A financial data risk analysis method and system based on large model

Abstract

The present invention provides a financial data risk analysis method and system based on a large model, relating to the field of intelligent risk control technology. The method includes obtaining financial data from participating parties and standardizing and aggregating them through a federated learning framework, generating federated feature vectors, constructing a two-way knowledge distillation model to receive vectors, and deploying teacher and student models to construct a risk transmission map. The system uses a generative adversarial network to monitor changes in enterprise association networks, adjusts risk transmission edge weights, aligns financial report time series and transaction event sequences through a spatiotemporal coupling encoder based on LSTM and Transformer architectures, generates multimodal fusion feature vectors and inputs them into the two-way knowledge distillation model. Combined with the map, a final risk score is obtained, and a hierarchical early warning mechanism is triggered to generate risk disposal recommendations and transmission paths, thereby improving data standardization and aggregation efficiency, enhancing the dynamic coupling capability of the multimodal fusion mechanism, and enhancing the adaptability of the risk transmission model.

Description

Technical Field

[0001] The present invention relates to the field of intelligent risk control technology, and in particular to a financial data risk analysis method and system based on a large model. Background Art

[0002] With the rapid development of the digital economy, corporate financial data has gradually shown characteristics such as multi-source heterogeneity, strong real-time nature, and a high proportion of unstructured data. Traditional financial data risk analysis methods face significant challenges.

[0003] In recent years, federated learning technology, through distributed privacy protection mechanisms, has enabled the collaborative optimization of model parameters while ensuring data localization, providing a theoretical foundation for cross-institutional financial data analysis. Large Language Models (LLMs), with their powerful semantic understanding capabilities, can effectively extract risk signals from unstructured text, but their black-box nature and high energy consumption limit practical model deployment. Furthermore, existing risk transmission mechanisms are often based on static graph models, such as PageRank, which struggle to dynamically respond to market fluctuations and the real-time evolution of corporate relationships. This leads to delayed risk warnings and high misjudgment rates.

[0004] Existing financial risk analysis methods still have flaws: First, traditional federated learning adopts a homogeneous data preprocessing strategy, ignoring the heterogeneity of the data of the participants, and is unable to meet the real-time requirements of high-frequency trading scenarios; second, although existing technologies attempt to combine LSTM and spatiotemporal graph networks for risk prediction, they have not built a unified encoding framework to achieve fine-grained alignment of temporal features and event sequences, which leads to overly coarse granularity in risk representation; finally, traditional methods based on static association matrices cannot effectively capture changes in the transmission paths of risks between enterprises. In addition, when introducing generative adversarial networks (GANs) to monitor association networks, they often face the dual challenges of pattern collapse risk and privacy leakage. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a financial data risk analysis method and system based on a large model, which is used to solve the problems of insufficient efficiency of data standardization and security aggregation under the traditional federated learning framework, lack of dynamic coupling capability of multimodal data fusion mechanism, and insufficient dynamic adaptability of risk transmission model.

[0006] The present invention provides a financial data risk analysis method based on a large model, the method comprising:

[0007] Obtain structured and unstructured financial data from multiple parties, standardize and securely aggregate the structured and unstructured financial data through a federated learning framework, and generate federated feature vectors;

[0008] Build a bidirectional knowledge distillation model and receive federated feature vectors. Deploy a central pre-trained large model as the teacher model, and each participant deploys a student model locally.

[0009] Construct a real-time risk transmission map, use generative adversarial networks to monitor changes in enterprise association networks, and dynamically adjust the risk transmission edge weights in the real-time risk transmission map;

[0010] Through a spatiotemporal coupled encoder based on LSTM and Transformer architecture, the financial report time series and transaction event series are aligned to generate a multimodal fusion feature vector.

[0011] The multimodal fusion feature vector is input into the bidirectional knowledge distillation model, combined with the real-time risk transmission map, and the final risk score is obtained through weighted calculation;

[0012] The hierarchical warning mechanism is triggered based on the final risk score and risk disposal recommendations are generated. At the same time, the risk transmission path is displayed through visualization tools.

[0013] Preferably, the method of performing standardization and secure aggregation on structured and unstructured financial data through a federated learning framework to generate a federated feature vector specifically includes:

[0014] Perform missing value processing and data trimming on structured financial data to generate structured financial features, and perform text cleaning and semantic feature extraction on unstructured financial data to generate unstructured financial features;

[0015] Standardize structured and unstructured financial features;

[0016] Obtain real-time features, sequentially concatenate structured financial features, unstructured financial features, and real-time features to generate a federated feature vector;

[0017] Adaptive layer normalization is used to normalize the federated feature vector, and the contribution weights of structured financial features, unstructured financial features, and real-time features to the federated feature vector are dynamically adjusted.

[0018] Preferably, the bidirectional knowledge distillation model is constructed and federated feature vectors are received, a central pre-trained large model is deployed as a teacher model, and each participant locally deploys a student model, specifically including:

[0019] Build a bidirectional knowledge distillation model and receive federated feature vectors. Use a hybrid architecture of BERT and LSTM to build a teacher model, and use a lightweight MobileNetV1 architecture to build a student model.

[0020] The federated feature vector is input into the teacher model and the first risk prediction probability distribution is generated through the fully connected layer;

[0021] The student model is fine-tuned based on local data and learns the first risk prediction probability distribution of the teacher model, while the prediction error of the student model is calculated through cross entropy loss;

[0022] After training with local data, the student model generates a second risk prediction probability distribution, which is encrypted using differential privacy protection and transmitted to the teacher model.

[0023] By comparing and learning the first risk prediction probability distribution and the second risk prediction probability distribution, the teacher model is updated to optimize the global risk prediction ability of the teacher model.

