A financial data processing system and method

Abstract

The application belongs to the field of financial data processing, and discloses a kind of financial data processing system and method, comprising: data acquisition module, for collecting the multi-source multi-modal data of listed company, multi-source multi-modal data includes enterprise internal financial data, enterprise internal non-financial data, social external index and macroeconomic variable;Pretreatment module;Characteristic screening module;Model construction and optimization module;Interpretability analysis module;Output module.This scheme constructs a set of multi-source data integration, automation processing, intelligent modeling and decision explanation in one financial early warning system.Through the standardization of module pipeline, the whole process closed loop processing is realized;At the same time, the high precision prediction and SHAP interpretability technology are fused, so that the model can output accurate risk probability, and the influence degree and direction of each risk factor can be clearly quantified, the traditional "black box" prediction is transformed into transparent, reliable and operable decision basis, and the unity of early warning accuracy and decision support is realized.

Description

Technical Field

[0001] This invention relates to the field of financial data processing, and more particularly to a financial data processing system and method. Background Technology

[0002] With the deepening application of fintech and big data technologies, building intelligent financial risk early warning models using corporate data has become an important trend.

[0003] Currently, most related research and practice revolve around specific financial ratio indicators, using traditional statistical models or single machine learning algorithms for prediction. Although these methods achieve risk quantification to some extent, their technical approach typically separates data collection, feature engineering, and model training, relying heavily on manual experience for data cleaning, feature selection, and parameter tuning. This results in low automation and poor reproducibility of the entire early warning process, and makes it difficult to cope with the massive, multi-source, and rapidly changing data environment of modern enterprises, exhibiting significant bottlenecks in real-time performance, stability, and intelligence. Summary of the Invention

[0004] The present invention aims to provide a financial data processing system and method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A financial data processing system, comprising: The data acquisition module is used to collect multi-source, multi-modal data from listed companies. This multi-source, multi-modal data includes internal financial data, internal non-financial data, external social indicators, and macroeconomic variables. The preprocessing module is used to perform missing value imputation, outlier handling, data balancing, and normalization on the collected data. The feature filtering module is used to filter out predictive indicators that have significant distinguishing ability against financial distress from the processed data based on statistical significance tests. The model building and optimization module is used to build a prediction model based on XGBoost as the classifier, and to use the Red-billed Blue Magpie optimization algorithm to perform global search optimization of the hyperparameters of the XGBoost model. The interpretability analysis module is used to perform feature importance analysis and attribution interpretation of prediction results on the optimized prediction model based on SHAP value theory. The output module is used to output the financial distress prediction results of listed companies and the corresponding risk factor analysis report.

[0007] Preferably, when the preprocessing module performs missing value imputation, it uses the K-nearest neighbor algorithm, with the Minkowski distance as the distance metric, and the calculation formula is as follows: ; Preferably, when the preprocessing module handles outliers, it uses a box plot method for judgment, and the judgment criterion is: if the data point satisfy or If it is, it is determined to be an outlier, and the data in that feature column is used. The value is replaced; where, and These are the lower and upper quartiles of the data, respectively. Interquartile range, The standard deviation is denoted as .

[0008] Preferably, when the preprocessing module performs data balancing, it uses the SMOTEENN hybrid sampling algorithm, wherein the formula for synthesizing new minority class samples in the SMOTE part is: ; in, For the selected minority class samples, It is a neighbor sample randomly selected from its K nearest neighbors. It is a random number within the interval [0,1].

[0009] Preferably, when the preprocessing module performs normalization, it adopts the minimum-maximum normalization method to linearly scale the value of each feature index to the [0,1] interval. The calculation formula is as follows: ; in, The normalized value. The original value, and These are the minimum and maximum values ​​of the feature in the dataset, respectively.

[0010] Preferably, the feature screening module screens predictive indicators based on significance tests, including: first, performing a Shapiro-Wilk normality test on each candidate indicator; for indicators that pass the normality test, using a paired T-test to determine whether the difference in mean between the financial distress group and the normal group is significant; for indicators that fail the normality test, using a Mann-Whitney U-test to determine whether the difference in distribution is significant; and finally, selecting indicators that have significant differences at a preset significance level to form a predictive indicator system.

