Enterprise financial analysis method and system

Abstract

The invention relates to an enterprise financial analysis method and system. The enterprise financial analysis method comprises the steps of obtaining structured financial data and unstructured market data of an enterprise; preprocessing the structured financial data and the unstructured market data; generating a joint embedded representation by using the pre-processed structured financial data and the pre-processed non-structured market data; constructing an enterprise relation graph, and generating a dynamic enterprise relation graph based on the enterprise relation graph; obtaining enterprise financial data; an LSTM-TCN mixed time sequence model is constructed; performing pre-training on the LSTM-TCN mixed time sequence model by using the macroeconomic fluctuation data; performing fine adjustment on the pre-trained LSTM-TCN mixed time sequence model by using enterprise financial data; and inputting the current enterprise financial data into the fine-tuned LSTM-TCN mixed time sequence model for risk assessment. According to the method and the device, enterprise risks caused by factors such as unstructured market data can be found in time.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to an enterprise financial analysis method and system. Background Art

[0002] Financial analysis is crucial for an enterprise to conduct risk control and achieve profitability. However, there are still problems with data dimension segmentation in current financial analysis: existing financial analysis systems (such as SAP Financials and Oracle EPM) mainly rely on structured financial statement data (such as balance sheets, cash flow statements, etc.), and have weak capabilities for integrating and processing unstructured market data (such as policies and regulations, etc.); as a result, enterprise risks caused by factors such as unstructured market data cannot be discovered in a timely manner. Summary of the Invention

[0003] Based on this, it is necessary to provide an enterprise financial analysis method and system for the problems in the related technologies.

[0004] To achieve the above object, in the first aspect, this application provides an enterprise financial analysis method, and the enterprise financial analysis method includes: Obtain the structured financial data and unstructured market data of the enterprise; Preprocess the obtained structured financial data and unstructured market data; Use the multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data; Construct an enterprise relationship graph, and generate a dynamic enterprise relationship graph based on the enterprise relationship graph; Fuse the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data; Construct an LSTM-TCN hybrid time series model; Obtain macroeconomic fluctuation data; use the macroeconomic fluctuation data to pre-train the LSTM-TCN hybrid time series model; Through a domain adaptation layer, use the enterprise financial data to fine-tune the pre-trained LSTM-TCN hybrid time series model; Input the current enterprise financial data into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain a risk assessment result.

[0005] In some of the embodiments, the formula for using the multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data is as follows: where Q is the query matrix, K is the key matrix, KT represents the transpose of matrix K, represents the matrix multiplication of the query matrix Q and the transpose of the key matrix K to calculate the similarity between the query and the key, V is the value matrix, d k is the dimension of the key vector.

[0006] In some embodiments, constructing an enterprise relationship graph and generating a dynamic enterprise relationship map based on the enterprise relationship graph includes: Taking enterprises as nodes and the business relationships between enterprises as edges to construct the enterprise relationship graph; Using CNN for graph convolution operations to capture the association information between enterprises to generate the dynamic enterprise relationship map; the formula for the graph convolution operation is as follows: where, is the adjacency matrix with self-loops, is the degree matrix, is the node feature matrix of the l-th layer, is the node feature matrix of the (l + 1)-th layer, is the learnable weight matrix, is the activation function.

[0007] In some embodiments, after obtaining the risk assessment result, it further includes: Based on the structured financial data of the enterprise, the risk assessment result and the market environment data, obtaining a multi-stage profit optimization strategy for the enterprise, including: Collecting the structured financial data of the enterprise, the risk assessment result and the market environment data; According to the Granger causality test and the Bayesian network, obtaining the causal relationship between the risk control measures and the profit indicators; Based on the causal relationship, obtaining an initial multi-stage profit optimization strategy; Using a reinforcement learning algorithm to optimize the initial multi-stage profit optimization strategy to obtain a multi-stage profit optimization strategy.

[0008] In some embodiments, using a profit correlation inference engine model, based on the structured financial data of the enterprise, the risk assessment result and the market environment data, obtaining a multi-stage profit optimization strategy for the enterprise; after obtaining the multi-stage profit optimization strategy for the enterprise, it further includes: Based on an online incremental learning mechanism, regularly updating the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model; Use SHAP values for analysis to evaluate the contribution of each feature to the LSTM-TCN hybrid time series model and the profit correlation inference engine model; According to the analysis results of SHAP values, optimize the feature engineering and update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model.

[0009] In some embodiments, it further includes: Deploy a load balancer to dynamically allocate computing resources according to the load; During the analysis process, adopt a fault tolerance mechanism to ensure the continuity of the analysis process; Collect user feedback information; Based on the user feedback information, perform iterative improvement on at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model.

[0010] In a second aspect, the present application also provides an enterprise financial analysis system, and the enterprise financial analysis system includes: A data acquisition module for acquiring the structured financial data and unstructured market data of an enterprise; A heterogeneous data fusion layer for preprocessing the acquired structured financial data and unstructured market data; using a multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data; constructing an enterprise relationship graph, generating a dynamic enterprise relationship graph based on the enterprise relationship graph; fusing the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data; A dynamic risk assessment module for constructing an LSTM-TCN hybrid time series model; acquiring macroeconomic fluctuation data; using the macroeconomic fluctuation data to pre-train the LSTM-TCN hybrid time series model; through a domain adaptation layer, using the enterprise financial data to fine-tune the pre-trained LSTM-TCN hybrid time series model; inputting the current enterprise financial data into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain a risk assessment result.

[0011] In some embodiments, the enterprise financial analysis system further includes: A profit correlation inference engine for obtaining a multi-stage profit optimization strategy of an enterprise based on the structured financial data of the enterprise, the risk assessment result, and market environment data.