[0024] Preferably, the construction of a real-time risk transmission map, using a generative adversarial network to monitor changes in enterprise association networks, and dynamically adjusting the risk transmission edge weights in the real-time risk transmission map specifically includes:

[0025] The nodes in the real-time risk transmission map are set to include enterprise nodes, associated nodes, and macro nodes, and the initial edge weights between nodes are calculated based on the correlation of historical default probabilities;

[0026] Generative adversarial networks are used to monitor changes in enterprise association networks. Risk events between enterprises are captured in real time through a streaming processing framework. When a risk event is detected, the structure of the real-time risk transmission map is adjusted, that is, the risk transmission edge weights in the real-time risk transmission map are dynamically adjusted.

[0027] Among them, the generator of the generative adversarial network uses a graph convolutional network to generate potential risk conduction edges, and the discriminator inputs the current real-time risk conduction map to identify and output high-risk conduction edges.

[0028] Preferably, the feature alignment of the financial report time series and the transaction event sequence is performed by using a spatiotemporal coupled encoder based on the LSTM and Transformer architecture to generate a multimodal fusion feature vector, specifically including:

[0029] Use LSTM to encode the financial report time series and generate financial report features;

[0030] Transformer is used to encode transaction event sequences and generate transaction features;

[0031] Financial report features and transaction features are aligned through the cross-attention mechanism, and a multimodal fusion feature vector is output.

[0032] Preferably, the multimodal fusion feature vector is input into the bidirectional knowledge distillation model, combined with the real-time risk transmission map, and the final risk score is obtained through weighted calculation, which specifically includes:

[0033] The multimodal fusion feature vector is input into the student model of the bidirectional knowledge distillation model to calculate the preliminary risk score;

[0034] Obtain the global risk score generated by the teacher model and train the student model through forward distillation;

[0035] The trained student model generates a local risk score and feeds it back to the teacher model through reverse distillation to optimize the teacher model's global risk prediction ability;

[0036] Combined with the real-time risk transmission map, a graph convolutional network is used to perform weighted calculation on the preliminary risk scores of neighboring companies to generate the final risk score of the company.

[0037] Preferably, triggering a graded early warning mechanism based on the final risk score and generating risk disposal suggestions, while displaying the risk transmission path through a visualization tool, specifically includes:

[0038] Set three-level warning thresholds for the hierarchical warning mechanism;

[0039] Calculate the enterprise's comprehensive early warning risk score based on the enterprise's final risk score and the final risk scores of neighboring enterprises;

[0040] Compare the enterprise's comprehensive early warning risk score with the three-level early warning threshold to determine the enterprise's early warning risk level and the corresponding risk disposal recommendations and risk transmission path, and display the risk transmission path through visualization tools.

[0041] A financial data risk analysis system based on a large model, the system comprising:

[0042] The data processing module is used to obtain structured and unstructured financial data from multiple parties, standardize and securely aggregate the structured and unstructured financial data through the federated learning framework, and generate federated feature vectors;

[0043] The model generation module is used to build a bidirectional knowledge distillation model and receive federated feature vectors. It deploys a central pre-trained large model as the teacher model, and each participant deploys a student model locally.

[0044] The graph construction module is used to construct a real-time risk transmission graph, use generative adversarial networks to monitor changes in enterprise association networks, and dynamically adjust the risk transmission edge weights in the real-time risk transmission graph;

[0045] The feature fusion module is used to align the features of the financial report time series and the transaction event series through a spatiotemporal coupled encoder based on the LSTM and Transformer architecture to generate a multimodal fusion feature vector;

[0046] The risk scoring module is used to input the multimodal fusion feature vector into the bidirectional knowledge distillation model, combine it with the real-time risk transmission map, and obtain the final risk score through weighted calculation;

[0047] The early warning trigger module is used to trigger the graded early warning mechanism based on the final risk score and generate risk disposal recommendations, while displaying the risk transmission path through visualization tools.

[0048] An electronic device includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the steps of a large model-based financial data risk analysis method as described above.

[0049] A readable storage medium having a computer program stored therein, wherein the computer program, when executed by a processor, is used to implement the steps of a large model-based financial data risk analysis method as described above.

[0050] Compared with related technologies, the financial data risk analysis method and system based on a large model provided by the present invention has the following beneficial effects:

[0051] The present invention ensures the normalization of model input data by cleaning and standardizing financial data, reduces the prediction error caused by data loss and excessive noise, and makes subsequent risk analysis more accurate and stable; by adopting the Paillier homomorphic encryption algorithm to encrypt the local feature vectors of each participant and performing distributed key management, it ensures that the data of each participant will not be directly shared but can still achieve joint calculation, thereby improving the comprehensiveness of risk assessment; by forward distillation, the prediction ability of the student model is improved, the computational burden of local deployment of complex models in enterprises is reduced, and the risk assessment ability is improved at the same time; by reverse distillation, the global prediction ability of the teacher model is optimized, so that the teacher model The model can dynamically adapt to the personalized risk assessment needs of various financial institutions, and is more industry-universal and personalized adaptable. It dynamically adjusts risk transmission paths through streaming monitoring and generative adversarial networks, and dynamically identifies risk transmission paths between enterprises, so that it can adjust risk relationships between enterprises according to the latest market changes, discover potential risk diffusion paths in advance, and reduce systemic financial risks. It aligns low-frequency financial data with high-frequency transaction data through a spatiotemporal coupling encoder, which improves the correlation between different types of data and reduces information loss in the time dimension. It optimizes risk scores through graph neural networks, improves the accuracy and interpretability of risk scores, and provides financial institutions with more valuable risk control data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a financial data risk analysis method based on a large model provided by an embodiment of the present invention;

[0053] Figure 2 A system block diagram of a financial data risk analysis system based on a large model provided by an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0056] like Figure 1 FIG. 1 is a flow chart of a method for analyzing financial data risk based on a large model according to an embodiment of the present invention. Figure 1The execution subject of the method shown may be a software and / or hardware device. The execution subject of the present application may include but is not limited to at least one of the following: user equipment, network equipment, etc. Among them, the user equipment may include but is not limited to computers, smart phones, personal digital assistants (PDAs) and the electronic devices mentioned above. Network equipment may include but is not limited to a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers, wherein cloud computing is a type of distributed computing, a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. It includes steps S1 to S6, as follows:

[0057] S1, obtains structured and unstructured financial data from multiple participants, standardizes and securely aggregates the structured and unstructured financial data through the federated learning framework, and generates federated feature vectors;

[0058] S2 builds a bidirectional knowledge distillation model and receives federated feature vectors, deploys a central pre-trained large model as the teacher model, and each participant deploys a student model locally;

[0059] S3, builds a real-time risk transmission map, uses generative adversarial networks to monitor changes in enterprise association networks, and dynamically adjusts the risk transmission edge weights in the real-time risk transmission map;

[0060] S4, through the spatiotemporal coupled encoder based on LSTM and Transformer architecture, aligns the features of financial report time series and transaction event series to generate a multimodal fusion feature vector;

[0061] S5: Input the multimodal fusion feature vector into the bidirectional knowledge distillation model, combine it with the real-time risk transmission map, and obtain the final risk score through weighted calculation;

[0062] S6 triggers the graded warning mechanism based on the final risk score and generates risk disposal recommendations, while displaying the risk transmission path through visualization tools.

[0063] In the specific implementation process, the federated learning framework is used to standardize and securely aggregate structured and unstructured financial data to generate federated feature vectors, specifically including:

[0064] Perform missing value processing and data trimming on structured financial data to generate structured financial features, and perform text cleaning and semantic feature extraction on unstructured financial data to generate unstructured financial features;

[0065] For example, the participants can be financial institutions, which collect structured financial data from multiple financial institutions, including financial statement data in XBRL format, such as balance sheets, income statements, cash flow statements, etc.; transaction flow data, that is, corporate transaction records, such as large transfers, overdue repayment records, etc.; credit rating data, such as the company's credit rating, historical rating change trends, etc.

[0066] Structured financial data can then be stored in a standardized format, preferably Parquet, which outperforms CSV and supports efficient querying and compressed storage. Furthermore, time field data should be formatted in ISO 8601, such as "YYYY-MM-DD HH:MM:SS." Numerical data should be rounded to four decimal places to avoid loss of precision.

[0067] For numeric fields, such as net profit and debt-to-asset ratio, linear interpolation is usually used to fill missing values; for categorical fields, such as industry classification, the mode is used to fill in missing values, such as filling "manufacturing" with "manufacturing industry".

[0068] Furthermore, you can obtain the annual report text in PDF format from relevant databases or corporate websites, and access the financial news platform through API to capture the headlines and text of news related to financial risks. Then, you can use regular expressions to remove HTML tags, such as 、 The NLTK library is used to remove stop words such as "the" and "and," thereby retaining core semantic units. After loading weights using the BERT-base-uncased pre-trained model, fine-tuning is performed on classification tasks such as "debt default" and "liquidity depletion." The input text is segmented, and the attention weight for each sentence is calculated. Sentences with an attention weight greater than 0.8 are identified as key risk statements. Furthermore, key risk statements are segmented and industry-specific word embeddings are trained using the Word2Vec model, with a window size of 5×5 and 10 iterations.

[0069] Then, high-frequency dynamic data from the financial market can be collected, including high-frequency trading data such as buy and sell orders, trading volume, and market information; macroeconomic indicators such as GDP growth and the Consumer Price Index (CPI); and supply chain data such as upstream raw material price fluctuations and logistics information. Furthermore, Apache Flink and Apache Kafka can be used for streaming data transmission to ensure low latency, and Delta Lake can be used for data storage, supporting data versioning and backtracking.

[0070] Standardize structured and unstructured financial features;

[0071] Specifically, you can use the Z-score to normalize numeric fields in structured financial data, such as net profit and debt-to-asset ratio, and perform one-hot encoding on categorical fields, such as company size and industry classification, to generate binary vectors. Then, you can perform text segmentation on the cleaned unstructured financial data, using a trained Word2Vec model to generate word vectors, which are then sequentially concatenated into a dense vector of length 1000.

[0072] Furthermore, the Paillier homomorphic encryption algorithm can be used to encrypt local feature vectors and generate ciphertext to ensure that the original information is not leaked during data transmission. Local feature vectors refer to the enterprise risk feature data calculated and stored locally by financial institutions. The FATE framework can also be used to implement distributed key management for data. Specifically, the FATE framework sends ciphertext to a central server, which sums and decrypts the ciphertexts of all institutions to generate global features. This ensures that each institution can only access the encrypted global features and cannot restore the original data.

[0073] Obtain real-time features, sequentially concatenate structured financial features, unstructured financial features, and real-time features to generate a federated feature vector;

[0074] It should be noted that the structured financial features are 512-dimensional and include financial ratios such as debt-to-asset ratio, current ratio, and historical default probability. The unstructured financial features are 768-dimensional and are generated by Word2Vec word embeddings and semantic attention weighting. The real-time features are 1-dimensional and represent the current volatility calculated by the Flink stream processing engine. These three features are sequentially concatenated into a federated feature vector of 1281 dimensions.

[0075] Adaptive layer normalization is used to normalize the federated feature vector, and the contribution weights of structured financial features, unstructured financial features, and real-time features to the federated feature vector are dynamically adjusted.

[0076] It should be noted that in order to improve the expressiveness of the federated feature vector under different modal combinations, adaptive layer normalization (AdaLN) can be used to normalize the federated feature vector, and trainable modal weight parameters can be assigned to the structured financial features, unstructured financial features and real-time features in turn. In this way, the contribution weights of the structured financial features, unstructured financial features and real-time features to the federated feature vector can be dynamically adjusted during the training process, effectively improving the convergence speed and generalization ability of the bidirectional knowledge distillation model.