[0011] Preferably, in the model building and optimization module, the loss function of the XGBoost classification model for binary classification problems is: ; in, For real labels, These are the raw scores predicted by the model.

[0012] Preferably, the position update strategy of the red-billed blue magpie optimization algorithm used in the model construction and optimization module includes a small group cooperation strategy and a group cooperation strategy, the mathematical expressions of which are as follows:

[0013] Small group collaboration strategies: ; Group collaboration strategies: ; in, For small group size, For group size, For a randomly selected search agent, and It is a random number.

[0014] Preferably, the interpretability analysis module, based on the SHAP value theory, decomposes the model prediction value of each sample into the sum of the contributions of each feature, the basic expression of which is: ; in, For the model to sample The predicted value, This represents the baseline expected value of the model across all samples. Features The SHAP value indicates the impact of this feature on the prediction result. The contribution of the value indicates that a positive value pushes the prediction toward the positive class, while a negative value does the opposite.

[0015] A financial data processing method includes the following steps: S1. Collect multi-source, multi-modal data from listed companies; S2. Preprocess the data, including using the KNN algorithm to fill in missing values, replacing outliers based on box plots, using the SMOTEENN algorithm to balance the classes, and performing min-max normalization. S3. Screening prediction indicators based on the significance test process of Shapiro-Wilk, T-test and Mann-Whitney U-test; S4. Construct an XGBoost prediction model and optimize its hyperparameters using the Red-billed Blue Magpie Optimization Algorithm. S5. Predict the financial distress of listed companies based on the optimized model; S6. Use SHAP value theory to interpret the prediction results and analyze the direction and intensity of the influence of each feature on the prediction results; S7. Output a report that includes a probability forecast of financial distress and an analysis of key risk factors.

[0016] The beneficial effects of this technical solution compared to existing technologies are as follows:

[0017] (1) This invention constructs a fully automated intelligent system for financial data processing and risk early warning, effectively solving the problems of fragmented processes, reliance on manual labor, and low efficiency in existing technologies. The system achieves full automation from raw multi-source data input to the generation of structured risk reports through the serial design and seamless connection of data acquisition, preprocessing, feature screening, model construction and optimization, interpretability analysis, and output modules. This design changes the situation in traditional solutions that rely on expert experience and frequent manual operations, significantly improving the efficiency, consistency, and scalability of risk identification and analysis, and can adapt to large-scale, continuously updated data application scenarios.

[0018] (2) This invention creatively integrates a metaheuristic optimization algorithm (Red-billed Blue Magpie Optimization Algorithm) with Explainable Artificial Intelligence (SHAP) technology into the financial early warning process, improving the accuracy of model predictions while ensuring the transparency and credibility of decision-making. The system does not simply apply existing models, but rather uses the Red-billed Blue Magpie Optimization Algorithm to adaptively optimize the key hyperparameters of the XGBoost model globally, ensuring that the model is always in its optimal performance state, thereby obtaining more accurate prediction results. Simultaneously, SHAP technology is used to quantitatively explain the prediction logic of the optimal model, making every risk judgment traceable. This dual guarantee of "accuracy optimization" and "explanatory transparency" makes the system not only an efficient prediction tool, but also a reliable intelligent platform that can provide clear decision-making basis and support in-depth business analysis. Attached Figure Description

[0019] Figure 1 The overall system architecture diagram provided by this invention; Figure 2 Flowchart of the financial data processing method provided by the present invention; Figure 3 This is a schematic diagram of the data acquisition module provided by the present invention; Figure 4 A flowchart of the preprocessing module provided by this invention; Figure 5 Flowchart of the feature filtering module provided by this invention; Figure 6 A schematic diagram of the model building and optimization module provided by this invention; Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:

[0021] like Figures 1 to 6This invention discloses a financial data processing system and method. The core design concept of this invention is to construct a fully automated financial distress early warning system that integrates "data input-processing-modeling-interpretation-output," achieving intelligent transformation from multi-source data to decision support through modular integration and closed-loop data flow.