[0012] In some of these embodiments, the profit-related inference engine is further configured to construct a profit-related inference engine model, which obtains a multi-stage profit optimization strategy for the enterprise based on the enterprise's structured financial data, the risk assessment results, and market environment data; the enterprise financial analysis system further includes: A dynamic knowledge iteration module, configured to periodically update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit-related inference engine model based on an online incremental learning mechanism; perform analysis using SHAP values to evaluate the contribution of each feature to the LSTM-TCN hybrid time series model and the profit-related inference engine model; optimize the feature engineering according to the analysis results of the SHAP values, and update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit-related inference engine model.

[0013] In some of these embodiments, the enterprise financial analysis system further includes: a system integration and optimization module, and the system integration and optimization module includes: Multiple standardized interfaces, located on each module, for ensuring smooth data exchange and communication between modules; A data flow management unit, for ensuring smooth data flow between modules; A load balancer, for dynamically allocating computing resources according to the load; A fault tolerance mechanism module, for adopting a fault tolerance mechanism to ensure that the enterprise financial system can still operate normally when some modules fail; A user feedback information collection module, for collecting user feedback information; An iterative improvement module, for iteratively improving at least one of the models including the LSTM-TCN hybrid time series model and the profit-related inference engine model based on user feedback information.

[0014] In the above enterprise financial analysis method and system, after obtaining and preprocessing the structured financial data and unstructured market data of the enterprise, a multi-head self-attention mechanism is used to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data, and the joint embedding representation is fused with the dynamic enterprise relationship graph to obtain enterprise financial data. Through the domain adaptation layer, the enterprise financial data is used to fine-tune the pre-trained LSTM-TCN hybrid time series model for macroeconomic fluctuation data. Finally, the current enterprise financial data is input into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain a risk assessment result. The enterprise financial data of this application can comprehensively reflect the financial status and market environment of the enterprise, break through the bottleneck of traditional single-modal data processing, can timely discover enterprise risks caused by factors such as unstructured market data, and can provide timely and accurate risk assessment and early warning for the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0016] Figure 1 It is a flowchart of the enterprise financial analysis method provided in an embodiment of the present application; Figure 2 and Figure 3 It is a structural block diagram of the enterprise financial analysis system provided in different embodiments of the present application.

[0017] Description of the reference numerals: 10, data acquisition module; 20, heterogeneous data fusion layer; 30, dynamic risk assessment module; 40, profit correlation inference engine. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will further describe the present application in detail with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0019] Financial analysis is crucial for enterprises to conduct risk control and achieve profitability. Some financial analysis systems (such as SAP Financials and Oracle EPM) have the problem of fragmented data dimensions, mainly relying on structured financial statement data (such as balance sheets, cash flow statements, etc.), and having weak capabilities in integrating and processing unstructured market data (such as policies and regulations, etc.); thus, they cannot timely detect enterprise risks caused by factors such as unstructured market data. Some other financial analysis solutions regard the risk identification and profitability analysis of enterprises as two independent versions, with no relevant connection between the two; this will result in the inability to discover the strong correlation between enterprise risks and profitability, and this financial analysis solution has the problem of missing risk-profit association. Some other financial analysis solutions adopt machine learning methods (logistic regression or random forest). The above machine learning methods use fixed-time slice modeling, unable to capture dynamic risk evolution and unable to identify the risk conduction lag effect caused by the end of the supply chain. This financial analysis solution has the defect of static modeling.

[0020] In one embodiment, please refer to Figure 1 , this application provides an enterprise financial analysis method, and the enterprise financial analysis method includes the following steps: S10~S90.

[0021] S10: Obtain the structured financial data and unstructured market data of the enterprise.

[0022] S20: Preprocess the obtained structured financial data and unstructured market data.

[0023] S30: Use the multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data.

[0024] S40: Construct an enterprise relationship graph and generate a dynamic enterprise relationship graph based on the enterprise relationship graph.

[0025] S50: Integrate the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data.

[0026] S60: Construct an LSTM-TCN hybrid time series model.

[0027] S70: Obtain macroeconomic fluctuation data; use the macroeconomic fluctuation data to pre-train the LSTM-TCN hybrid time series model.

[0028] S80: Through the domain adaptation layer, use the enterprise financial data to fine-tune the pre-trained LSTM-TCN hybrid time series model.

[0029] S90: Input the current enterprise financial data into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain the risk assessment result.

[0030] In the enterprise financial analysis method of this application, after obtaining the structured financial data and unstructured market data of the enterprise and preprocessing them, the multi-head self-attention mechanism is used to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data, and the joint embedding representation and the dynamic enterprise relationship graph are fused to obtain enterprise financial data. Through the domain adaptation layer, the enterprise financial data is used to fine-tune the pre-trained LSTM-TCN hybrid time series model for macroeconomic fluctuation data. Finally, the current enterprise financial data is input into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain the risk assessment result. The enterprise financial data of this application can comprehensively reflect the financial status and market environment of the enterprise, break through the bottleneck of traditional single-modal data processing, can timely discover enterprise risks caused by factors such as unstructured market data, and can provide timely and accurate risk assessment and early warning for the enterprise.

[0031] In step S10, refer to step S10 in Figure 1 to obtain the structured financial data and unstructured market data of the enterprise.

[0032] As an example, the structured financial data of the enterprise (such as balance sheet, cash flow statement, etc.) and unstructured market data can be collected. The unstructured market data can be unstructured text data, such as industry research reports, policies and regulations, etc.

[0033] As an example, while collecting the structured financial data and unstructured market data of the enterprise, the associated data of the upstream and downstream enterprises of the enterprise can also be collected, including but not limited to supplier and customer information, etc.

[0034] In step S20, refer to Figure 1 step S20 in Figure 1 to preprocess the obtained structured financial data and unstructured market data.

