Financial risk prediction method and system based on artificial intelligence

Abstract

The invention provides a financial risk prediction method and system based on artificial intelligence, and the method comprises the steps: inputting real-time transaction data, real-time macroeconomic indexes and real-time enterprise financial report data of a target financial market into a trained risk prediction model, and obtaining a risk prediction result; based on the risk prediction result and a preset risk result threshold interval, outputting a target financial market risk degree grade prediction result; wherein the training process of the risk prediction model comprises the steps of constructing a multi-modal feature matrix based on historical transaction data, historical macroeconomic indexes and historical enterprise financial report data of a target financial market, and then processing the multi-modal feature matrix according to matrix operation and principal component analysis to obtain a dimension reduction feature matrix; constructing a risk prediction model; and carrying out training and cross validation on the preliminary hybrid prediction model by using the dimension reduction feature matrix to obtain a trained risk prediction model.

Description

Technical Field

[0001] This invention belongs to the field of financial technology, specifically relating to a financial risk prediction method and system based on artificial intelligence. Background Art

[0002] Financial risk refers to the risk in financial activities that may lead to asset losses or failure to realize expected returns due to changes in the market environment, policy adjustments, operational errors, etc. Artificial intelligence, by learning from large amounts of historical data, can effectively predict changes in the financial market and assist financial institutions in identifying and managing financial risks. Common AI technologies include machine learning, deep learning, and natural language processing, which can help identify potential financial risks through pattern recognition, predictive modeling, and anomaly detection. However, existing financial risk prediction methods using deep learning models may overfit to the training data. This is because the volatility and complexity of financial data mean that even small market changes can have a significant impact, thus preventing effective generalization in practical applications. Summary of the Invention

[0003] To overcome the shortcomings of existing financial risk prediction methods that suffer from overfitting in deep learning models and thus fail to generalize effectively in practical applications, this invention provides an artificial intelligence-based financial risk prediction method and system. It comprehensively considers various economic data from the financial market and establishes a multimodal feature matrix to capture multi-dimensional dynamic alignment information of the financial market, improving the subsequent model's understanding of market information at different levels. Principal component analysis is used for dimensionality reduction to avoid overfitting from single features. By introducing appropriate composite regularization terms into the initial prediction model of the financial market to construct a hybrid model, overfitting is prevented, and the complexity can be adaptively adjusted according to data characteristics, improving prediction accuracy. Cross-validation comprehensively evaluates the model's stability and accuracy, ensuring that the model can not only handle complex and diverse market data but also adapt to dynamic market changes, thereby better achieving the prediction and control of financial risks.

[0004] According to one aspect of the present invention, an artificial intelligence-based financial risk prediction method is provided, comprising:

[0005] Input real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data from the target financial market into the trained risk prediction model to obtain risk prediction results.

[0006] Based on the risk prediction results and the preset risk result threshold range, the predicted risk level of the target financial market is output.

[0007] The training process of the risk prediction model includes:

[0008] A multimodal feature matrix is ​​constructed based on historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market. Then, the multimodal feature matrix is ​​processed by matrix operations and principal component analysis to obtain a dimension-reduced feature matrix.

[0009] A preliminary hybrid prediction model is constructed, which is a preliminary prediction model based on the updated parameters of the composite regularization term. The construction steps are as follows:

[0010] Obtain the target loss function of the preliminary prediction model, then obtain the first derivative of the target loss function with respect to the model output, and the second derivative with respect to the model parameters;

[0011] The gradient of the loss function is obtained by combining the first and second derivatives;

[0012] A composite regularization term is sequentially introduced into the preliminary prediction model. The composite regularization term is a matching pair of the first regularization term and the second regularization term. The gradients of the first regularization term and the second regularization term in the composite regularization term are obtained respectively.

[0013] The parameters of the preliminary prediction model are updated based on the gradients of the first regularization term, the second regularization term, and the loss function to obtain the preliminary hybrid prediction model.

[0014] Different preliminary mixed prediction models are constructed by sequentially selecting composite regularization terms according to preset rules. Simultaneously, the preliminary mixed prediction models are trained and cross-validated using dimensionality reduction feature matrices until the model output meets the preset requirements. The trained risk prediction model is then output.

[0015] As a further implementation scheme, the steps for constructing the multimodal feature matrix include:

[0016] Extract market transaction dynamics and transaction timestamps from historical transaction data;

[0017] Market behavior analysis features and their behavior timestamps are extracted from historical macroeconomic indicators. Then, based on the alignment of behavior timestamps with transaction timestamps, the market behavior analysis features are converted into market behavior analysis features that are synchronized with transaction dynamics.

[0018] Extract the financial health characteristics and financial timestamps of enterprises from historical corporate financial reports, and then convert the financial health characteristics into enterprise financial health characteristics that are dynamically synchronized with transaction timestamps based on the alignment of financial timestamps and transaction timestamps.

[0019] The market transaction dynamics, market behavior analysis features synchronized with transaction dynamics, and corporate financial health features are standardized and normalized sequentially. Then, they are spliced ​​and arranged according to the transaction timestamps and their order to obtain a multimodal feature matrix.

[0020] As a further implementation plan, the steps for constructing dimensionality reduction features are as follows:

[0021] Each feature in the multimodal feature matrix is ​​standardized to obtain a standardized multimodal feature matrix;

[0022] The standardized multimodal feature matrix is ​​transposed to obtain the transposed multimodal feature matrix, and the covariance matrix is ​​obtained based on the transposed multimodal feature matrix and the standardized multimodal feature matrix.

[0023] Eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvectors and eigenvalues. The eigenvalues ​​are sorted from largest to smallest and a preset number of the largest eigenvalues ​​are selected. The eigenvectors corresponding to the selected eigenvalues ​​are used as principal component eigenvectors.

[0024] The principal component matrix is ​​obtained by stacking the principal component feature vectors column by column. Then, the multimodal feature matrix is ​​mapped and projected onto the principal component matrix to obtain the dimensionality-reduced feature matrix.

[0025] As a further implementation plan, the initial prediction model is a machine learning algorithm model based on the financial market, including but not limited to linear regression models, decision tree models, support vector machines, or random forests.

[0026] As a further implementation plan, the steps to obtain the trained risk prediction model include:

[0027] The initial hybrid prediction model is trained using a dimensionality-reduced feature matrix. After training, the validation error of the results is obtained, and it is determined whether the minimum validation error is greater than a preset error threshold.

[0028] If yes, then according to the preset rules, reselect the target matching pairs of the first and second regularization terms in the composite regularization terms to construct a new preliminary mixed prediction model, and return to the training process to train the newly constructed preliminary mixed prediction model; otherwise, output the current preliminary mixed prediction model to obtain the trained risk prediction model.

[0029] According to another aspect of the present invention, an artificial intelligence-based financial risk prediction system is provided, comprising:

[0030] The risk prediction module is used to input real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data from the target financial market into the trained risk prediction model to obtain risk prediction results.