[0022] The system adopts a loosely coupled modular design, with each module enabling real-time data flow through standardized data interfaces. The core modules include data acquisition, preprocessing, feature selection, model building and optimization, interpretability analysis, and output. Each module independently performs a specific function while also collaborating to complete the entire processing flow, ensuring the system's flexibility, scalability, and maintainability.

[0023] Data Acquisition Module: This module serves as the system's data input terminal. Its core objective is to comprehensively, timely, and accurately acquire multi-source, multi-modal data from listed companies, providing a complete data source for subsequent processing. The implementation process strictly adheres to three principles: "full coverage of data sources, adaptation of acquisition strategies, and verification of data quality."

[0024] Data source composition and acquisition method: Internal financial data: By accessing authoritative financial database interfaces, we regularly capture publicly disclosed balance sheets (total assets, total liabilities, owners' equity, and other core indicators), income statements (operating revenue, operating costs, net profit, and other indicators), and cash flow statements (net cash flow from the three major activities, and other indicators) of listed companies; at the same time, we configure a PDF parsing engine to automatically extract structured financial data from annual and quarterly report PDF files of the disclosure platforms designated by the China Securities Regulatory Commission, ensuring the authority of the data source.

[0025] Internal non-financial data of enterprises: Obtain corporate governance and operational data of listed companies through system interfaces; use natural language processing (NLP) technology to extract keywords (such as "declining performance" and "tight funds") from texts such as management discussion and analysis (MD&A) and earnings reports, and perform sentiment scoring (positive / negative / neutral) to quantify unstructured information.

[0026] External social indicators: Real-time crawling of company-related news through mainstream financial news APIs (such as Caixin.com and Eastmoney.com) to monitor negative events (such as penalties for violations and lawsuits); collection of investor discussion heat (number of posts and comments) on platforms such as stock forums and Xueqiu using web crawlers; acquisition of cooperation stability data of upstream and downstream enterprises from supply chain databases, combined with the competition intensity index (CRn) released by industry associations, to construct an external environment indicator system.

[0027] Macroeconomic variables: Connecting to the public data interfaces of the National Bureau of Statistics and the People's Bank of China, we regularly collect indicators such as GDP growth rate, CPI, PPI, 1-year LPR, money supply M2, and industry prosperity index, and update them quarterly to capture the systematic impact of the macro environment on corporate financial conditions.

[0028] Data collection strategy and data verification: Data collection mode: A dual mode of "scheduled batch collection + event-triggered collection" is adopted. Scheduled tasks are executed periodically (financial data quarterly, macroeconomic data monthly, and public opinion data daily) to ensure the timeliness of basic data; event triggers monitor key nodes such as listed company announcements and major news pushes, and capture sudden information in real time (such as temporary suspension announcements and major contract signings) to avoid data lag.

[0029] Data verification: After collection, the system automatically performs three verifications: format verification (to ensure that numerical and text data are correctly classified), integrity verification (to check that no required indicators are missing), and consistency verification (to compare the same indicator value from different data sources, and to trigger manual review if the deviation exceeds 5%). After the verification is passed, the data is stored in the system's preset data storage unit in a standardized format, supporting data traceability and backtracking query.

[0030] Preprocessing module: The presence of missing, anomalies, inconsistent units, and class imbalances in the original data can directly affect the accuracy of the model. This module outputs high-quality standardized data through systematic cleaning and transformation.

[0031] Missing value imputation: Set the nearest neighbor number K=5 (this can be preset within the range of 3-10 based on experience or feature dimensions, with K=5 being preferred); for samples containing missing values, calculate their Minkowski distance to all complete samples (formula: , where p=2 (Euclidean distance), quantifies sample similarity; the 5 closest complete samples are selected, numerical features are filled with weighted mean (the weight is the reciprocal of the distance, and the sum of the weights is 1), and categorical features are filled with mode to ensure that the filled values ​​fit the local distribution of the data.