[0035] As an example, the structured financial data can be standardized to ensure the consistency of the dimensions of different financial indicators. For unstructured market data, which is unstructured text data, pre-trained language models such as BERT can be used for text embedding to convert the text into a high-dimensional vector representation.

[0036] In step S30, refer to Figure 1 step S30 in Figure 1 to use the multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data.

[0037] As an example, a cross-modal attention mechanism, which can be but is not limited to the Transformer architecture, can be used to jointly embed the processed structured financial data and unstructured market data. The core of the Transformer architecture is the multi-head self-attention mechanism, which can capture the correlation between different modal data.

[0038] As an example, the vector representations of the preprocessed structured financial data and the preprocessed unstructured market data can be input into the Transformer model. The Transformer model calculates the correlation weights between different modal data through the multi-head self-attention mechanism and generates a joint embedding representation. The specific formula can be as follows: where Q is the query matrix, K is the key matrix, and K T represents the transpose of matrix K, represents the matrix multiplication of the query matrix Q and the transpose of the key matrix K to calculate the similarity between the query and the key, V is the value matrix, and d k is the dimension of the key vector. Through the multi-head self-attention mechanism, the model can capture the complex correlations between different modal data.

[0039] In step S40, please refer to Figure 1 step S40 in

[0040] As an example, step S40 may include the following steps: S401~S402.

[0041] S401: Use enterprises as nodes and the business relationships between enterprises as edges to construct the enterprise relationship graph.

[0042] S402: Use CNN for graph convolution operations to capture the correlation information between enterprises to generate the dynamic enterprise relationship graph. The formula for the graph convolution operation is as follows: where is the adjacency matrix with self-loops, is the degree matrix, is the node feature matrix of the l-th layer, is the node feature matrix of the (l + 1)-th layer, is the learnable weight matrix, is the activation function.

[0043] As an example, in step S401, the business relationships between enterprises can include supplier and customer relationships, etc.

[0044] As an example, in step S401, in the constructed enterprise relationship graph, each edge can carry a weight, indicating the strength of the relationship or the business volume.

[0045] As an example, in step S402, deploy a graph neural network (GNN) to process the associated data of upstream and downstream enterprises and construct a dynamic enterprise relationship graph; the GNN can capture the complex relationships between enterprises and reflect the positions and correlations of enterprises in the industrial chain.

[0046] As an example, in step S402, use multi-layer graph convolution so that the GNN can capture the high-order relationships between enterprises.

[0047] In step S50, please refer to Figure 1 step S50 in, and fuse the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data.

[0048] As an example, steps S20 to S50 can be implemented by constructing a heterogeneous data fusion layer. The heterogeneous data fusion layer uses the cross-modal attention mechanism of the Transformer architecture to jointly embed structured financial data and unstructured market data, and deploys a graph neural network (GNN) to process the associated data of upstream and downstream enterprises and construct a dynamic enterprise relationship graph, aiming to solve the problem of data dimension fragmentation in enterprise financial analysis. Through the deep fusion of multi-modal data, the financial situation and market environment of enterprises can be comprehensively reflected. Through the above method, a dynamic enterprise relationship graph is constructed to reflect the positions and correlations of enterprises in the industrial chain in real time. The dynamic enterprise relationship graph can be used to analyze the supply chain risks and market competitiveness of enterprises, etc.

[0049] As an example, in step S50, fuse the joint embedding representation and the dynamic enterprise relationship graph, and the obtained enterprise financial data is used as the output of the heterogeneous data fusion layer for subsequent financial risk identification and profit analysis.

[0050] In a specific example, taking a manufacturing enterprise as an example, the relevant implementation steps of the heterogeneous data fusion layer can include data preparation, data preprocessing, joint embedding, graph structure construction, graph convolution operation, and data fusion.

[0051] Data preparation: Collect the structured financial data of the enterprise (such as balance sheets, cash flow statements, etc.) and unstructured text data (such as industry research reports, policies and regulations, etc.); at the same time, collect the associated data of the upstream and downstream enterprises of the enterprise, including supplier and customer information, etc.

[0052] Data preprocessing: Standardize the structured financial data to ensure the consistency of the dimensions of different financial indicators; for the unstructured text data, use the BERT model for text embedding to convert the text into a high-dimensional vector representation.

[0053] Joint embedding: Input the preprocessed structured financial data and unstructured text data into the Transformer model, and generate a joint embedding representation through the multi-head self-attention mechanism.

[0054] Graph structure construction: Take the enterprise and its upstream and downstream enterprises as nodes and the business relationships as edges to construct an enterprise relationship graph.

[0055] Graph convolution operation: Use the GNN for graph convolution operations to capture the association information between enterprises and generate a dynamic enterprise relationship graph.

[0056] Data fusion: Fuse the joint embedding representation and the dynamic enterprise relationship graph to generate the output of the heterogeneous data fusion layer for subsequent financial risk identification and profitability analysis.

[0057] Through the above steps, the heterogeneous data fusion layer can comprehensively reflect the financial status and market environment of the enterprise, providing a rich information basis for subsequent financial risk identification and profitability analysis.

[0058] In step S60, please refer to Figure 1 step S60 in

[0059] As an example, LSTM is a special type of recurrent neural network (RNN) that can capture long-term dependencies in time series. LSTM effectively solves the gradient vanishing problem of traditional RNNs by introducing memory cells and gating mechanisms (input gate, forget gate, output gate). The update formula of LSTM can be as follows: Among them, is the activation value of the forget gate, is the activation value of the input gate, is the activation value of the output gate, is the state of the memory cell, is the updated state of the memory cell, h t is the hidden state at the current time step, ht-1 is the hidden state of the previous time step, W f is the weight matrix of the forget gate, W i is the weight matrix of the input gate, W c is the weight matrix of the cell state, W o is the weight matrix of the output gate, b f is the bias vector of the forget gate, b i is the bias vector of the input gate, b c is the bias vector of the cell state, b o is the bias vector of the output gate, x t is the input vector of the current time step, is the sigmoid activation function, tanh is the hyperbolic tangent function.