[0031] The risk rating module is used to output the risk level prediction results of the target financial market based on the risk prediction results and the preset risk result threshold range.

[0032] The training process of the risk prediction model includes:

[0033] A multimodal feature matrix is ​​constructed based on historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market. Then, the multimodal feature matrix is ​​processed by matrix operations and principal component analysis to obtain a dimension-reduced feature matrix.

[0034] A preliminary hybrid prediction model is constructed, which is a preliminary prediction model based on the updated parameters of the composite regularization term. The construction steps are as follows:

[0035] Obtain the target loss function of the preliminary prediction model, then obtain the first derivative of the target loss function with respect to the model output, and the second derivative with respect to the model parameters;

[0036] The gradient of the loss function is obtained by combining the first and second derivatives;

[0037] A composite regularization term is sequentially introduced into the preliminary prediction model. The composite regularization term is a matching pair of the first regularization term and the second regularization term. The gradients of the first regularization term and the second regularization term in the composite regularization term are obtained respectively.

[0038] The parameters of the preliminary prediction model are updated based on the gradients of the first regularization term, the second regularization term, and the loss function to obtain the preliminary hybrid prediction model.

[0039] Different preliminary mixed prediction models are constructed by sequentially selecting composite regularization terms according to preset rules. Simultaneously, the preliminary mixed prediction models are trained and cross-validated using dimensionality reduction feature matrices until the model output meets the preset requirements. The trained risk prediction model is then output.

[0040] As a further implementation scheme, the training process of the risk prediction model is completed within a pre-training module. The pre-training module includes: a first acquisition module, used to acquire historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market, and construct a multimodal feature matrix based on these data; a dimensionality reduction module, used to acquire a covariance matrix based on the multimodal feature matrix constructed by the first acquisition module, and perform principal component analysis on the covariance matrix to reduce its dimensionality, obtaining a dimensionality-reduced feature matrix; a second acquisition module, used to acquire a preliminary financial market prediction model and a composite regularization term; a selection module, used to select the composite regularization term acquired by the second acquisition module according to preset rules; a preliminary mixed prediction model construction module, used to introduce the composite regularization term selected by the selection module into the preliminary financial market prediction model acquired by the second acquisition module, obtaining a preliminary mixed prediction model; and a cross-validation module, used to train and cross-validate the preliminary mixed prediction model obtained by the preliminary mixed prediction model construction module based on the dimensionality-reduced feature matrix obtained by the dimensionality reduction module. The trained risk prediction model is obtained when the model output meets preset requirements.

[0041] As a further implementation, the dimensionality reduction module includes: a preprocessing unit, used to standardize each feature in the multimodal feature matrix to obtain a standardized multimodal feature matrix; a transpose unit, used to transpose the standardized multimodal feature matrix obtained by the preprocessing unit to obtain a transposed multimodal feature matrix, and obtain a covariance matrix based on the transposed multimodal feature matrix and the standardized multimodal feature matrix; a principal component filtering unit, used to perform eigenvalue decomposition on the covariance matrix obtained by the transpose unit to obtain multiple eigenvectors and eigenvalues, sort the eigenvalues ​​from largest to smallest and filter a preset number of the largest eigenvalues, and use the eigenvectors corresponding to the filtered eigenvalues ​​as principal component eigenvectors; and a mapping unit, used to stack the principal component eigenvectors obtained by the principal component filtering unit column-wise to obtain a principal component matrix, and then map and project the multimodal feature matrix onto the principal component matrix to obtain a dimensionality-reduced feature matrix.

[0042] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to execute an artificial intelligence-based financial risk prediction method.

[0043] According to another aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute an artificial intelligence-based financial risk prediction method.

[0044] Compared with existing technologies, the advantages of this invention are as follows: This invention comprehensively considers various economic data from the financial market and establishes a multimodal feature matrix to capture multi-dimensional dynamic alignment information of the financial market, improving the subsequent model's understanding of market information at different levels. Dimensionality reduction through principal component analysis avoids overfitting problems caused by single features. By introducing appropriate composite regularization terms into the initial prediction model of the financial market to construct a hybrid model, overfitting is prevented and the complexity can be adaptively adjusted according to data characteristics, improving prediction accuracy. Cross-validation comprehensively evaluates the stability and accuracy of the model, ensuring that the model can not only handle complex and diverse market data but also adapt to dynamic market changes, thereby better achieving the prediction and control of financial risks. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating an artificial intelligence-based financial risk prediction method provided in an embodiment of the present invention;

[0047] Figure 2 A flowchart illustrating an example of an artificial intelligence-based financial risk prediction method provided in an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of the structure of an artificial intelligence-based financial risk prediction system provided in an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of the structure of the pre-training module provided in an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the preprocessing module provided in an embodiment of the present invention;

[0051] Figure 6 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed Implementation

[0052] It should be noted that:

[0053] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, AI-based financial risk prediction method, system, product, or device that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, AI-based financial risk prediction methods, products, or devices.

[0054] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0056] like Figure 1 As shown, Figure 1 A flowchart illustrating an artificial intelligence-based financial risk prediction method provided in this embodiment of the invention is shown. The method includes:

[0057] Input real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data from the target financial market into the trained risk prediction model to obtain risk prediction results.

[0058] Based on the risk prediction results and the preset risk result threshold range, the predicted risk level of the target financial market is output.

[0059] The training process of the risk prediction model includes:

[0060] A multimodal feature matrix is ​​constructed based on historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market. Then, the multimodal feature matrix is ​​processed by matrix operations and principal component analysis to obtain a dimension-reduced feature matrix.

[0061] A preliminary hybrid prediction model is constructed, which is a preliminary prediction model based on the updated parameters of the composite regularization term. The construction steps are as follows:

[0062] Obtain the target loss function of the preliminary prediction model, then obtain the first derivative of the target loss function with respect to the model output, and the second derivative with respect to the model parameters;

[0063] The gradient of the loss function is obtained by combining the first and second derivatives;

[0064] A composite regularization term is sequentially introduced into the preliminary prediction model. The composite regularization term is a matching pair of the first regularization term and the second regularization term. The gradients of the first regularization term and the second regularization term in the composite regularization term are obtained respectively.

[0065] The parameters of the preliminary prediction model are updated based on the gradients of the first regularization term, the second regularization term, and the loss function, resulting in a preliminary hybrid prediction model.

[0066] Different preliminary mixed prediction models are constructed by sequentially selecting composite regularization terms according to preset rules. Simultaneously, the preliminary mixed prediction models are trained and cross-validated using dimensionality reduction feature matrices until the model output meets the preset requirements. The trained risk prediction model is then output.