[0077] The bidirectional knowledge distillation model is constructed and federated feature vectors are received. A central pre-trained large model is deployed as a teacher model, and each participant locally deploys a student model. Specifically, the following steps are performed:

[0078] Build a bidirectional knowledge distillation model and receive federated feature vectors. Use a hybrid architecture of BERT and LSTM to build a teacher model, and use a lightweight MobileNetV1 architecture to build a student model.

[0079] The teacher model is constructed using a hybrid architecture of BERT and LSTM, combining natural language processing (NLP) and time series modeling capabilities. The BERT component extracts deep semantic features from corporate text information, capturing global contextual associations. The LSTM component models financial time series data, such as revenue growth rates and debt ratios, capturing temporal dependencies. Finally, a fully connected layer (FC) generates the first risk prediction probability distribution.

[0080] It should be noted that the BERT layer count is set to 12, with a hidden layer dimension of 768; the LSTM layer count is set to 3, with a hidden layer dimension of 512. The total number of parameters of the BERT and LSTM hybrid architecture is approximately 500M, deployed on a cloud-based GPU server. BERT is responsible for text encoding, with approximately 350M parameters; LSTM is responsible for time series modeling of financial data, with approximately 150M parameters. Furthermore, federated feature vectors can be used as input for training on a cloud-based GPU server, periodically updating the parameters of the BERT and LSTM hybrid architecture.

[0081] Then, we can use a lightweight MobileNetV1 architecture with 5M parameters, based on the 512-dimensional feature input layer output by federated feature engineering. By removing the last fully connected layer and replacing it with a risk prediction layer, we can build a student model and deploy it on the local servers of various financial institutions. The output category of the risk prediction layer is "default" or "normal", corresponding to a dimension of 2.

[0082] The federated feature vector is input into the teacher model and the first risk prediction probability distribution is generated through the fully connected layer;

[0083] In actual applications, the teacher model receives the 512-dimensional federated feature vector output by federated feature engineering, which contains financial ratios, semantic metaphors, and real-time volatility. The first 384-dimensional federated feature vector is input into the BERT module for processing unstructured features; the second 128-dimensional federated feature vector is input into the LSTM module for processing time series features. Finally, the first risk prediction probability distribution is generated through the fully connected layer. The corresponding expression is:

[0084] ;

[0085] in, is the first risk prediction probability distribution; is the federated feature vector; Multi-source financial data for enterprises; is the risk category, such as normal, warning, and default; represents the activation function of the fully connected layer.

[0086] The student model is fine-tuned based on local data and learns the first risk prediction probability distribution of the teacher model, while the prediction error of the student model is calculated through cross entropy loss;

[0087] After training with local data, the student model generates a second risk prediction probability distribution, which is encrypted using differential privacy protection and transmitted to the teacher model.

[0088] The expression of the second risk prediction probability distribution is:

[0089] ;

[0090] in, Predicted probability distribution for the second risk of the Student model.

[0091] The student model's prediction error is calculated using cross-entropy loss, and the model is optimized to make its predictions as close as possible to those of the teacher model. The student model updates its weights using a gradient descent algorithm, such as the Adam optimizer. After multiple rounds of iterative training, the student model's prediction ability gradually approaches that of the teacher model.

[0092] After training on local data, the student model can generate a personalized second risk prediction probability distribution and send it to the central server. Furthermore, a differential privacy protection mechanism can be employed to add noise to the second risk prediction probability distribution to prevent data leakage from a single institution. Specifically, Laplace noise can be added to the second risk prediction probability distribution for the "default" category, with a scale parameter Δ=0.5 / ε, where ε=2; and Gaussian noise can be added to the second risk prediction probability distribution for the "normal" category, with a mean of 0 and a standard deviation η=0.3. The noisy second risk prediction probability distribution can then be clipped to the interval [0,1] to ensure the rationality of the probability values.

[0093] By comparing and learning the first risk prediction probability distribution and the second risk prediction probability distribution, the teacher model is updated to optimize the global risk prediction ability of the teacher model.

[0094] Specifically, the teacher model is optimized by cross entropy loss and Kullback-Leibler divergence (KL divergence), and the prediction error of the optimized teacher model is for:

[0095] ;

[0096] ;

[0097] in, is the regularization coefficient, which is set to 0.1 to ensure the balance between cross entropy loss and Kullback-Leibler divergence.

[0098] Among them, KL divergence is used to calculate the difference in prediction distribution between the student model and the teacher model, encouraging the teacher model to absorb more personalized risk features.

[0099] It should be noted that during the first round of training, the teacher model is trained independently to generate initial weights, while the student model is trained locally using the teacher model's predictions. After this first round of training, multiple rounds of interactive training are required. Each round of training includes forward distillation, where the teacher model transmits risk prediction results to the student model; student model training, where the student model generates personalized risk predictions after training on local data; and backward distillation, where the student model feeds personalized risk predictions back to the teacher model. After each round of training, the student model is trained locally, a validation set is divided on the local data, and the loss is calculated. To prevent overfitting of the student model during fine-tuning, an early stopping mechanism is introduced to dynamically monitor the model training process. If the validation set loss does not show a significant downward trend or a fluctuating upward trend over several consecutive rounds, the training round is terminated early to prevent the student model parameters from falling into the overfitting zone. The student model parameters that perform best on the validation set are retained, and the personalized risk prediction probability distribution output by the student model is used as feedback. This is then fed back to the teacher model via a differential privacy mechanism for continuous optimization of the backward distillation and global model.

[0100] It's important to note that the bidirectional knowledge distillation model ensures the student model's personalization while enhancing the teacher model's global predictive capabilities. Differential privacy and homomorphic encryption ensure the security of data exchange. Using MobileNetV1 as the student model reduces local computing resource consumption, making the federated learning framework suitable for small and medium-sized financial institutions. Optimizing the teacher model using KL divergence accelerates model convergence and improves model iteration efficiency.