[0032] Example: A sample is missing the "R&D investment ratio". The values ​​of this indicator for the 5 nearest neighbor samples are 4%, 5%, 6%, 5%, and 7%, with corresponding distance weights of 0.2, 0.25, 0.3, 0.15, and 0.1. The imputed value = 4%×0.2 + 5%×0.25 + 6%×0.3 + 5%×0.15 + 7%×0.1 = 5.3%.

[0033] Outlier handling: Sort each feature column in ascending order, calculate the first quartile (Q1), the third quartile (Q3), and the interquartile range (IQR = Q3 - Q1); determine data points that satisfy "x > Q3 + 1.5 × IQR" or "x < Q1 - 1.5 × IQR" as outliers to avoid interference from extreme values.

[0034] Outlier replacement: Instead of using the deletion method (to prevent loss of data information), calculate the mean μ and standard deviation σ of the feature column after excluding outliers, and replace according to the 3σ principle: when the outlier is greater than the upper limit, replace it with μ + 3σ; when it is less than the lower limit, replace it with μ - 3σ to ensure the stability of the data distribution.

[0035] Example: For a certain feature column, Q1 = 15, Q3 = 35, IQR = 20, the outlier thresholds are 35 + 1.5 × 20 = 65 and 15 - 1.5 × 20 = -15; for an outlier of 70, μ = 25 and σ = 10 for this column, and the 3σ upper limit = 25 + 3 × 10 = 55, so replace 70 with 55.

[0036] Data balancing processing: For the imbalance problem where the number of financial distress samples (minority class) is much less than that of normal samples (majority class), implement a strategy combining "oversampling + undersampling": SMOTE oversampling: Set the number of neighbors K = 3, randomly select minority class samples and randomly select a sample among its 3 neighbors Generate a random number δ in the range [0, 1] (such as 0.4), and according to the formula ; Synthesize new samples to ensure that the new samples are on the connection line of the two-sample feature space and fit the original minority class distribution.

[0037] ENN undersampling: For the synthesized full set of samples, calculate the 5-neighbor classes of each sample; if more than half of the neighbors are inconsistent with the sample class, then determine it as a noise sample and remove it to clean the data in the class overlap area and optimize the decision boundary.

[0038] Balancing goal: After applying the SMOTEENN algorithm, the class imbalance problem can be effectively alleviated, the proportion of minority class samples can be significantly increased, and the impact of class imbalance on model training can be eliminated.

[0039] Data normalization: To eliminate the difference in feature dimensions (for example, the unit of total assets is 100 million yuan and the asset-liability ratio is a percentage), linearly map all features to the range [0, 1], and the formula is where and Calculate based on the training set, and use the same parameters for normalization of the validation set and the test set to ensure the unity of the data scale and accelerate model convergence.

[0040] Example: A training set of certain features =10、 =100, the original value of a sample =40, after normalization = =0.333.

[0041] Feature filtering module: The preprocessed data may still contain redundant or undiscriminative features. This module uses a three-level process of "normality test - inter-group difference test - indicator purification" to screen out core indicators that have significant distinguishing power for financial distress.

[0042] Screening process and implementation details: Step 1: Normality test (Shapiro-Wilk test): For each candidate feature, perform the test separately in the financial distress group and the normal group. The null hypothesis is "the data follows a normal distribution", and the significance level is set to α=0.05. If the p-value is >0.05, accept the null hypothesis and determine that the data in this group follows a normal distribution; if the p-value is ≤0.05, reject the null hypothesis and determine that the data is not normally distributed.

[0043] Step 2: Between-group difference test: Normal distribution characteristics: An independent samples t-test was used, with the null hypothesis being "no significant difference between the two group means"; the t-statistic was calculated (formula: ,in , The two sets of means are given. , There are two sets of variances. , (For two groups of samples), if the p-value is <0.05, the null hypothesis is rejected, and the difference in means between the feature groups is considered significant.