[0067] As a preferred implementation, the applicability of L1 regularization (first regularization) and L2 regularization (second regularization) can be determined based on the characteristics of the data and the task objectives. If the data contains many redundant or unimportant features, L1 regularization will have an advantage; if the stability of all features is the focus, L2 regularization is more suitable. The preset rule is usually a standard used to select L1 and L2 regularization terms from the composite regularization terms according to specific conditions. For example, based on the regularization strength rule, the weight coefficients can be used to determine which term should be used as the L1 regularization term (first regularization term) and which term should be used as the L2 regularization term (second regularization term). Alternatively, based on the model requirements rule, if the model prefers a sparser solution, then L1 regularization will have an advantage; if the model prefers a smoother solution, then L2 regularization will have an advantage.

[0068] Furthermore, the steps for constructing the multimodal feature matrix include:

[0069] Extract market transaction dynamics and transaction timestamps from historical transaction data;

[0070] Market behavior analysis features and their behavior timestamps are extracted from historical macroeconomic indicators. Then, based on the alignment of behavior timestamps with transaction timestamps, the market behavior analysis features are converted into market behavior analysis features that are synchronized with transaction dynamics.

[0071] Extract the financial health characteristics and financial timestamps of enterprises from historical corporate financial reports, and then convert the financial health characteristics into enterprise financial health characteristics that are dynamically synchronized with transaction timestamps based on the alignment of financial timestamps and transaction timestamps.

[0072] The market transaction dynamics, market behavior analysis features synchronized with transaction dynamics, and corporate financial health features are standardized and normalized sequentially. Then, they are spliced ​​and arranged according to the transaction timestamps and their order to obtain a multimodal feature matrix.

[0073] Furthermore, the steps for constructing the dimensionality reduction features are as follows:

[0074] Each feature in the multimodal feature matrix is ​​standardized to obtain a standardized multimodal feature matrix;

[0075] The standardized multimodal feature matrix is ​​transposed to obtain the transposed multimodal feature matrix, and the covariance matrix is ​​obtained based on the transposed multimodal feature matrix and the standardized multimodal feature matrix.

[0076] Eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvectors and eigenvalues. The eigenvalues ​​are sorted from largest to smallest and a preset number of the largest eigenvalues ​​are selected. The eigenvectors corresponding to the selected eigenvalues ​​are used as principal component eigenvectors.

[0077] The principal component matrix is ​​obtained by stacking the principal component feature vectors column by column. Then, the multimodal feature matrix is ​​mapped and projected onto the principal component matrix to obtain the dimensionality-reduced feature matrix.

[0078] Furthermore, the preliminary prediction model is a machine learning algorithm model based on the financial market, including but not limited to linear regression models, decision tree models, support vector machines, or random forests.

[0079] Furthermore, the steps to obtain the trained risk prediction model include:

[0080] The initial hybrid prediction model is trained using a dimensionality-reduced feature matrix. After training, the validation error of the results is obtained, and it is determined whether the minimum validation error is greater than a preset error threshold.

[0081] If yes, then according to the preset rules, reselect the target matching pairs of the first and second regularization terms in the composite regularization terms to construct a new preliminary mixed prediction model, and return to the training process to train the newly constructed preliminary mixed prediction model; otherwise, output the current preliminary mixed prediction model to obtain the trained risk prediction model.

[0082] Furthermore, based on the risk prediction results and the preset risk result threshold range, the predicted risk level of the target financial market is output, specifically as follows:

[0083] Determine the relationship between the risk prediction results and the preset threshold range;

[0084] If the risk prediction result is greater than the upper limit of the preset threshold range, the target financial market is judged to have a high level of risk.

[0085] If the risk prediction result is within the preset threshold range, the risk level of the target financial market is determined to be moderate.

[0086] If the risk prediction result is less than the lower limit of the preset threshold range, the risk level of the target financial market is determined to be low.

[0087] like Figure 2 As shown, Figure 2 A flowchart illustrating an example of an artificial intelligence-based financial risk prediction method provided in this embodiment of the invention. The example process includes:

[0088] S1. Obtain historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market, and construct a multimodal feature matrix based on the historical transaction data, historical macroeconomic indicators, and historical corporate financial report data;

[0089] S2. Obtain the covariance matrix based on the multimodal feature matrix, and perform principal component analysis on the covariance matrix to reduce its dimensionality, thereby obtaining the dimensionality-reduced feature matrix.

[0090] S3. Obtain a preliminary forecasting model for the financial market and multiple compound regularization terms;

[0091] S4. Select composite regularization terms according to preset rules, and use the first and second regularization terms in the selected composite regularization terms as target L1 regularization terms and target L2 regularization terms.

[0092] S5. Introduce objective L1 regularization and objective L2 regularization terms into the preliminary prediction model to obtain a preliminary hybrid prediction model;

[0093] S6. Cross-validate the preliminary hybrid prediction model based on the dimensionality-reduced feature matrix, output the trained preliminary hybrid prediction model, and obtain the risk prediction model.

[0094] S7. Obtain real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data of the target financial market, and input these data into the risk prediction model to obtain the risk prediction results.

[0095] Furthermore, step S1, which involves constructing a multimodal feature matrix based on historical transaction data, historical macroeconomic indicators, and historical corporate financial statements, includes:

[0096] S11. Extract market transaction dynamic features from historical transaction data and obtain the transaction timestamps of the market transaction dynamic features;

[0097] S12. Extract market behavior analysis features from historical macroeconomic indicators and obtain the behavior timestamps of the market behavior analysis features;

[0098] S13. Align and transform the market behavior analysis features according to the behavior timestamp and the transaction timestamp to obtain the market behavior analysis features that are dynamically synchronized with the transaction.

[0099] S14. Extract the financial health characteristics of enterprises from historical enterprise financial report data and obtain the financial timestamps of the financial health characteristics of enterprises.

[0100] S15. Align and transform the enterprise's financial health characteristics according to the financial timestamp and the transaction timestamp to obtain the enterprise's financial health characteristics that are dynamically synchronized with the transaction.

[0101] S16. The market transaction dynamic characteristics, the market behavior analysis characteristics synchronized with the transaction dynamics, and the corporate financial health characteristics are standardized and normalized in sequence to obtain the corresponding synchronized market transaction dynamic characteristics, synchronized converted market behavior analysis characteristics, and synchronized converted corporate financial characteristics.

[0102] S17. The synchronous market transaction dynamic features, synchronous conversion market behavior analysis features, and synchronous conversion enterprise financial features are concatenated in chronological order according to the transaction timestamps to obtain a multimodal feature matrix.