[0101] The construction of a real-time risk transmission map, using a generative adversarial network to monitor changes in enterprise association networks, and dynamically adjusting the risk transmission edge weights in the real-time risk transmission map specifically includes:

[0102] The nodes in the real-time risk transmission map are set to include enterprise nodes, associated nodes, and macro nodes, and the initial edge weights between nodes are calculated based on the correlation of historical default probabilities;

[0103] The real-time risk transmission map consists of three types of nodes: Company Node, which is used to represent individual companies, such as "Company A" and "Company B"; Relation Node, which is used to represent key connections between companies, such as "equity pledge", "supply chain relationship", "debt guarantee", etc.; Macro Node, which is used to represent macro factors affecting the entire industry, such as "market fluctuations" and "interest rate adjustments".

[0104] Complete real-time risk transmission map It can be expressed as:

[0105] ;

[0106] in, is the set of all nodes, , For enterprise nodes, is the associated node, is a macro node; is the risk transmission edge connecting these nodes.

[0107] The risk transmission edge weight represents the risk transmission intensity between enterprises. The initial state risk transmission edge weight, i.e., the initial edge weight The expression is as follows:

[0108] ;

[0109] in, Any risk transmission edge in the real-time risk transmission map, that is, the possible risk connection between enterprises, and The risk transmission edge and The correlation of historical default probabilities among connected companies, The expression is:

[0110] ;

[0111] in, is the default probability of enterprise A; is the default probability of enterprise B; is the set of all enterprise pairs; is the probability of enterprise defaulting on AB;

[0112] Generative adversarial networks are used to monitor changes in enterprise association networks. Risk events between enterprises are captured in real time through a streaming processing framework. When a risk event is detected, the structure of the real-time risk transmission map is adjusted, that is, the risk transmission edge weights in the real-time risk transmission map are dynamically adjusted.

[0113] Among them, the streaming processing framework is used to capture risk events between enterprises in real time, including: supply chain disruption events, that is, the default or bankruptcy of the upstream suppliers of the enterprise, which affects the production of downstream enterprises; equity pledge risk events, that is, the pledge ratio of major shareholders is too high, which can easily cause the stock price to plummet.

[0114] When a supply chain disruption is detected, a high-risk edge is added between the upstream and downstream companies. Furthermore, if the upstream company goes bankrupt, the associated supply chain edge is deleted. When a change in equity pledge is detected, a high-risk edge is added if the shareholder pledge ratio exceeds 60%. If the stock price fluctuates significantly, with a fluctuation greater than 10%, the edge weight is increased by 30%.

[0115] Among them, the generator of the generative adversarial network uses a graph convolutional network to generate potential risk conduction edges, and the discriminator inputs the current real-time risk conduction map to identify and output high-risk conduction edges.

[0116] The generator of the generative adversarial network uses a graph convolutional network to generate potential risk transmission edges and dynamically adjusts the initial edge weights. To ensure that the real-time risk transmission map maintains real-time and accuracy, the adjusted risk transmission edge weight for:

[0117] ;

[0118] in, The risk transmission edge weights predicted by the generative adversarial network, is the risk transmission edge weight at the previous moment, is the learning rate, set to 0.5, to balance new and old information.

[0119] It's important to note that Flink stream processing enables low-latency risk transmission calculations, keeping latency to less than 50ms. Using generative adversarial networks to predict potential high-risk relationships improves the adaptability of real-time risk transmission maps. By monitoring events like supply chain disruptions and equity pledges in real time, risk transmission relationships can be dynamically adjusted. Combining historical data with real-time events can adjust the intensity of risk transmission between enterprises, improving the accuracy of risk predictions.

[0120] The spatiotemporal coupled encoder based on the LSTM and Transformer architecture is used to align the features of the financial report time series and the transaction event series to generate a multimodal fusion feature vector. Specifically, the following steps are performed:

[0121] Use LSTM to encode the financial report time series and generate financial report features;

[0122] Transformer is used to encode transaction event sequences and generate transaction features;

[0123] Financial report features and transaction features are aligned through the cross-attention mechanism, and a multimodal fusion feature vector is output.

[0124] In practical applications, quarterly financial data, such as operating income, net profit, debt-to-asset ratio, etc., can be defined as ,in, Represents the quarter, Representing feature dimensions, such as revenue and cost, generates a time series of corporate financial reports. LSTM can then be used to encode the time series of corporate financial reports, generate time series hidden states, and ultimately output financial report features for subsequent feature fusion.

[0125] Then, key trading indicators such as buy-sell ratio and transaction volume fluctuation can be extracted from high-frequency trading data to construct trading event sequences. Then, Transformer can be used to encode the transaction event sequence and extract time-dependent information, ultimately obtaining transaction features for fusion with financial report features.

[0126] Since financial report data is low-frequency and often updated quarterly, while transaction data is high-frequency and typically changes in real time, a sliding window average of the transaction data is performed based on a time alignment mechanism and converted to quarterly time steps. Finally, a cross-attention mechanism can be used to fuse financial report features and transaction features. The attention mechanism calculates the weight distribution between the financial report features and transaction features to ensure the rationality of information fusion and form the final multimodal fusion feature vector.

[0127] It should be noted that the spatiotemporal coupled encoder based on the LSTM and Transformer architecture consists of two parallel input modules: one input module inputs temporal structured features such as corporate financial reports, which are encoded via LSTM to generate a time-dependent embedding representation; the other input module inputs a sequence of transaction events, which are encoded via Transformer to generate semantic contextual features. Subsequently, a cross-attention mechanism is used to couple and align the outputs of the two modalities, explicitly expressing the response of transaction behavior to historical financial indicators in the temporal dimension, ultimately generating a fused multimodal feature vector. This multimodal feature vector possesses temporal dynamics and behavioral semantic expression capabilities, which helps improve the accuracy of risk scoring calculations.