[0060] As an example, TCN is a time series model based on convolutional neural network that can capture dependencies at different time scales through dilated convolution. The formula of dilated convolution of TCN is as follows: Among them, is the t-th element of the output sequence, is the weight of the convolutional kernel, d is the dilation factor, K is the size of the convolutional kernel, x t-d is the input at time step t - d; by adjusting the dilation factor, TCN can capture dependencies at different time scales.

[0061] As an example, in step S60, LSTM and TCN are combined. LSTM is used to capture long-term dependencies, and TCN is used to capture short-term dependencies at different time scales; through this LSTM-TCN hybrid time series model, the dynamic evolution law of risks can be captured more comprehensively.

[0062] In step S70, please refer to Figure 1 in step S70 to obtain macroeconomic volatility data; use the macroeconomic volatility data to pre-train the LSTM-TCN hybrid time series model.

[0063] As an example, macroeconomic volatility data (such as GDP growth rate, inflation rate, and interest rate, etc.) can be used to pre-train the risk conduction model; macroeconomic volatility data can reflect the changes in the overall economic environment, and the pre-trained LSTM-TCN hybrid time series model can capture the conduction law of risks in different economic environments.

[0064] In step S80, please refer to Figure 1In step S80, through the domain adaptation layer, the pre-trained LSTM-TCN hybrid time series model is fine-tuned using the enterprise financial data.

[0065] As an example, enterprise-level fine-tuning can be performed through the Domain Adaptation layer. The goal of the domain adaptation layer is to minimize the distribution difference between the source domain (macroeconomic data) and the target domain (enterprise financial data), thereby improving the adaptability of the model to different enterprises. The loss function of the domain adaptation layer is as follows: Among them, is the task loss (such as the mean squared error of risk prediction), is the domain difference loss (such as the maximum mean difference, MMD), and λ is the trade-off parameter. By minimizing the domain adaptation loss, the model can better adapt to enterprise-level data.

[0066] In step S90, refer to Figure 1 In step S90, the current enterprise financial data is input into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain the risk assessment result.

[0067] As an example, step S90 can include two parts: the risk conduction mode and the real-time risk assessment.

[0068] Risk conduction mode: Through the fine-tuned LSTM-TCN hybrid time series model, capture the risk conduction mode at different time scales. The risk conduction mode can reflect the propagation path of risks within and outside the enterprise, helping the enterprise identify potential risks in advance.

[0069] Real-time risk assessment: Through the LSTM-TCN hybrid time series model fine-tuned at the enterprise level, the financial risks of the enterprise can be evaluated in real time. The fine-tuned LSTM-TCN hybrid time series model can dynamically adjust the risk assessment result according to the latest financial data and market environment, providing timely risk warnings for the enterprise.

[0070] As an example, steps S60 to S90 can be implemented through the dynamic risk assessment module; in the dynamic risk assessment module, the hybrid time series model integrating LSTM and TCN captures the risk evolution law at different time scales, and through creating a new transfer learning framework (including the content related to steps S70 and S80), uses the macroeconomic fluctuation data to pre-train the risk conduction mode and performs enterprise-level fine-tuning.

[0071] In a specific example, taking a manufacturing enterprise as an example, the steps corresponding to the dynamic risk assessment module can include: data preparation, pre-training stage, fine-tuning stage, real-time risk assessment, and risk warning.

[0072] Data Preparation: Collect macroeconomic volatility data (such as GDP growth rate, inflation rate, interest rate, etc.) and the financial data of the enterprise (such as balance sheet and cash flow statement, etc.).

[0073] Pre-training Stage: Use macroeconomic volatility data to pre-train the LSTM-TCN hybrid model to capture the conduction patterns of risks in different economic environments.

[0074] Fine-tuning Stage: Through the domain adaptation layer, use the financial data of the enterprise to fine-tune the trained model to minimize the distribution difference between the source domain and the target domain.

[0075] Real-time Risk Assessment: Input the latest financial data into the fine-tuned model to assess the financial risks of the enterprise in real time and generate a risk assessment report.

[0076] Risk Warning: According to the risk assessment results, provide risk warnings to the enterprise in a timely manner to help the enterprise take countermeasures.

[0077] Through the above steps, the dynamic risk assessment module can comprehensively capture the dynamic evolution law of risks and provide timely and accurate risk assessment and warning for the enterprise.

[0078] As an example, after step S90, the following steps may also be included: S100: Obtain the multi-stage profit optimization strategy of the enterprise based on the structured financial data of the enterprise, the risk assessment results, and the market environment data.

[0079] As an example, obtaining the multi-stage profit optimization strategy of the enterprise based on the structured financial data of the enterprise, the risk assessment results, and the market environment data may include the following steps: S1011~S1004.

[0080] S1001: Collect the structured financial data, the risk assessment results, and the market environment data of the enterprise.

[0081] S1002: Obtain the causal relationship between the risk control measures and the profit indicators according to the Granger causality test and the Bayesian network.

[0082] S1003: Obtain the initial multi-stage profit optimization strategy based on the causal relationship.

[0083] S1004: Use the reinforcement learning algorithm to optimize the initial multi-stage profit optimization strategy to obtain the multi-stage profit optimization strategy.

[0084] As an example, in step S1001, the structured financial data of the enterprise collected may include but is not limited to sales, profit margins, etc.; the risk assessment results may include but are not limited to risk data, such as inventory risk, supply chain risk, etc.; the market environment data may include but is not limited to consumer information index, industry growth rate, etc.