[0103] Specifically, this invention extracts dynamic market transaction features from historical transaction data and obtains the transaction timestamps of these features. Extracting these dynamic features allows for a deeper understanding of market operating patterns and short-term behavior, such as trading volume and price fluctuations. These features directly reflect the market's immediate volatility and potential risks. Extracting transaction timestamps helps accurately track the temporal changes in market dynamics, enabling the model to understand and predict short-term market trends without ignoring the timeliness of market changes. In financial risk prediction, accurately capturing dynamic market transaction features improves the model's sensitivity to market fluctuations, thereby avoiding overfitting caused by excessive reliance on a single data source. This is achieved by extracting market behavior features from historical macroeconomic indicators. To analyze features and obtain behavioral timestamps for market behavior analysis features, macroeconomic indicators (such as GDP, interest rates, unemployment rates, etc.) provide broader background information on market behavior. Market behavior is influenced by the macroeconomic environment, and macroeconomic features can provide a more robust perspective, helping to avoid overfitting predictions based solely on local market features. The extracted behavioral timestamps enable the market behavior analysis features to be aligned with market transaction dynamics in time, enhancing the timeliness correlation between features and ensuring that the macroeconomic impact at different points in time can be accurately incorporated into the model. By aligning and transforming market behavior analysis features according to behavioral timestamps and transaction timestamps, transformed market behavior analysis features synchronized with transaction dynamics are obtained. By extracting corporate financial health characteristics from historical corporate financial reports and obtaining their financial timestamps, these characteristics can help capture internal risk signals (such as debt ratio, profitability, and cash flow). These factors are intrinsic drivers of market dynamics. Through this data, the model can not only see external market factors but also integrate the company's own financial situation, reducing the risk of overfitting by relying solely on market data. The extraction of financial timestamps ensures synchronization with other features over time, thereby enhancing the model's timely understanding of the company's health status. By aligning and transforming corporate financial health characteristics according to transaction timestamps, transaction-synchronized corporate financial health characteristics are obtained.Aligning and transforming the timestamps of different features ensures the consistency of various features over time. This means that all features can reflect the market situation at the same point in time, avoiding prediction errors caused by time mismatch. Time alignment also avoids the "timeliness" problem that may be caused by different feature time spans, enabling the model to better identify the relationships between features and thus avoid over-reliance on a single time window of historical data. By standardizing and normalizing the market transaction dynamic features, the market behavior analysis features synchronized with the transaction dynamics, and the corporate financial features synchronized with the transaction dynamics, we obtain the corresponding synchronized market transaction dynamic features, synchronized transformed market behavior analysis features, and synchronized transformed corporate financial features. By concatenating the synchronized market transaction dynamic features, synchronized transformed market behavior analysis features, and synchronized transformed corporate financial features according to the order of transaction timestamps, we obtain a multimodal feature matrix. Standardization and normalization are effective methods to avoid the unbalanced impact of different features on the model due to scale differences. Through these processes, all features are converted to the same dimensions and range, ensuring that the model gives equal importance to all types of features, thereby avoiding overfitting of the model to certain features. Standardization and normalization processes enable different types of features to exert equal influence within the same model, reducing overfitting caused by excessive reliance on certain features (such as the dramatic fluctuations in market transaction dynamics). By concatenating multiple features (market transaction dynamics, market behavior analysis, and corporate financial health) in chronological order into a multimodal feature matrix, the advantages of information from different data sources can be leveraged to improve the comprehensiveness and accuracy of the model. Multimodal fusion not only enhances the understanding of market information at different levels but also strengthens the model's predictive ability, avoiding overfitting caused by a single feature. This concatenation process ensures the synergistic effect between multiple information sources, enabling the model to learn the complex relationships between multiple feature interactions, thereby avoiding overfitting predictions based solely on one type of information.

[0104] Further, step S2, which involves obtaining the covariance matrix based on the multimodal feature matrix and performing dimensionality reduction on the covariance matrix to obtain the dimensionality-reduced feature matrix, includes:

[0105] S21. Standardize each feature in the multimodal feature matrix to obtain a standardized multimodal feature matrix;

[0106] S22. Transpose the standardized multimodal feature matrix to obtain the transposed multimodal feature matrix, and obtain the covariance matrix based on the transposed multimodal feature matrix and the standardized multimodal feature matrix.

[0107] S23. Perform eigenvalue decomposition on the covariance matrix to obtain multiple eigenvectors and eigenvalues;

[0108] S24. Sort multiple feature values ​​from largest to smallest, and select the feature vectors corresponding to a preset number of feature values ​​as principal component features;

[0109] S25. Stack the principal component features column by column to construct the principal component matrix, and map and project the multimodal feature matrix onto the principal component matrix to obtain the dimensionality-reduced feature matrix.

[0110] Specifically, this invention obtains a standardized multimodal feature matrix by standardizing each feature in the multimodal feature matrix. Each feature is standardized to have a mean of 0 and a variance of 1, ensuring consistent feature scale. Different features may have different dimensions and ranges; direct use could lead to some features dominating the prediction model. Standardization ensures that different features contribute equally to the model, avoiding the impact of scale differences on the model results. Standardized data helps improve the convergence speed of optimization algorithms such as gradient descent. If some features have large numerical ranges, they can disproportionately affect the model weights. Standardization makes the contributions of all features to the model more balanced, reducing the risk of overfitting. The standardized multimodal feature matrix is ​​then further processed... Row transpose yields the transposed multimodal feature matrix, which simplifies subsequent covariance matrix calculation, making the computation more efficient. The transpose operation ensures that each column corresponds to a feature and each row to a sample, making covariance matrix calculation more consistent with common machine learning theoretical frameworks. The covariance matrix is ​​obtained from the transposed and standardized multimodal feature matrices. The covariance matrix is ​​then calculated based on the standardized feature matrix and its transpose. The covariance matrix reveals the correlation between different features. By capturing this correlation, the model can better understand the relationships between features, thus reducing redundant information during dimensionality reduction and prediction. Through covariance matrix calculation, highly correlated features can be identified, which is crucial for dimensionality reduction. During the process, these features are merged or eliminated to reduce collinearity and improve model stability. Eigenvalue decomposition of the covariance matrix yields multiple eigenvectors and eigenvalues. Eigenvalues ​​represent the variance of each principal component in the covariance matrix; larger eigenvalues ​​indicate greater importance of that component in the data, and vice versa. Through decomposition, the most representative features of the data can be identified. By selecting the eigenvectors corresponding to larger eigenvalues, the most informative parts of the data can be retained, reducing noise and unimportant information, thereby reducing model complexity and preventing overfitting. Multiple eigenvalues ​​are sorted from largest to smallest, and eigenvectors corresponding to a predetermined number of eigenvalues ​​are selected as principal component features. The eigenvectors corresponding to the largest eigenvalue are then selected. Principal component features (PCMs) can retain the most significant variations in data. These features typically represent the core information of the data, thus avoiding excessive redundant information, preventing model complexity, reducing interference from irrelevant features, improving the model's predictive ability on unknown data, and avoiding overfitting to training data. PCMs are constructed by stacking PCMs column-wise. By stacking important PCMs, the high-dimensional feature space can be compressed into a low-dimensional space, making the data simpler, reducing computational and storage overhead, and improving processing efficiency, especially for large-scale datasets. Dimensionality reduction significantly improves the speed of model training and inference. Furthermore, the multimodal feature matrix is ​​mapped and projected onto the PCM matrix to obtain the dimensionality-reduced feature matrix.Mapping can significantly reduce the dimensionality of the feature space, which helps simplify the model and speed up training. It also reduces data noise. The reduced number of features in the dimensionality-reduced model decreases the chance of the model learning unimportant or noisy features, effectively preventing overfitting. Dimensionally reduced data typically has lower noise and redundancy, better reflecting the essential patterns of the data and enhancing the model's stability under different environments.

