a directed edge e is constructed. ij Weight w ij Calculation formula: In the formula, w ij For enterprise v i With v j The edge weights between them, and the total annual transaction amount for enterprise v i With v j Annual transaction amount, historical number of deferred payments for enterprise v i To v j The historical number of delayed payments; the total number of transactions is the total number of historical transactions between the two parties. Step S23: Dynamic time series slicing, generating time series snapshots G by quarter. (t) ={V (t) E (t) ,X (t) }, retaining data from eight consecutive quarters to form a graph sequence {G (t-7) ,…,G (t) }, where G (t) For a snapshot of the time series at time t, V (t) Let E be the set of nodes at time t. (t) Let X be the set of edges at time t. (t) Let be the node attribute matrix at time t.

4. The AI-based automated enterprise financial analysis method according to claim 1, characterized in that, Step S3, feature fusion, specifically includes the following steps: Step S31: Extract time-series features and process the financial indicator sequence of each enterprise using a bidirectional LSTM, specifically: In the formula, Forward LSTM to v i The coding results of financial indicators over the past 8 quarters, This is the reverse encoding result of the same data using an inverse LSTM. To create a vector that incorporates bidirectional temporal features, This is a vector concatenation operation; Step S32: Extract graph structure features and aggregate neighbor information using the GraphSAGE algorithm, specifically: In the formula, For enterprise node v i The neighbor aggregation feature, where N(i) represents the number of enterprises v. i The set of neighboring nodes, W agg A trainable weight matrix aggregated from neighbor features. For enterprise node v i The final graph structure features, W comb It is a trainable weight matrix that combines its own features with the features of its neighbors; Step S33: Modal fusion, calculate the modal weights α according to the formula. time ,α graph ,α market The specific formula is as follows: In the formula, W represents the fusion weights of the m-th mode at time t. a For trainable weight matrix, Let g be the eigenvector of the m-th mode at time t. (t) For global temporal states, σ is the LeakyReLU activation function, and M is the total number of modes; Regenerated fusion features: In the formula, For enterprise v i The fused feature vector at time t, α is the vector extracted from market data features using a CNN. time ,α graph ,α market These are the fusion weights for time series, graph structure, and market modality, respectively.

5. The AI-based automated enterprise financial analysis method according to claim 1, characterized in that, Step S4, risk prediction, specifically includes the following steps: Step S41: Design a two-layer attention mechanism, specifically as follows: Time-attention layer: Calculates the contribution of each historical quarter to the current forecast. In the formula, β (k) Let q be the time attention weight for the prediction in the k-th historical quarter, q be the trainable query vector, T be the transpose operator, and W be the time attention weight for the prediction. t This is the time attention weight matrix. For enterprise v i The fused feature vector at time tk; Cross-enterprise attention layer: Calculating the impact weights of supply chain related enterprises: In the formula, γ ij For supply chain related enterprises v j For target company v i Influence weight, W r A trainable weight matrix for modeling cross-enterprise relationships; Step S42: Predict the output, and input the weighted features into the fully connected layer: In the formula, For enterprise v i In the risk probability score for the next quarter, W p Let be the weight matrix of the fully connected layer, b be the bias term, and τ be the Sigmoid function.

6. The AI-based automated enterprise financial analysis method according to claim 1, characterized in that, Step S5, the early warning output, specifically includes the following steps: Step S51, Threshold Determination: Dynamic threshold adjustment based on the industry risk median μ and standard deviation. Step S52: Automatic Report Generation: Key Factor Attribution, Extraction of Attention Weights and β (k) Generate explanatory text and charts.

7. An AI-based automated enterprise financial analysis system, characterized in that, The analysis system includes a data acquisition module, a spatiotemporal graph modeling module, a feature fusion module, a risk prediction module, and an early warning output module; The data acquisition module, spatiotemporal graph modeling module, feature fusion module, risk prediction module, and early warning output module are connected in sequence.

8. The AI-based automated enterprise financial analysis system according to claim 7, characterized in that, The data acquisition module is used to collect structured and unstructured data from multiple channels, and to complete data cleaning, alignment, and classification storage. The spatiotemporal graph modeling module is used to construct a dynamic enterprise relationship graph, define node attributes and edge weights, and generate a time sequence graph sequence to represent the dynamic relationship between enterprises. The feature fusion module is used to extract financial time-series features, supply chain graph structure features, and market enterprise-related information features respectively, and to achieve adaptive weighted fusion of cross-modal features through dynamic modal fusion weights; The risk prediction module captures key risk signals based on a hierarchical attention mechanism and outputs the probability value of financial risk in the next quarter through a fully connected network. The early warning output module is used to trigger risk warnings based on dynamic thresholds and automatically generate standardized reports that include attribution analysis, visualization charts, and response suggestions.

Images