[0085] As an example, in step S1002, the Granger causality test is a statistical method used to test the causal relationship between variables in time series data; its basic idea is that if the past values of variable X can significantly predict the current value of variable Y, then X is called the Granger cause of Y. The formula for the Granger causality test can be as follows: Where, is the current value of the target variable, is the past value of the target variable, is the past value of the explanatory variable, is the constant term, and are the regression coefficients, p is the lag order, is the error term. By testing whether it is significantly non-zero, it can be judged whether X is the Granger cause of Y.

[0086] As an example, the Bayesian network is a causal inference method based on a probabilistic graphical model, which can represent the causal relationship between variables through a directed acyclic graph (DAG). The joint probability distribution of the Bayesian network can be expressed as: Where, is the node in the network, is the parent node of, n is the total number of nodes in the network. Through the Bayesian network, the causal relationship between risk control measures and profit indicators can be inferred.

[0087] As an example, step S1001 and step S1002 can be executed through a dual-channel causal discovery model.

[0088] As an example, in step S1003, an initial multi-stage profit optimization strategy can be generated according to the causal relationship discovered in step S1002. For example, when the risk is low, increase investment to improve profit; when the risk is high, reduce investment to control risk.

[0089] As an example, in step S1004, it may include the following related content: state space, action space, reward function, and policy optimization.

[0090] State space: The state space S consists of variables such as the financial indicators, risk indicators, and market environment of the enterprise, representing the state of the enterprise at a certain moment.

[0091] Action space: The action space A consists of risk control measures and profit strategies that the enterprise can take, such as adjusting the investment portfolio, optimizing the supply chain, and so on.

[0092] Reward function: The reward function R is used to evaluate the profit effect of the enterprise after taking a certain action, and can be expressed as: where, is the profit of the enterprise after taking action a in state s, is the corresponding risk, and λ is the risk aversion coefficient.

[0093] Policy optimization: Through reinforcement learning algorithms (such as Q-learning or Deep Q-Network, DQN), optimize the multi-stage profit optimization policy of the enterprise. The update formula of Q-learning is as follows: where, is the Q value of collecting action a in state s, is the learning rate, γ is the discount factor, is the next state, is the next action; by continuously updating the Q value, the reinforcement learning algorithm can find the optimal multi-stage profit optimization policy.

[0094] As an example, step S100 can be executed by a profit correlation inference engine, which constructs a two-channel causal discovery model, combines Granger causality test and Bayesian network to analyze the dynamic correlation between risk control measures and profit indicators, and designs a reinforcement learning decision-making module to generate a multi-stage profit optimization policy.

[0095] In a specific example, taking a retail enterprise as an example, the steps corresponding to the profit correlation inference engine can be as follows: data preparation, causal discovery, policy generation, policy optimization, and policy implementation.

[0096] Data preparation: Collect the financial data (such as sales volume, profit margin, etc.), risk data (such as inventory risk, supply chain risk, etc.), and market environment data (such as consumer confidence index, industry growth rate, etc.) of the enterprise. It should be noted that the data obtained in the previous steps can be directly used in this step without repeated data preparation.

[0097] Causal discovery: Through Granger causality test and Bayesian network, it is found that reducing inventory risk can significantly improve the profitability of the enterprise.

[0098] Strategy generation: Based on the discovered causal relationships, generate a preliminary multi-stage profit optimization strategy. For example, when the inventory risk is low, increase the inventory to boost sales; when the inventory risk is high, reduce the inventory to control risks.

[0099] Strategy optimization: Through reinforcement learning algorithms, optimize the enterprise's profit strategy to obtain a multi-stage profit optimization strategy. For example, dynamically adjust the inventory level and investment strategy according to changes in the market environment to maximize profits.

[0100] Strategy implementation: Apply the optimized multi-stage profit optimization strategy to the actual operation of the enterprise, monitor the changes in profits and risks in real time, and make adjustments as needed.

[0101] Through the above steps, the profit-correlation inference engine can deeply analyze the dynamic parallel relationship between risk control measures and profit indicators, and generate a multi-stage profit optimization strategy to help the enterprise achieve coordinated management of risks and profits.

[0102] As an example, using the profit-correlation inference engine model, based on the enterprise's structured financial data, the risk assessment results, and market environment data, obtain the enterprise's multi-stage profit optimization strategy; after obtaining the enterprise's multi-stage profit optimization strategy, that is, after step S100, it may further include: steps S101 to S103.

[0103] S101: Regularly update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit-correlation inference engine model based on an online incremental learning mechanism.

[0104] S102: Use SHAP values for analysis to evaluate the contribution of each feature to the LSTM-TCN hybrid time series model and the profit-correlation inference engine model.

[0105] S103: According to the analysis results of the SHAP values, optimize the feature engineering and update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit-correlation inference engine model.

[0106] As an example, in step S101, it may include two parts: sliding time window update and incremental learning algorithm.

[0107] Sliding time window update: Adopt a sliding time window mechanism to regularly update the model parameters (which can include the parameters of the LSTM-TCN hybrid time series model and the profit correlation inference engine model, and of course can also be the parameters of other models used in the above steps) to absorb the latest market changes. The size W of the sliding time window can be adjusted according to the data frequency and business requirements. For each time window t, the parameters of the model can be updated through the following formula: where, is the learning rate, are the parameters before model update, is the gradient operator, is the loss function within time window t, is the data within time window t. In this way, the model can adapt to market changes in real time and maintain sensitivity to the latest data.

[0108] Incremental learning algorithm: Adopt an incremental learning algorithm (such as Online Gradient Descent) to update the model parameters. The core idea of the incremental learning algorithm is to update the model parameters using only one data point or a small batch of data each time, thereby reducing the computational overhead and improving the real-time performance of the model.