[0111] Further, step S5, which introduces objective L1 regularization and objective L2 regularization terms into the preliminary prediction model to obtain the preliminary hybrid prediction model, includes:

[0112] S51. Obtain the target loss function of the preliminary prediction model, and then obtain the first derivative of the target loss function with respect to the model output, and the second derivative with respect to the model parameters.

[0113] Specifically, the preliminary prediction model is obtained as follows: ;in, This indicates the preliminary predicted output. Indicates the model input, Indicates the initial model parameters. This represents the initial bias term; the objective loss function is: ;in, Indicates the model loss. Indicate quantity, Represents the actual value. Indicates the predicted value. Indicates the serial number.

[0114] S52. Obtain the second derivative of the target loss function with respect to the parameters of the preliminary prediction model, and obtain the gradient of the loss function based on the second derivative and the first derivative;

[0115] Specifically, the objective loss function after introducing the regularization term is: ;in, This represents the target loss function after introducing the regularization term; the expression for the target L1 regularization term (the first regularization term) is: The L1 gradient (gradient of the first regularization term) is The expression for the target L2 regularization term (second regularization term) is: The L2 gradient (gradient of the second regularization term) is 2. .

[0116] S53. Introduce the target L1 regularization term and the target L2 regularization term into the preliminary prediction model in sequence, and obtain the L1 gradient of the target L1 regularization term and the L2 gradient of the target L2 regularization term respectively.

[0117] S54. Update the parameters of the preliminary prediction model based on the L1 gradient, L2 gradient and loss function gradient to obtain the preliminary hybrid prediction model.

[0118] Specifically, the updated preliminary mixed prediction model is as follows: ;in, This represents the prediction output of the updated preliminary mixed prediction model. Indicates the model input, This represents the updated model parameters. This represents the updated bias term.

[0119] Specifically, this invention obtains the target loss function through a preliminary prediction model. The target loss function is typically used to measure the error between the model's prediction and the actual value. This function quantifies the model's performance. In financial risk prediction, the loss function effectively captures model bias, helps quantify the model's prediction error and generalization ability, and thus provides direction for optimization. Furthermore, it obtains the first derivative of the target loss function with respect to the output of the preliminary prediction model. The first derivative reflects the sensitivity of the loss function to changes in the model output, helping us find the direction and rate of change of the loss function. In financial risk prediction, the model's output needs to respond quickly to market changes, and the first derivative can optimize the sensitivity and stability of the output, avoiding overfitting while maintaining the model's generalization ability. The ability to obtain the second derivative of the target loss function with respect to the initial model parameters is crucial. This second derivative captures the acceleration of the loss function's change, helping to determine the rate and direction of parameter updates. In complex financial risk prediction, excessive gradient updates can lead to overfitting on the training set. Adjusting the second derivative allows for smoother parameter updates, avoiding unnecessary oscillations or over-optimization, further controlling overfitting. The loss function gradient is obtained based on the first and second derivatives. By combining the first derivative of the model output with the second derivative of the model parameters, the loss function gradient can more accurately guide the parameter optimization process. This optimization method avoids the local minima problem in simple gradient descent algorithms, improving model stability. In terms of performance and convergence speed, this gradient update method, which combines the first derivative of the model output and the second derivative of the model parameters, can improve the model's adaptability to complex market conditions, especially in financial risk prediction where data noise is high. By sequentially introducing a target L1 regularization term and a target L2 regularization term into the initial prediction model, and obtaining the L1 gradient of the target L1 regularization term and the L2 gradient of the target L2 regularization term respectively, L1 regularization can cause some parameters to become zero, thereby enabling feature selection during model training and reducing the influence of unimportant features. L2 regularization helps prevent excessively large parameters, thus avoiding overfitting of the model to the training data. By introducing L1 and L2 regularization terms, model complexity can be controlled while preventing parameter overfitting. Overfitting, especially in financial data, often involves significant noise and irrelevant features. Regularization can effectively mitigate the impact of these factors. L1 and L2 regularization gradients help constrain model parameters in different ways, providing two distinct methods to control model complexity and improve generalization ability. Particularly in high-noise data for financial risk prediction, this strategy effectively reduces the risk of overfitting. By updating the parameters of the initial prediction model using L1, L2, and loss function gradients, a preliminary mixed prediction model is obtained. Combining the updates of L1, L2, and loss function gradients during optimization not only reduces overfitting but also makes the parameters more stable.To prevent models from becoming overly sensitive to the training set, in financial risk prediction, model performance should not rely excessively on training data but should possess good generalization ability. Combining multiple gradient optimization methods can effectively improve model performance on unseen data. By integrating multiple gradient updates and regularization adjustments, the resulting preliminary hybrid prediction model exhibits strong generalization ability and can better cope with the volatility of financial markets. Hybrid models typically utilize the advantages of multiple features or models to mitigate the biases that may arise from single models, improving prediction accuracy and stability, and better predicting complex market risks.

[0120] Further, step S6, which involves cross-validating the preliminary hybrid prediction model based on the dimensionality-reduced feature matrix to obtain the risk prediction model, includes:

[0121] S61. Divide the dimensionality-reduced feature matrix into multiple data subsets in equal proportions, select one data subset from the multiple data subsets as the validation set, and use the remaining data subsets as the training set;

[0122] S62. Train the preliminary mixed prediction model based on the training set and validation set until each data subset is used as a validation set for training.

[0123] S63. Obtain the verification error of each training result in real time, and determine whether the minimum verification error is greater than the preset error.

[0124] If the minimum verification error is greater than the preset error, return to the step of selecting a first regularization term and a second regularization term from the composite regularization term according to the preset rules as the target L1 regularization term and the target L2 regularization term, until the minimum verification error is not greater than the preset error.

[0125] S64. Take the target L1 regularization term and the target L2 regularization term at this time as the final L1 regularization term and the final L2 regularization term, and take the model parameter corresponding to the minimum verification error at this time as the final model parameter.

[0126] S65. Adjust the preliminary hybrid prediction model according to the final model parameters to obtain the risk prediction model.