[0128] The multimodal fusion feature vector is input into the bidirectional knowledge distillation model, combined with the real-time risk transmission map, and the final risk score is obtained through weighted calculation, specifically including:

[0129] The multimodal fusion feature vector is input into the student model of the bidirectional knowledge distillation model to calculate the preliminary risk score;

[0130] Obtain the global risk score generated by the teacher model and train the student model through forward distillation;

[0131] The trained student model generates a local risk score and feeds it back to the teacher model through reverse distillation to optimize the teacher model's global risk prediction ability;

[0132] Combined with the real-time risk transmission map, a graph convolutional network is used to perform weighted calculation on the preliminary risk scores of neighboring companies to generate the final risk score of the company.

[0133] It should be noted that the input data for calculating the final risk score is the multimodal fusion feature vector, the risk conduction adjacency matrix, and the inter-firm edge weight matrix. The multimodal fusion feature vector is generated by a spatiotemporal coupled encoder based on LSTM and Transformer. Specifically, it includes the company's quarterly financial time series characteristics, which are generated by LSTM modeling, high-frequency trading behavior characteristics, which are generated by Transformer modeling, and unstructured text data such as announcements and industry news.

[0134] The risk association between enterprises is generated by a real-time risk transmission map, which uses a node-edge model to represent the direct or indirect risk transmission relationship between enterprises. Define the risk transmission adjacency matrix ,in, is the number of all enterprises; is the set of all risk conduction adjacency matrices.

[0135] In the risk transmission graph, each edge is represented by the inter-enterprise edge weight matrix Quantify the edge weight matrix between enterprises The expression is:

[0136] ;

[0137] in, Represents element-by-element multiplication, i.e., Hadamard product; represents the intensity of risk transmission, The calculation of is mainly based on the historical default correlation of the enterprise, such as historical data showing that the enterprise is highly correlated with a high-risk enterprise; supply chain dependence, such as enterprise 1 relies on enterprise 2 as a supplier. If enterprise 2 defaults, enterprise 1 may face a production crisis; market transaction behavior, such as frequent large-scale transactions between enterprises 3 and 4, indicating strong capital transactions.

[0138] The multimodal fusion feature vector Input into each participant’s locally deployed student model and calculate preliminary risk scores , the corresponding expression is:

[0139] ;

[0140] in, is the Sigmoid function, which is used to ensure that the initial risk score is normalized to the interval [0,1]. Indicates the impact of real-time risk transmission map on enterprise risks.

[0141] The teacher model provides a global risk score, and the student model is trained by absorbing this global information through forward knowledge distillation. After training, the student model generates local risk scores, which are fed back to the teacher model through backward distillation to optimize the teacher model's global risk prediction capabilities.

[0142] The final risk score is generated by weighting the risk score of the enterprise and the risk scores of neighboring enterprises. , the corresponding expression is:

[0143] ;

[0144] in, For enterprises The set of neighboring enterprises includes the enterprise Directly related enterprises, such as supply chain partners, investment parties, industry competitors, etc.; For enterprises With neighboring businesses the intensity of risk transmission between them; For enterprises Initial risk score.

[0145] In order to enable the risk score to be reasonably spread among enterprises, a graph convolutional network is used for calculation to output the optimized final risk score.

[0146] The final risk score is used to trigger a graded early warning mechanism and generate risk disposal recommendations. The risk transmission path is displayed through a visualization tool, specifically including:

[0147] Set three-level warning thresholds for the hierarchical warning mechanism;

[0148] Calculate the enterprise's comprehensive early warning risk score based on the enterprise's final risk score and the final risk scores of neighboring enterprises;

[0149] Compare the enterprise's comprehensive early warning risk score with the three-level early warning threshold to determine the enterprise's early warning risk level and the corresponding risk disposal recommendations and risk transmission path, and display the risk transmission path through visualization tools.

[0150] Based on the calculated final risk score and actual needs, the three-level warning thresholds of the hierarchical warning mechanism can be set:

[0151] ;

[0152] Among them, the primary warning means that the company has a certain default risk, but has not yet reached a dangerous level and needs to be continuously observed; the intermediate warning means that the company has entered a high-risk state, but still has the ability to save itself, and financial institutions should consider measures such as restricting financing quotas; the advanced warning means that the company is in a serious credit crisis and may face default or bankruptcy risks. It is necessary to freeze financing, increase collateral requirements, and even report to regulatory authorities.

[0153] Based on the final risk score of the enterprise and the final risk level of neighboring enterprises, the comprehensive early warning risk score of the enterprise is calculated :

[0154] ;

[0155] in, For neighboring businesses Final risk score of It is the weight coefficient, which is set to 0.7 by default and is used to control and balance the risks of the enterprise itself and the risks of neighboring enterprises.

[0156] Score the enterprise's comprehensive early warning risk Compare with the third level warning threshold, if If the warning threshold of a certain level is exceeded, the enterprise will enter the corresponding warning risk level and affect the warning status of its associated enterprises.

[0157] In addition, Neo4j can be used to store and manage real-time risk transmission maps, support dynamic query of risk transmission paths, and combine D3.js or Cytoscape.js to render network relationship diagrams, and dynamically display the risk transmission process and path through visualization tools.

[0158] In the visualization tool interface, each node represents an enterprise, among which red nodes represent high-level warning enterprises; orange nodes represent intermediate-level warning enterprises; yellow nodes represent primary-level warning enterprises; and the edge color depth represents the intensity of risk transmission.

[0159] In actual applications, corresponding disposal suggestions can also be formulated according to the enterprise risk level: for primary warnings, strengthen the financial review of the enterprise, pay close attention to its cash flow situation, and recommend that financial institutions reduce exposure to short-term financing risks; for intermediate warnings, limit the enterprise's new financing quota and require additional guarantees. If the enterprise's credit rating declines, the credit strategy needs to be adjusted in advance; for high-level warnings, freeze the enterprise's supply chain financing quota to prevent risk spillover, and require the enterprise to provide additional collateral to reduce the losses of financial institutions. At the same time, notify the regulatory authorities and report the enterprise's credit crisis to avoid the spread of systemic risks.