[0109] As an example, in step S102, use SHAP (Shapley Additive exPlanations) value analysis to evaluate the contribution of each feature to the model prediction. The SHAP value represents the contribution of feature i to the model prediction and can be obtained through the following formula: where, is the SHAP value of feature i, F is the set of all features, S is the subset of features that does not include feature i, is the model prediction value applicable only to the feature subset S, is the model prediction value applicable to the feature subset S plus feature i.

[0110] Through SHAP value analysis, the features that have the greatest impact on the model prediction can be identified, thereby finding the optimization of feature engineering.

[0111] As an example, in step S103, optimize the feature engineering according to the results of SHAP value analysis. For example, for features with large contributions, their representations can be further refined; for features with small contributions, they can be considered for removal or merging. In this way, the feature representation ability of the model can be improved, thereby improving the prediction performance of the model.

[0112] As an example, in step S103, the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model are also updated through an online incremental learning mechanism.

[0113] Specifically, steps S102 and S103 embody the model self-evolution ability, which specifically includes feedback loop construction. By constructing a closed-loop feedback loop, the results of SHAP value analysis are fed back into feature engineering and model parameter update. Specifically, first, the features that have the greatest impact on model prediction are identified through SHAP value analysis, then feature engineering optimization is performed based on these features, and finally, the model parameters are updated through an online incremental learning mechanism. Through the above method, the model can continuously optimize itself and form the self-evolution ability.

[0114] As an example, the enterprise financial analysis method of the present application may further include the step of regularly monitoring the performance indicators of each model. Through this step, performance indicators such as accuracy and recall rate of the model can be regularly monitored, and the model parameters and feature engineering can be adjusted according to the performance changes. For example, if the accuracy of the model decreases, the performance of the model can be improved by increasing the complexity of feature engineering or adjusting the learning rate.

[0115] As an example, the content of the above steps S101 to S103 can be executed based on a dynamic knowledge iteration module. The dynamic knowledge iteration module can deploy an online incremental learning mechanism, update the model parameters through a sliding time window, and construct an interpretable feedback loop to drive feature engineering optimization using SHAP value analysis, forming the model self-evolution ability.

[0116] In a specific example, taking a retail enterprise as an example, the steps involved in the dynamic knowledge iteration module can be as follows: data preparation, sliding time window update, SHAP value analysis, feature engineering optimization, model performance monitoring, and feedback loop construction.

[0117] Data preparation: Collect the financial data (such as sales volume, profit margin, etc.), risk data (such as inventory risk, supply chain risk, etc.), and market environment data (such as consumer information index, industry growth rate, etc.) of the enterprise. It should be noted that the data obtained in the previous steps can be directly used in this step without repeated data preparation.

[0118] Sliding time window update: Set the size W of the sliding time window to one month, and update the model parameters once a month. For each time window t, update the model parameters through the online gradient descent algorithm .

[0119] SHAP Value Analysis: Use SHAP value analysis to evaluate the contribution of each feature to the model prediction. For example, it is found that the consumer confidence index has the greatest contribution to the sales prediction.

[0120] Feature Engineering Optimization: Optimize the feature engineering according to the results of SHAP value analysis. For example, increase the representation dimension of the consumer confidence index or remove features with less contribution to the prediction.

[0121] Model Performance Monitoring: Regularly monitor the performance metrics of the model, such as accuracy and recall. If the accuracy of the model decreases, the model performance can be improved by increasing the complexity of the feature engineering or adjusting the learning rate.

[0122] Feedback Loop Construction: Feed the results of SHAP value analysis back into the feature engineering and model parameter updates to form a closed-loop feedback loop. In this way, the model can continuously optimize itself and form the ability of self-evolution.

[0123] Through the above steps, the dynamic knowledge iteration module can achieve the continuous optimization and self-evolution of the model through online incremental learning and interpretable feedback loops, thereby improving the prediction performance and real-time performance of the model.

[0124] As an example, the enterprise financial analysis method of this application may further include the following steps: S104~S107.

[0125] S104: Deploy a load balancer to dynamically allocate computing resources according to the load.

[0126] S105: During the analysis process, adopt a fault tolerance mechanism to ensure the continuity of the analysis process.

[0127] S106: Collect the feedback information of users.

[0128] S107: Iteratively improve at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model based on the feedback information of users.

[0129] As an example, before step S104, it further includes the steps of module interface design and data flow management.

[0130] As an example, in the module interface design step, design a standardized interface for each module to ensure smooth data exchange and communication between modules. For example, the output data of the heterogeneous data fusion layer should be able to be seamlessly transmitted to the dynamic risk assessment module, and the results of the dynamic risk assessment module can be used by the profit correlation inference engine. The interface design can adopt technologies such as RESTful API or message queues (such as Kafka) to ensure the efficiency and reliability of data transmission.

[0131] As an example, in the data flow management step, a unified data flow management module can be built to ensure smooth data flow between various modules. The data flow management module uses ETL (Extract, Transform, Load) technology to extract, transform, and load data to ensure data consistency and integrity. At the same time, data version control technology is used to ensure the traceability of historical data versions, which is convenient for the system (enterprise financial analysis system) to roll back and troubleshoot faults.

[0132] As an example, in step S104, the deployed load balancer can dynamically allocate computing resources according to the load to ensure the stability of the system (enterprise financial analysis system) under high concurrency.

[0133] As an example, in step S105, a fault tolerance mechanism (such as redundant backup and failover, etc.) can be adopted to ensure that the system can still operate normally when some modules fail.

[0134] As an example, after step S105, various performance indicators of each module (such as response time, throughput, and CPU usage, etc.) can also be monitored in real time. According to the monitoring results, each module can be optimized. For example, if it is found that the response time of a certain module is too long, its performance can be improved by increasing computing resources or optimizing the algorithm.

[0135] As an example, in step S106, a user feedback collection mechanism can be designed to collect users' usage experiences and improvement suggestions through methods such as questionnaires and user interviews. At the same time, users' operation behaviors and model prediction results are automatically recorded as feedback data.