[0127] Specifically, this invention divides the dimensionality-reduced feature matrix into multiple data subsets proportionally. One subset is selected as the validation set, and the remaining subsets are used as the training set. By dividing the data into multiple subsets, it ensures that the model does not merely memorize the details of the training data, thus avoiding the risk of overfitting. Using different data subsets for each training iteration enhances the model's generalization ability. All data subsets are used for both training and validation, improving data utilization efficiency and preventing any data waste. Different subsets may contain different samples, allowing for better verification of the model's adaptability to different types of data and preventing biased learning on certain data. Cross-validation effectively evaluates the model's performance on different data, thereby... By identifying potential overfitting issues and implementing multiple training and validation cycles, model bias caused by different data splits was reduced, resulting in a more adaptable final model to all data. Alternating between different subsets stabilized the validation error, preventing overfitting. The initial mixed prediction model was trained using both training and validation sets until each data subset was used as a validation set. The validation error was monitored in real-time for each training result, and the minimum validation error was determined to be greater than a preset error. Real-time monitoring of the validation error allowed for timely detection of overfitting or underfitting, enabling dynamic adjustments to model parameters or training methods to avoid overfitting during training. Overfitting is achieved by setting a preset error threshold to ensure the training process remains within a reasonable range, preventing excessively large or small errors that could affect model stability. If the minimum validation error exceeds the preset error, model parameters can be adjusted promptly for targeted optimization, avoiding unnecessary complexity during training. If the minimum validation error is greater than the preset error, the process returns to selecting a first and second regularization term from the composite regularization terms according to preset rules as the target L1 and target L2 regularization terms, respectively, until the minimum validation error is no greater than the preset error. By adjusting the regularization terms based on the validation error after each iteration, overfitting due to excessive complexity during model training is effectively avoided. Optimizing the composite regularization term allows for flexible adjustment of the regularization strength based on actual performance during model training, avoiding over-constraint or over-relaxation of the regularization parameters. Continuously adjusting the regularization term during training maintains sufficient complexity while mitigating the model's sensitivity to noise, enhancing its robustness across diverse data. By selecting appropriate target L1 and target L2 regularization terms, the model's complexity can be adaptively adjusted, preventing it from becoming too complex or too simple, thus reducing the possibility of overfitting or underfitting. Continuous optimization until the validation error is no greater than a preset error ensures that the model consistently moves towards reducing error during training, avoiding the risk of overfitting. The preset error serves as a standard.This method allows for precise control of the error range during training, ensuring the model ultimately meets the requirements of practical applications. By using the target L1 and L2 regularization terms as the final L1 and L2 regularization terms, and the model parameters corresponding to the minimum validation error as the final model parameters, the initial mixed prediction model is adjusted to obtain the risk prediction model. Selecting the model parameters corresponding to the minimum validation error ensures the final model performs optimally in practical applications, reducing overfitting. The role of L1 and L2 regularization in the final model helps control model complexity, ensuring it captures effective patterns in the data without overlearning noise. Adjusting the initial mixed prediction model ensures consistent performance across multiple data subsets, further reducing the possibility of overfitting. After multiple training and adjustments, the final risk prediction model demonstrates stronger predictive power and stability when processing financial data. The resulting risk prediction model provides more accurate prediction results, offering reliable support for financial decision-making and reducing decision bias caused by overfitting or data inconsistency.

[0128] Furthermore, after step S6, which involves inputting real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data into the risk prediction model to obtain the risk prediction results, the process also includes:

[0129] S66. Determine the relationship between the risk prediction results and the preset threshold range;

[0130] If the risk prediction result is greater than the upper limit of the preset threshold range, the target financial market is judged to have a high level of risk.

[0131] If the risk prediction result is within the preset threshold range, the risk level of the target financial market is determined to be moderate.

[0132] If the risk prediction result is less than the lower limit of the preset threshold range, the risk level of the target financial market is determined to be low.

[0133] In summary, this invention constructs a multimodal feature matrix using historical transaction data, historical macroeconomic indicators, and historical corporate financial report data from the target financial market. By comprehensively considering historical data from multiple aspects such as the financial market, macroeconomics, and corporate financial reports, it helps to establish a more comprehensive feature matrix that reflects the complexity and diversity of the market. The multimodal feature matrix integrates different types of data, capturing multi-dimensional alignment information from the financial market, providing more predictive signals, and avoiding the limitations of a single data source. The covariance matrix is ​​obtained through the multimodal feature matrix, and dimensionality reduction processing is performed to obtain a dimensionality-reduced feature matrix. The covariance matrix helps reveal the relationships between various features, identifying the correlations and dependencies between different features. For financial markets, the interaction and correlation between different features are crucial for risk prediction. Furthermore, the covariance matrix helps capture the volatility relationships between different financial indicators, which is essential for modeling market volatility and risk. Dimensionality reduction can remove redundant or unnecessary information, reduce data noise, improve the model's generalization ability, and avoid overfitting. Dimensionality reduction reduces the dimensionality of the data, making subsequent calculations more efficient and avoiding wasted computational resources. After dimensionality reduction, the model can focus more on the most important features, reducing excessive attention to details and thus avoiding overfitting. This is achieved by obtaining a preliminary prediction model of the financial market and multiple first and second regularization terms of the composite regularization term, and selecting one from the composite regularization term according to preset rules. The first and second regularization terms are used as the target L1 and target L2 regularization terms, respectively. The introduction of L1 and L2 regularization in the hybrid model further constrains the model parameters, prevents overfitting, and makes the model focus more on the true signal rather than noise. By introducing the target L1 and target L2 regularization terms into the initial prediction model, a preliminary hybrid prediction model is obtained. L1 regularization helps with feature selection, compressing the coefficients of unimportant features to zero, thereby simplifying the model and reducing the risk of overfitting. L2 regularization helps smooth the model, making parameter estimation more stable and preventing the model from being overly sensitive to noise in the training data. This strategy automatically selects the most suitable regularization terms (L1 and L2) through preset rules, making... The model can adaptively adjust its complexity based on data characteristics, improving prediction accuracy. L1 and L2 regularization each have their advantages: L1 allows for feature selection, while L2 prevents excessive weights. Combining both allows for better control of model complexity and reduces overfitting. A risk prediction model is obtained by cross-validating the initial hybrid prediction model using a dimensionality-reduced feature matrix. Cross-validation, by dividing the data into training and validation sets, ensures the model performs well on multiple subsets, thus improving its generalization ability and preventing overfitting to only a specific dataset. Cross-validation comprehensively evaluates the model's stability and accuracy, verifying its performance from multiple perspectives and reducing its dependence on a single dataset.The model can adjust its parameters based on validation set results to further improve prediction accuracy. It acquires real-time transaction data, macroeconomic indicators, and corporate financial statements from the target financial market and inputs these data into the risk prediction model to obtain prediction results. By using real-time data, the model can dynamically adapt to market changes, improving the timeliness and accuracy of predictions. Real-time data helps predict the latest market risks, providing timely warnings and enabling decision-makers to react quickly to potential financial crises. The model updates itself with the input of real-time data, enhancing its adaptability to market changes. This ensures that the model can not only handle complex and diverse market data but also adapt to dynamic market changes, thereby better enabling the prediction and control of financial risks.