[0044] Non-normal distribution characteristics: The Mann-Whitney U test (a non-parametric test) was used, with the null hypothesis being "no significant difference between the two distributions"; the U statistic was obtained by calculating the sample rank sum. If the p-value was <0.05, the null hypothesis was rejected, and the distribution difference between the groups was determined to be significant.

[0045] Step 3: Indicator purification: Summarize all significant features with p-values ​​< 0.05. To further improve the robustness of the model, the variance inflation factor (VIF) of the filtered features can be calculated. Features with excessively high VIF values ​​(such as > 10) can be removed as appropriate to reduce the impact of multicollinearity. Finally, a prediction system with 20-30 core indicators is formed, which balances effectiveness and simplicity.

[0046] Model building and optimization module: This module is the core prediction unit of the system. It constructs a high-precision financial distress prediction model through a three-stage process of "basic model building - intelligent hyperparameter optimization - model performance verification".

[0047] Base model: XGBoost classifier

[0048] Model initialization: A binary classification model is built based on the gradient boosting decision tree (GBDT) framework. The initial parameters are set as follows: number of trees (n_estimators) = 100, learning rate (learning_rate) = 0.1, maximum tree depth (max_depth) = 6, minimum child weight (min_child_weight) = 1, subsample ratio (subsample) = 0.8, and column sampling ratio (colsample_bytree) = 0.8.

[0049] Loss function: Log loss function (LogLoss) is used, the formula is as follows: ,in The labels are for real situations (1 = financial distress, 0 = normal). This function represents the original predicted score of the model. It can effectively quantify the prediction bias, has good differentiability, and supports gradient descent optimization.

[0050] Hyperparameter optimization: Red-billed Blue Magpie optimization algorithm

[0051] Optimization objective: To maximize the AUC (area under the curve) of the model on the validation set, search for the optimal combination of hyperparameters. The hyperparameter search space is as follows: learning rate (0.01-0.3), maximum tree depth (3-10), minimum sample weight sum (0.1-1.0), subsample ratio (0.5-1.0), column sampling ratio (0.5-1.0), regularization parameter λ (0-10), and regularization parameter γ (0-5).

[0052] Algorithm implementation: Initialization: Generate 30 search agents (each agent corresponds to a set of hyperparameters), and set the number of iterations to 100.

[0053] Location update strategy: Small group collaboration strategy: Divide the 30 agents into 6 small groups (p=5), according to the formula Update location, where For the first The first agent Substitute position, Rand2 is a random number in the range [0,1], which enables a fine-grained local search.

[0054] Group collaboration strategy: Based on all 30 agents (q=30), according to the formula Update the position to achieve a global broad search and avoid getting trapped in local optima.

[0055] Iterative optimization: After each iteration, the agent fitness (AUC value) is calculated. The top 10% of the best agents are retained, and the bottom 10% of the worst agents are replaced to ensure the efficiency of population evolution. After the iteration ends, the hyperparameter combination corresponding to the agent with the highest fitness is selected as the optimal solution.

[0056] Model Validation: Hierarchical 5-fold cross-validation is used (keeping the class ratio of the training set and the validation set consistent). Validation metrics include AUC, accuracy (ACC), precision (P), recall (R), and F1 score. The model performance is evaluated through cross-validation. If the key metrics (such as AUC and F1 score) do not meet the expectations, the search space or number of iterations of the optimization algorithm can be adjusted, and hyperparameters can be re-tuned until the model performance meets the requirements.

[0057] Interpretability Analysis Module: To address the trust issues inherent in "black box" models, this module, based on the SHAP value theory, makes the model's prediction results transparent and interpretable.

[0058] SHAP value calculation and core principles: Calculation method: The TreeSHAP algorithm (adapted to XGBoost model) is used. The optimized model and prediction data are loaded, each decision tree node is traversed, the contribution value of the feature on the node is calculated, and the SHAP value of each feature of each sample is obtained by summarizing.

[0059] Core formula: Decompose the sample prediction value into the sum of feature contributions, the formula is as follows. ,in This represents the model's predicted value (log odds). The average expected value (baseline value) of all sample predictions. Features The SHAP value; a positive value indicates that the feature shifts the prediction toward "financial distress", a negative value shifts it toward "normal", and the absolute value indicates the strength of the contribution.