[0136] As an example, in step S107, at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model can be iteratively improved according to user feedback and the automatically recorded data. For example, if users feedback that the prediction results of the model are inaccurate, the prediction accuracy can be improved by increasing training data or adjusting model parameters. At the same time, A / B testing technology is adopted to compare the effects of different models or algorithms and select the optimal solution.

[0137] As an example, the above steps S104 to S107 can be executed based on the system integration and optimization module, which aims to organically integrate each functional module to ensure the overall stability and collaborative working ability of the system, and continuously improve the model and algorithm according to the feedback in actual applications to adapt to the changing market environment and enterprise needs.

[0138] In a specific example, taking a manufacturing enterprise as an example, the system integration and optimization module can correspond to the following steps: module interface design, data flow management, load balancing and fault tolerance mechanism, performance monitoring and optimization, user feedback collection, model and algorithm iterative improvement.

[0139] Module interface design: Design standardized RESTful API interfaces for each module such as the heterogeneous data fusion layer, dynamic risk assessment module, profit correlation inference engine, etc., to ensure smooth data exchange and communication between modules.

[0140] Data flow management: Build a unified data flow management system, and use ETL technology to extract, transform, and load data to ensure data consistency and integrity. At the same time, use data version control technology to ensure the traceability of data historical versions.

[0141] Load balancing and fault tolerance mechanism: Deploy a load balancer to dynamically allocate computing resources according to system load to ensure the stability of the system under high concurrency. At the same time, adopt redundant backup and failover mechanisms to ensure that the system can still operate normally when some modules fail.

[0142] Performance monitoring and optimization: Build a performance monitoring module to monitor the performance of each module in the system in real time, and optimize the system according to the monitoring results. For example, if it is found that the response time of the dynamic risk assessment module is too long, its performance can be improved by increasing computing resources or optimizing the algorithm.

[0143] User feedback collection: Design a user feedback collection mechanism to collect users' usage experiences and improvement suggestions for the system through questionnaires and system automatic records.

[0144] Model and algorithm iterative improvement: Iteratively improve the model and algorithm according to user feedback and data automatically recorded by the system. For example, if users feedback that the prediction results of the model are inaccurate, the prediction accuracy can be improved by increasing training data or adjusting model parameters. At the same time, adopt A / B testing technology to compare the effects of different models or algorithms and select the optimal solution.

[0145] Through the above steps, the system integration and optimization module can organically integrate each functional module, ensure the overall stability and collaborative working ability of the system, and continuously improve the model and algorithm according to the feedback in actual applications to adapt to the changing market environment and enterprise requirements.

[0146] In another embodiment, please refer to Figure 2, this application also provides an enterprise financial analysis system, which may include: a data acquisition module 10, a heterogeneous data fusion layer 20, and a dynamic risk assessment module 30; the data acquisition module 10 is used to acquire the structured financial data and unstructured market data of the enterprise; the heterogeneous data fusion layer 20 is used to preprocess the acquired structured financial data and unstructured market data; use the multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data; construct an enterprise relationship graph, and generate a dynamic enterprise relationship graph based on the enterprise relationship graph; fuse the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data; the dynamic risk assessment module 30 is used to construct an LSTM-TCN hybrid time series model; obtain macroeconomic fluctuation data; use the macroeconomic fluctuation data to pre-train the LSTM-TCN hybrid time series model; through a domain adaptation layer, use the enterprise financial data to fine-tune the pre-trained LSTM-TCN hybrid time series model; input the current enterprise financial data into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain a risk assessment result.

[0147] As an example, please combine Figure 2 participate in Figure 3 , the enterprise financial analysis system may further include: a profit correlation inference engine 40, and the profit correlation inference engine 40 is used to obtain a multi-stage profit optimization strategy for the enterprise based on the structured financial data of the enterprise, the risk assessment result, and the market environment data.

[0148] As an example, the profit correlation inference engine 40 is further used to construct a profit correlation inference engine model, and the profit correlation inference engine model obtains a multi-stage profit optimization strategy for the enterprise based on the structured financial data of the enterprise, the risk assessment result, and the market environment data; the enterprise financial analysis system may further include: a dynamic knowledge iteration module (not shown), and the dynamic knowledge iteration module is used to periodically update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model based on an online incremental learning mechanism; use SHAP values for analysis to evaluate the contribution of each feature to the LSTM-TCN hybrid time series model and the profit correlation inference engine model; according to the analysis results of the SHAP values, optimize the feature engineering and update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model.

[0149] As an example, the enterprise financial analysis system may further include: a system integration and optimization module (not shown), and the system integration and optimization module may include: a plurality of standardized interfaces (not shown), and a plurality of standardized structures are respectively located on each module to ensure smooth data exchange and communication between each module; a data flow management unit (not shown), and the data flow management unit is used to ensure smooth data flow between each module; a load balancer (not shown), and the load balancer is used to dynamically allocate computing resources according to the load; a fault tolerance mechanism module (not shown), and the fault tolerance mechanism module is used to adopt a fault tolerance mechanism to ensure that the enterprise financial system can still operate normally when some modules fail; a user feedback information collection module (not shown), and the user feedback information collection module is used to collect user feedback information; an iterative improvement module (not shown), and the iterative improvement module is used to iteratively improve at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model based on the user feedback information.

[0150] As an example, the enterprise financial analysis system in this embodiment can be used to execute as Figure 1 the enterprise financial analysis method described in the related embodiments.