[0134] The implementation of the various embodiments of this invention is based on programmed processing through a system with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of this invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of this invention provide an artificial intelligence-based financial risk prediction system, which is used to execute one of the artificial intelligence-based financial risk prediction methods described in the above embodiments of the artificial intelligence-based financial risk prediction method.

[0135] See Figure 3 The system includes:

[0136] The risk prediction module is used to input real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data from the target financial market into the trained risk prediction model to obtain risk prediction results. The training process of the risk prediction model includes: constructing a multimodal feature matrix based on historical transaction data, historical macroeconomic indicators, and historical corporate financial report data from the target financial market; then processing the multimodal feature matrix using matrix operations and principal component analysis to obtain a dimensionality-reduced feature matrix; and constructing a preliminary hybrid prediction model, which is a preliminary prediction model updated based on the parameters of a composite regularization term. The construction steps are as follows: First, obtain the target loss function of the preliminary prediction model. Then, obtain the first derivative of the target loss function with respect to the model output and the second derivative with respect to the model parameters. Combine the first and second derivatives to obtain the gradient of the loss function. Next, sequentially introduce composite regularization terms into the preliminary prediction model. These composite regularization terms are matching pairs of the first and second regularization terms, and obtain the gradients of the first and second regularization terms in the composite regularization terms respectively. Update the parameters of the preliminary prediction model based on the gradients of the first and second regularization terms and the loss function to obtain the preliminary hybrid prediction model. Select composite regularization terms sequentially according to preset rules to construct different preliminary hybrid prediction models. Simultaneously, use a dimensionality-reduced feature matrix to sequentially train and cross-validate the preliminary hybrid prediction models until the model output meets preset requirements. Output the model at this point to obtain the trained risk prediction model.

[0137] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other artificial intelligence-based financial risk prediction method embodiments provided by the present invention. The only difference is the setting of corresponding functional modules. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the modules in the above system embodiments by referring to the specific technical solutions in other method embodiments, combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, and ensuring the practicality of the technical solutions, they can obtain corresponding system-class embodiments for implementing the methods in other method-class embodiments.

[0138] Based on the above system embodiments, as a preferred embodiment, this invention provides an artificial intelligence-based financial risk prediction system, which further includes:

[0139] Furthermore, such as Figure 4As shown, the training process of the risk prediction model is completed within the pre-training module, which includes: a first acquisition module, used to acquire historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market, and construct a multimodal feature matrix based on the historical transaction data, historical macroeconomic indicators, and historical corporate financial report data; a dimensionality reduction module, used to acquire the covariance matrix based on the multimodal feature matrix constructed by the first acquisition module, and perform principal component analysis on the covariance matrix to reduce its dimensionality, thus obtaining a dimensionality-reduced feature matrix; a second acquisition module, used to acquire a preliminary financial market prediction model and a composite regularization term; a selection module, used to select the composite regularization term acquired by the second acquisition module according to preset rules; a preliminary mixed prediction model construction module, used to introduce the composite regularization term selected by the selection module into the preliminary financial market prediction model acquired by the second acquisition module to obtain a preliminary mixed prediction model; and a cross-validation module, used to train and cross-validate the preliminary mixed prediction model obtained by the preliminary mixed prediction model construction module based on the dimensionality-reduced feature matrix obtained by the dimensionality reduction module. When the model output meets preset requirements, the trained risk prediction model is obtained.

[0140] Among them, such as Figure 5 As shown, the dimensionality reduction module includes: a preprocessing unit, used to standardize each feature in the multimodal feature matrix to obtain a standardized multimodal feature matrix; a transpose unit, used to transpose the standardized multimodal feature matrix obtained by the preprocessing unit to obtain a transposed multimodal feature matrix, and obtain a covariance matrix based on the transposed multimodal feature matrix and the standardized multimodal feature matrix; a principal component filtering unit, used to perform eigenvalue decomposition on the covariance matrix obtained by the transpose unit to obtain multiple eigenvectors and eigenvalues, sort the eigenvalues ​​from largest to smallest and filter a preset number of the largest eigenvalues, and use the eigenvectors corresponding to the filtered eigenvalues ​​as principal component eigenvectors; and a mapping unit, used to stack the principal component eigenvectors obtained by the principal component filtering unit column-wise to obtain a principal component matrix, and then map and project the multimodal feature matrix onto the principal component matrix to obtain a dimensionality-reduced feature matrix.

[0141] The method in this embodiment of the invention is implemented using an electronic device; therefore, it is necessary to introduce the relevant electronic device. For this purpose, embodiments of the present invention provide an electronic device, such as... Figure 6 As shown, the electronic device includes: at least one processor, a communication interface, at least one memory, and a communication bus, wherein the at least one processor, the communication interface, and the at least one memory communicate with each other via the communication bus. The at least one processor invokes logical instructions stored in the at least one memory to execute all or part of the steps of the methods provided in the foregoing method embodiments.

[0142] Furthermore, when the logical instructions in at least one of the aforementioned memories are implemented as software functional units and sold or used as independent products, they are stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, is embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (a personal computer, server, or network device) to execute all or part of the steps of the methods described in the various method embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks—various media for storing program code.

[0143] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, located in one place, or distributed across multiple network units. The purpose of this embodiment is achieved by selecting some or all of the modules according to actual needs. Those skilled in the art will understand and implement this without any inventive effort.

[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as an artificial intelligence-based financial risk prediction method, system, or computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of an artificial intelligence-based financial risk prediction method, apparatus (system), and computer program product according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A financial risk prediction method based on artificial intelligence, characterized in that, include: Input real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data from the target financial market into the trained risk prediction model to obtain risk prediction results. Based on the risk prediction results and the preset risk result threshold range, the predicted risk level of the target financial market is output. The training process of the risk prediction model includes: A multimodal feature matrix is ​​constructed based on historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market. Then, the multimodal feature matrix is ​​processed by matrix operations and principal component analysis to obtain a dimension-reduced feature matrix. A preliminary hybrid prediction model is constructed, which is a preliminary prediction model based on the updated parameters of the composite regularization term. The construction steps are as follows: Obtain the target loss function of the preliminary prediction model, then obtain the first derivative of the target loss function with respect to the model output, and the second derivative with respect to the model parameters; The gradient of the loss function is obtained by combining the first and second derivatives; A composite regularization term is sequentially introduced into the preliminary prediction model. The composite regularization term is a matching pair of the first regularization term and the second regularization term. The gradients of the first regularization term and the second regularization term in the composite regularization term are obtained respectively. The parameters of the preliminary prediction model are updated based on the gradients of the first regularization term, the second regularization term, and the loss function to obtain the preliminary hybrid prediction model. Different preliminary mixed prediction models are constructed by sequentially selecting composite regularization terms according to preset rules. Simultaneously, the preliminary mixed prediction models are trained and cross-validated using dimensionality reduction feature matrices until the model output meets the preset requirements. The trained risk prediction model is then output.