[0060] Calculate the mean of the absolute values ​​of the SHAP values ​​of all features for all samples, sort them by mean to generate a "core risk factor list", and intuitively display the indicators that have the greatest impact on the prediction of financial distress (such as debt-to-equity ratio, cash flow ratio, etc.).

[0061] For a single high-risk enterprise, output the top 5 features that affect its forecast results and their corresponding SHAP values, such as "debt-to-equity ratio (78%) SHAP value = 0.35 (positive contribution), net profit growth rate (-12%) SHAP value = 0.28 (positive contribution)", to clarify the source of risk.

[0062] Plot a feature dependency graph, with the feature value on the horizontal axis and the SHAP value on the vertical axis. By fitting a curve, the non-linear relationship between the feature and risk can be shown. For example, "when the debt-to-equity ratio is below 50%, the SHAP value is close to 0; after it exceeds 70%, the SHAP value rises rapidly, and the risk increases sharply."

[0063] Generate a SHAP summary diagram (bee diagram) that displays feature importance, SHAP value distribution, and feature values ​​(red = high value, blue = low value), making it easy to intuitively understand the influence of features on prediction.

[0064] Output module: This module serves as the interaction terminal between the system and the user, integrating the prediction results and interpretability analysis into a structured report to meet the diverse needs of decision-making scenarios. The output content strictly corresponds to the "prediction results + risk factor analysis report" limitation of claim 1.

[0065] Sort by financial distress prediction probability from high to low, output the company name, prediction probability, and risk level of all analyzed objects (e.g., probability > 0.7 = high risk, 0.3-0.7 = medium risk, < 0.3 = low risk), support export in CSV and Excel formats, and facilitate batch filtering of key objects of focus.

[0066] For high-risk / key enterprises, a structured report is generated, with core content including: basic enterprise information (code, industry, size), core forecast data (probability, confidence level), top 5 risk / advantage factors (including SHAP value, indicator value, industry comparison), feature dependency analysis chart, and historical data trend comparison, comprehensively supporting risk decision-making.

[0067] It supports three output formats: PDF (standardized layout, including cover, table of contents, and charts, suitable for formal reports), HTML (interactive visualization, with clickable sample details, suitable for internal analysis), and Dashboard (real-time updates of risk data, suitable for risk control and monitoring scenarios).

[0068] Based on the above system, the data processing method of the present invention forms a complete closed-loop process (as shown in the appendix). Figure 2 As shown in the figure, the implementation process is as follows: S1: Data Acquisition: Start the data acquisition module and collect the target listed company's internal financial / non-financial data, external social indicators, and macroeconomic variables according to the "timed + event triggered" strategy. The data is then stored after triple verification.

[0069] S2: Data Preprocessing: Call the preprocessing module and execute the following steps in sequence: K-nearest neighbor algorithm to fill missing values ​​→ box plot method to identify outliers + 3σ principle to replace them → SMOTEENN mixed sampling to balance categories → min-max normalization, and output a standardized data matrix.

[0070] S3: Feature Selection: Run the feature selection module to perform the Shapiro-Wilk normality test on the standardized data. Based on the test results, use the T-test / Mann-Whitney U test to select significant features, eliminate multicollinearity indicators, and form a predictive indicator system.

[0071] S4: Model Building and Optimization: Initialize the XGBoost model with the selected feature data, start the Red-billed Blue Magpie optimization algorithm to search for the optimal combination in the preset hyperparameter space, and verify the model performance through hierarchical 5-fold cross-validation until the target is met.

[0072] S5: Financial Distress Prediction: Process the latest data of the company to be predicted according to the S2-S3 process, input it into the optimized XGBoost model, and output the financial distress prediction probability and risk level.

[0073] S6: Interpretability Analysis: Call the interpretability analysis module to calculate the SHAP values ​​of each feature in the prediction results, and generate a global feature importance list, individual attribution results, and visualization charts.