[0160] It's important to note that by visualizing risk transmission paths, the scope of enterprise risk impact can be displayed in real time, enabling financial institutions to take proactive measures. Intelligent decision-making provides risk response strategies at different levels, guiding enterprises and financial institutions in optimizing risk management decisions. Ultimately, a corporate risk report is generated, including risk scores, early warning levels, and transmission path analysis.

[0161] like Figure 2 FIG. 1 is a system block diagram of a financial data risk analysis system based on a large model provided by an embodiment of the present invention. The system includes:

[0162] The data processing module is used to obtain structured and unstructured financial data from multiple parties, standardize and securely aggregate the structured and unstructured financial data through the federated learning framework, and generate federated feature vectors;

[0163] The model generation module is used to build a bidirectional knowledge distillation model and receive federated feature vectors. It deploys a central pre-trained large model as the teacher model, and each participant deploys a student model locally.

[0164] The graph construction module is used to construct a real-time risk transmission graph, use generative adversarial networks to monitor changes in enterprise association networks, and dynamically adjust the risk transmission edge weights in the real-time risk transmission graph;

[0165] The feature fusion module is used to align the features of the financial report time series and the transaction event series through a spatiotemporal coupled encoder based on the LSTM and Transformer architecture to generate a multimodal fusion feature vector;

[0166] The risk scoring module is used to input the multimodal fusion feature vector into the bidirectional knowledge distillation model, combine it with the real-time risk transmission map, and obtain the final risk score through weighted calculation;

[0167] The early warning trigger module is used to trigger the graded early warning mechanism based on the final risk score and generate risk disposal recommendations, while displaying the risk transmission path through visualization tools.

[0168] Figure 2 The system of the embodiment shown can be used to perform Figure 1 The implementation principles and technical effects of the steps in the method embodiment shown are similar and will not be repeated here.

[0169] An electronic device includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the steps of a large model-based financial data risk analysis method as described above.

[0170] like Figure 3 FIG. 1 is a schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32 and a computer program;

[0171] The memory 32 is used to store the computer program, which may also be a flash memory. The computer program is, for example, an application program or a functional module for implementing the above method.

[0172] The processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant description in the above method embodiment.

[0173] Optionally, the memory 32 may be independent or integrated with the processor 31 .

[0174] When the memory 32 is a device independent of the processor 31, the device may further include:

[0175] The bus 33 is used to connect the memory 32 and the processor 31 .

[0176] A readable storage medium having a computer program stored therein, wherein the computer program, when executed by a processor, is used to implement the steps of a large model-based financial data risk analysis method as described above.

[0177] The readable storage medium may be a computer storage medium or a communication medium. Communication media include any medium that facilitates the transfer of computer programs from one location to another. Computer storage media may be any available medium that can be accessed by a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium may also be an integral part of the processor. The processor and the readable storage medium may be located in an application-specific integrated circuit (ASIC). In addition, the ASIC may be located in a user device. Of course, the processor and the readable storage medium may also exist as discrete components in a communication device. The readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0178] The present invention also provides a program product, which includes execution instructions stored in a readable storage medium. At least one processor of a device can read the execution instructions from the readable storage medium, and at least one processor executes the execution instructions so that the device implements the methods provided in the various embodiments described above.

[0179] In the embodiments of the above-mentioned devices, it should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), etc. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly executed by a hardware processor or by a combination of hardware and software modules within the processor.

[0180] Through the introduction of the above embodiments, the present invention uses a financial data risk analysis method and system based on a large model, and through data cleaning and standardization processing of financial data, ensures the standardization of model input data, reduces the prediction error caused by data missing and excessive noise, and makes subsequent risk analysis more accurate and stable; by adopting the Paillier homomorphic encryption algorithm to encrypt the local feature vectors of each participant and perform distributed key management, it ensures that the data of each participant will not be directly shared but can still achieve joint calculation, thereby improving the comprehensiveness of risk assessment; through forward distillation, the predictive ability of the student model is improved, the computational burden of the enterprise when deploying complex models locally is reduced, and the risk assessment ability is improved; through reverse distillation, the optimization The global prediction capability of the teacher model is optimized, so that the teacher model can dynamically adapt to the personalized risk assessment needs of various financial institutions, and is more industry-universal and personalized. Through streaming monitoring and GAN, the risk transmission path is dynamically adjusted, and the risk transmission path between enterprises is dynamically identified, so that the risk relationship between enterprises can be adjusted according to the latest market changes, potential risk diffusion paths can be discovered in advance, and systemic financial risks can be reduced. Through the spatiotemporal coupling encoder, low-frequency financial data and high-frequency transaction data are aligned, which improves the correlation between different types of data and reduces information loss in the time dimension. The risk score is optimized through the graph neural network, which improves the accuracy and interpretability of the risk score, and provides financial institutions with more valuable reference risk control data.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A financial data risk analysis method based on a large model, characterized in that: The method comprises: Obtain structured and unstructured financial data from multiple parties, standardize and securely aggregate the structured and unstructured financial data through a federated learning framework, and generate federated feature vectors; Build a bidirectional knowledge distillation model and receive federated feature vectors. Deploy a central pre-trained large model as the teacher model, and each participant deploys a student model locally. Construct a real-time risk transmission map, use generative adversarial networks to monitor changes in enterprise association networks, and dynamically adjust the risk transmission edge weights in the real-time risk transmission map; Through a spatiotemporal coupled encoder based on LSTM and Transformer architecture, the financial report time series and transaction event series are aligned to generate a multimodal fusion feature vector. The multimodal fusion feature vector is input into the bidirectional knowledge distillation model, combined with the real-time risk transmission map, and the final risk score is obtained through weighted calculation; The hierarchical warning mechanism is triggered based on the final risk score and risk disposal recommendations are generated. At the same time, the risk transmission path is displayed through visualization tools.