2. The financial risk prediction method based on artificial intelligence as described in claim 1, characterized in that, The steps for constructing the multimodal feature matrix include: Extract market transaction dynamics and transaction timestamps from historical transaction data; Market behavior analysis features and their behavior timestamps are extracted from historical macroeconomic indicators. Then, based on the alignment of behavior timestamps with transaction timestamps, the market behavior analysis features are converted into market behavior analysis features that are synchronized with transaction dynamics. Extract the financial health characteristics and financial timestamps of enterprises from historical corporate financial reports, and then convert the financial health characteristics into enterprise financial health characteristics that are dynamically synchronized with transaction timestamps based on the alignment of financial timestamps and transaction timestamps. The market transaction dynamics, market behavior analysis features synchronized with transaction dynamics, and corporate financial health features are standardized and normalized sequentially. Then, they are spliced ​​and arranged according to the transaction timestamps and their order to obtain a multimodal feature matrix.

3. The financial risk prediction method based on artificial intelligence as described in claim 1, characterized in that, The steps for constructing the dimensionality reduction features are as follows: Each feature in the multimodal feature matrix is ​​standardized to obtain a standardized multimodal feature matrix; The standardized multimodal feature matrix is ​​transposed to obtain the transposed multimodal feature matrix, and the covariance matrix is ​​obtained based on the transposed multimodal feature matrix and the standardized multimodal feature matrix. Eigenvalue decomposition is performed on the covariance matrix to obtain multiple eigenvectors and eigenvalues. The eigenvalues ​​are sorted from largest to smallest and a preset number of the largest eigenvalues ​​are selected. The eigenvectors corresponding to the selected eigenvalues ​​are used as principal component eigenvectors. The principal component matrix is ​​obtained by stacking the principal component feature vectors column by column. Then, the multimodal feature matrix is ​​mapped and projected onto the principal component matrix to obtain the dimensionality-reduced feature matrix.

4. The financial risk prediction method based on artificial intelligence as described in claim 1, characterized in that, The preliminary prediction model is a machine learning algorithm model based on the financial market, including but not limited to linear regression model, decision tree model, support vector machine or random forest.

5. The financial risk prediction method based on artificial intelligence as described in claim 4, characterized in that, The steps for obtaining the trained risk prediction model include: The initial hybrid prediction model is trained using a dimensionality-reduced feature matrix. After training, the validation error of the results is obtained, and it is determined whether the minimum validation error is greater than a preset error threshold. If yes, then according to the preset rules, reselect the target matching pairs of the first and second regularization terms in the composite regularization terms to construct a new preliminary mixed prediction model, and return to the training process to train the newly constructed preliminary mixed prediction model; otherwise, output the current preliminary mixed prediction model to obtain the trained risk prediction model.

6. A financial risk prediction system based on artificial intelligence, characterized in that, include: The risk prediction module is used to input real-time transaction data, real-time macroeconomic indicators, and real-time corporate financial report data from the target financial market into the trained risk prediction model to obtain risk prediction results. The risk rating module is used to output the risk level prediction results of the target financial market based on the risk prediction results and the preset risk result threshold range. The training process of the risk prediction model includes: A multimodal feature matrix is ​​constructed based on historical transaction data, historical macroeconomic indicators, and historical corporate financial report data of the target financial market. Then, the multimodal feature matrix is ​​processed by matrix operations and principal component analysis to obtain a dimension-reduced feature matrix. A preliminary hybrid prediction model is constructed, which is a preliminary prediction model based on the updated parameters of the composite regularization term. The construction steps are as follows: Obtain the target loss function of the preliminary prediction model, then obtain the first derivative of the target loss function with respect to the model output, and the second derivative with respect to the model parameters; The gradient of the loss function is obtained by combining the first and second derivatives; A composite regularization term is sequentially introduced into the preliminary prediction model. The composite regularization term is a matching pair of the first regularization term and the second regularization term. The gradients of the first regularization term and the second regularization term in the composite regularization term are obtained respectively. The parameters of the preliminary prediction model are updated based on the gradients of the first regularization term, the second regularization term, and the loss function to obtain the preliminary hybrid prediction model. Different preliminary mixed prediction models are constructed by sequentially selecting composite regularization terms according to preset rules. Simultaneously, the preliminary mixed prediction models are trained and cross-validated using dimensionality reduction feature matrices until the model output meets the preset requirements. The trained risk prediction model is then output.

7. The artificial intelligence-based financial risk prediction system as described in claim 6, characterized in that, The training process of the risk prediction model is completed within a pre-training module, which includes: The first acquisition module is used to acquire historical transaction data, historical macroeconomic indicators and historical corporate financial report data of the target financial market, and to construct a multimodal feature matrix based on the historical transaction data, historical macroeconomic indicators and historical corporate financial report data; The dimensionality reduction module is used to obtain the covariance matrix based on the multimodal feature matrix constructed by the first acquisition module, and to perform principal component analysis on the covariance matrix to reduce its dimensionality, thereby obtaining the dimensionality-reduced feature matrix. The second acquisition module is used to acquire the preliminary prediction model and compound regularization term for the financial market. The selection module is used to select the composite regularization terms obtained by the second acquisition module according to preset rules; The preliminary hybrid prediction model construction module is used to introduce the composite regularization term selected by the selection module into the preliminary financial market prediction model obtained by the second acquisition module to obtain the preliminary hybrid prediction model. The cross-validation module is used to train and cross-validate the preliminary mixed prediction model obtained by the preliminary mixed prediction model construction module based on the dimensionality reduction feature matrix obtained by the dimensionality reduction module. When the model output meets the preset requirements, the trained risk prediction model is obtained.

8. The artificial intelligence-based financial risk prediction system as described in claim 7, characterized in that, The dimensionality reduction module includes: The preprocessing unit is used to standardize each feature in the multimodal feature matrix to obtain a standardized multimodal feature matrix. The transpose unit is used to transpose the standardized multimodal feature matrix obtained by the preprocessing unit to obtain the transposed multimodal feature matrix, and to obtain the covariance matrix based on the transposed multimodal feature matrix and the standardized multimodal feature matrix. The principal component screening unit is used to perform eigenvalue decomposition on the covariance matrix obtained by the transpose unit to obtain multiple eigenvectors and eigenvalues, sort the eigenvalues ​​from largest to smallest and screen a preset number of the largest eigenvalues, and use the eigenvectors corresponding to the screened eigenvalues ​​as principal component eigenvectors. The mapping unit is used to stack the principal component feature vectors obtained by the principal component screening unit column by column to obtain the principal component matrix, and then map and project the multimodal feature matrix onto the principal component matrix to obtain the dimensionality-reduced feature matrix.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing program instructions that are executed by the processor, and the processor calling the program instructions to execute the artificial intelligence-based financial risk prediction method according to any one of claims 1 to 5.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the artificial intelligence-based financial risk prediction method according to any one of claims 1 to 5.

Images