[0151] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features of the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0152] The above embodiments only represent several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation to the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. An enterprise financial analysis method, characterized in that, including: obtaining the structured financial data and unstructured market data of an enterprise; preprocessing the obtained structured financial data and unstructured market data; using a multi-head self-attention mechanism to generate a joint embedding representation from the preprocessed structured financial data and unstructured market data; constructing an enterprise relationship graph and generating a dynamic enterprise relationship graph based on the enterprise relationship graph; fusing the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data; constructing an LSTM-TCN hybrid time series model; obtaining macroeconomic fluctuation data; using the macroeconomic fluctuation data to pre-train the LSTM-TCN hybrid time series model; through a domain adaptation layer, using the enterprise financial data to fine-tune the pre-trained LSTM-TCN hybrid time series model; inputting the current enterprise financial data into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain a risk assessment result.

2. The method according to claim 1, characterized in that, The formula for using a multi-head self-attention mechanism to generate a joint embedding representation from the preprocessed structured financial data and unstructured market data is as follows: Among them, Q is the query matrix, K is the key matrix, and K T represents the transpose of matrix K, represents the matrix multiplication of the query matrix Q and the transpose of the key matrix K to calculate the similarity between the query and the key. V is the value matrix, and d k is the dimension of the key vector.

3. The method according to claim 1, wherein Constructing an enterprise relationship graph and generating a dynamic enterprise relationship graph based on the enterprise relationship graph, including: taking enterprises as nodes and the business relationships between enterprises as edges to construct the enterprise relationship graph; using CNN for graph convolution operations to capture the correlation information between enterprises to generate the dynamic enterprise relationship graph; the formula for the graph convolution operation is as follows: Among them, is the adjacency matrix with self-loops, is the degree matrix, is the node feature matrix of the l-th layer, is the node feature matrix of the (l + 1)-th layer, is the learnable weight matrix, is the activation function.

4. The method according to any one of claims 1 to 3, characterized in that After obtaining the risk assessment result, it further includes: Based on the structured financial data of the enterprise, the risk assessment result, and market environment data, obtaining a multi-stage profit optimization strategy for the enterprise, including: collecting the structured financial data of the enterprise, the risk assessment result, and the market environment data; obtaining the causal relationship between risk control measures and profit indicators according to Granger causality test and Bayesian network; obtaining an initial multi-stage profit optimization strategy based on the causal relationship; using a reinforcement learning algorithm to optimize the initial multi-stage profit optimization strategy to obtain a multi-stage profit optimization strategy.

5. The method according to claim 4, wherein using a profit correlation inference engine model to obtain a multi-stage profit optimization strategy for the enterprise based on the structured financial data of the enterprise, the risk assessment result, and market environment data; After obtaining the multi-stage profit optimization strategy for the enterprise, it further includes: periodically updating the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model based on an online incremental learning mechanism; using SHAP values for analysis to evaluate the contribution of each feature to the LSTM-TCN hybrid time series model and the profit correlation inference engine model; according to the analysis results of SHAP values, optimizing the feature engineering and updating the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model.

6. The method according to claim 5, characterized in that, It further includes: deploying a load balancer to dynamically allocate computing resources according to the load; during the analysis process, adopting a fault tolerance mechanism to ensure the continuity of the analysis process; collecting the feedback information of users; Iteratively improve at least one of the models including the LSTM-TCN hybrid time series model and the profit-related inference engine model based on the user feedback information.

7. An enterprise financial analysis system, characterized in that, It includes: A data acquisition module for acquiring the structured financial data and unstructured market data of an enterprise; A heterogeneous data fusion layer for preprocessing the acquired structured financial data and unstructured market data; Use the multi-head self-attention mechanism to generate a joint embedding representation of the preprocessed structured financial data and unstructured market data; Construct an enterprise relationship graph and generate a dynamic enterprise relationship graph based on the enterprise relationship graph; Fuse the joint embedding representation and the dynamic enterprise relationship graph to obtain enterprise financial data; A dynamic risk assessment module for constructing an LSTM-TCN hybrid time series model; acquiring macroeconomic fluctuation data; Use the macroeconomic fluctuation data to pre-train the LSTM-TCN hybrid time series model; Through the domain adaptation layer, use the enterprise financial data to fine-tune the pre-trained LSTM-TCN hybrid time series model; Input the current enterprise financial data into the fine-tuned LSTM-TCN hybrid time series model for risk assessment to obtain a risk assessment result.

8. The enterprise financial analysis system according to claim 7, characterized in that It also includes: A profit-related inference engine for obtaining a multi-stage profit optimization strategy of an enterprise based on the structured financial data of the enterprise, the risk assessment result, and market environment data.

9. The enterprise financial analysis system according to claim 8, characterized in that, The profit-related inference engine is also used to construct a profit-related inference engine model, and the profit-related inference engine model obtains a multi-stage profit optimization strategy of an enterprise based on the structured financial data of the enterprise, the risk assessment result, and market environment data; The enterprise financial analysis system also includes: A dynamic knowledge iteration module for regularly updating the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit-related inference engine model based on the online incremental learning mechanism; using SHAP values for analysis to evaluate the contribution of each feature to the LSTM-TCN hybrid time series model and the profit-related inference engine model; According to the analysis results of the SHAP values, optimize the feature engineering and update the model parameters of at least one of the models including the LSTM-TCN hybrid time series model and the profit-related inference engine model.

10. The enterprise financial analysis system according to claim 9, characterized in that It also includes: A system integration and optimization module, and the system integration and optimization module includes: Multiple standardized interfaces located on each module to ensure smooth data exchange and communication between modules; A data flow management unit for ensuring the smooth flow of data between modules; A load balancer for dynamically allocating computing resources according to the load; A fault tolerance mechanism module for adopting a fault tolerance mechanism to ensure that the enterprise financial system can still operate normally when some modules fail; A user feedback information collection module for collecting user feedback information; An iterative improvement module for iteratively improving at least one of the models including the LSTM-TCN hybrid time series model and the profit correlation inference engine model based on user feedback information.

Images