2. The method for risk analysis of financial data based on a large model according to claim 1, characterized in that: The federated learning framework is used to standardize and securely aggregate structured and unstructured financial data to generate federated feature vectors, specifically including: Perform missing value processing and data trimming on structured financial data to generate structured financial features, and perform text cleaning and semantic feature extraction on unstructured financial data to generate unstructured financial features; Standardize structured and unstructured financial features; Obtain real-time features, sequentially concatenate structured financial features, unstructured financial features, and real-time features to generate a federated feature vector; Adaptive layer normalization is used to normalize the federated feature vector, and the contribution weights of structured financial features, unstructured financial features, and real-time features to the federated feature vector are dynamically adjusted.

3. The method for risk analysis of financial data based on a large model according to claim 1, characterized in that: The bidirectional knowledge distillation model is constructed and federated feature vectors are received. A central pre-trained large model is deployed as a teacher model, and each participant locally deploys a student model. Specifically, the following steps are performed: Build a bidirectional knowledge distillation model and receive federated feature vectors. Use a hybrid architecture of BERT and LSTM to build a teacher model, and use a lightweight MobileNetV1 architecture to build a student model. The federated feature vector is input into the teacher model and the first risk prediction probability distribution is generated through the fully connected layer; The student model is fine-tuned based on local data and learns the first risk prediction probability distribution of the teacher model, while the prediction error of the student model is calculated through cross entropy loss; After training with local data, the student model generates a second risk prediction probability distribution, which is encrypted using differential privacy protection and transmitted to the teacher model. By comparing and learning the first risk prediction probability distribution and the second risk prediction probability distribution, the teacher model is updated to optimize the global risk prediction ability of the teacher model.

4. The method for risk analysis of financial data based on a large model according to claim 1, characterized in that: The construction of a real-time risk transmission map, using a generative adversarial network to monitor changes in enterprise association networks, and dynamically adjusting the risk transmission edge weights in the real-time risk transmission map specifically includes: The nodes in the real-time risk transmission map are set to include enterprise nodes, associated nodes, and macro nodes, and the initial edge weights between nodes are calculated based on the correlation of historical default probabilities; Generative adversarial networks are used to monitor changes in enterprise association networks. Risk events between enterprises are captured in real time through a streaming processing framework. When a risk event is detected, the structure of the real-time risk transmission map is adjusted, that is, the risk transmission edge weights in the real-time risk transmission map are dynamically adjusted. Among them, the generator of the generative adversarial network uses a graph convolutional network to generate potential risk conduction edges, and the discriminator inputs the current real-time risk conduction map to identify and output high-risk conduction edges.

5. The method for risk analysis of financial data based on a large model according to claim 1, characterized in that: The spatiotemporal coupled encoder based on the LSTM and Transformer architecture is used to align the features of the financial report time series and the transaction event series to generate a multimodal fusion feature vector. Specifically, the following steps are performed: Use LSTM to encode the financial report time series and generate financial report features; Transformer is used to encode transaction event sequences and generate transaction features; Financial report features and transaction features are aligned through the cross-attention mechanism, and a multimodal fusion feature vector is output.

6. The method for risk analysis of financial data based on a large model according to claim 1, characterized in that: The multimodal fusion feature vector is input into the bidirectional knowledge distillation model, combined with the real-time risk transmission map, and the final risk score is obtained through weighted calculation, specifically including: The multimodal fusion feature vector is input into the student model of the bidirectional knowledge distillation model to calculate the preliminary risk score; Obtain the global risk score generated by the teacher model and train the student model through forward distillation; The trained student model generates a local risk score and feeds it back to the teacher model through reverse distillation to optimize the teacher model's global risk prediction ability; Combined with the real-time risk transmission map, a graph convolutional network is used to perform weighted calculation on the preliminary risk scores of neighboring companies to generate the final risk score of the company.

7. The method for risk analysis of financial data based on a large model according to claim 1, characterized in that: The final risk score is used to trigger a graded early warning mechanism and generate risk disposal recommendations. The risk transmission path is displayed through a visualization tool, specifically including: Set three-level warning thresholds for the hierarchical warning mechanism; Calculate the enterprise's comprehensive early warning risk score based on the enterprise's final risk score and the final risk scores of neighboring enterprises; Compare the enterprise's comprehensive early warning risk score with the three-level early warning threshold to determine the enterprise's early warning risk level and the corresponding risk disposal recommendations and risk transmission path, and display the risk transmission path through visualization tools.

8. A financial data risk analysis system based on a large model, applied to a financial data risk analysis method based on a large model as claimed in any one of claims 1 to 7, characterized in that: The system comprises: The data processing module is used to obtain structured and unstructured financial data from multiple parties, standardize and securely aggregate the structured and unstructured financial data through the federated learning framework, and generate federated feature vectors; The model generation module is used to build a bidirectional knowledge distillation model and receive federated feature vectors. It deploys a central pre-trained large model as the teacher model, and each participant deploys a student model locally. The graph construction module is used to construct a real-time risk transmission graph, use generative adversarial networks to monitor changes in enterprise association networks, and dynamically adjust the risk transmission edge weights in the real-time risk transmission graph; The feature fusion module is used to align the features of the financial report time series and the transaction event series through a spatiotemporal coupled encoder based on the LSTM and Transformer architecture to generate a multimodal fusion feature vector; The risk scoring module is used to input the multimodal fusion feature vector into the bidirectional knowledge distillation model, combine it with the real-time risk transmission map, and obtain the final risk score through weighted calculation; The early warning trigger module is used to trigger the graded early warning mechanism based on the final risk score and generate risk disposal recommendations, while displaying the risk transmission path through visualization tools.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor runs the computer program stored in the memory, the processor executes the steps of the large model-based financial data risk analysis method according to any one of claims 1 to 7.

10. A readable storage medium storing a computer program, wherein: When the computer program is executed by a processor, it is used to implement the steps of a financial data risk analysis method based on a large model as described in any one of claims 1 to 7.

Images