[0074] S7: Report Output: Integrate the prediction results of S5 with the analysis conclusions of S6, and generate a comprehensive report in multiple formats through the output module for users to use for decision support.

[0075] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A financial data processing system, characterized by: include: The data acquisition module is used to collect multi-source, multi-modal data from listed companies. The multi-source, multi-modal data includes internal financial data, internal non-financial data, external social indicators, and macroeconomic variables. The preprocessing module is used to perform missing value imputation, outlier handling, data balancing, and normalization on the collected data. The feature filtering module is used to filter out predictive indicators that have significant distinguishing ability against financial distress from the processed data based on statistical significance tests. The model building and optimization module is used to build a prediction model based on XGBoost as the classifier, and to use the Red-billed Blue Magpie optimization algorithm to perform global search optimization on the hyperparameters of the XGBoost model. The interpretability analysis module is used to perform feature importance analysis and attribution interpretation of prediction results on the optimized prediction model based on SHAP value theory. The output module is used to output the financial distress prediction results of listed companies and the corresponding risk factor analysis report.

2. A financial data processing system as claimed in claim 1, characterized in that: When performing missing value imputation, the preprocessing module uses the K-nearest neighbor algorithm, with the Minkowski distance as the distance metric, calculated as follows:

3. The financial data processing system as described in claim 1, characterized in that: When the preprocessing module handles outliers, it uses a box plot method for judgment. The judgment criterion is: if the data point satisfy or If it is, it is determined to be an outlier, and the data in that feature column is used. The value is replaced; where, and These are the lower and upper quartiles of the data, respectively. Interquartile range, The standard deviation is denoted as .

4. The financial data processing system as described in claim 1, characterized in that: When performing data balancing, the preprocessing module uses the SMOTEENN hybrid sampling algorithm, where the formula for synthesizing new minority class samples in the SMOTE part is: in, For the selected minority class samples, It is a neighbor sample randomly selected from its K nearest neighbors. It is a random number in the interval [0,1].

5. The financial data processing system as described in claim 1, characterized in that: When performing normalization in the preprocessing module, the minimum-maximum normalization method is used to linearly scale the value of each feature index to the [0,1] interval. The calculation formula is as follows: in, The normalized value. The original value, and These are the minimum and maximum values ​​of the feature in the dataset, respectively.

6. The financial data processing system as described in claim 1, characterized in that: The feature screening module's process for screening predictive indicators based on significance testing includes: first, performing a Shapiro-Wilk normality test on each candidate indicator; for indicators that pass the normality test, using a paired T-test to determine whether the difference in their means between the financial distress group and the normal group is significant; for indicators that fail the normality test, using a Mann-Whitney U-test to determine whether their distribution differences are significant; and finally, selecting indicators that have significant differences at a preset significance level to form a predictive indicator system.

7. The financial data processing system as described in claim 1, characterized in that: In the model building and optimization module, the loss function of the XGBoost classification model used for binary classification problems is: in, For real labels, These are the raw scores predicted by the model.

8. The financial data processing system as described in claim 1, characterized in that: The red-billed blue magpie optimization algorithm used in the model construction and optimization module employs two position update strategies during the search phase: a small group cooperation strategy and a group cooperation strategy. Their mathematical expressions are as follows: Small group collaboration strategies: Group collaboration strategies: in, For small group size, For group size, For randomly selected search agents, and It is a random number.

9. A financial data processing system as described in claim 1, characterized in that: The interpretability analysis module, based on the SHAP value theory, decomposes the model prediction value of each sample into the sum of the contributions of each feature. Its basic expression is: in, For the model to sample The predicted value, This represents the baseline expected value of the model across all samples. Features The SHAP value indicates the impact of this feature on the prediction result. The contribution of the value indicates that a positive value pushes the prediction toward the positive class, while a negative value does the opposite.

10. A financial data processing method, applied to a financial data processing system according to any one of claims 1 to 9, characterized in that: The method includes the following steps: S1. Collect multi-source, multi-modal data from listed companies; S2. Preprocess the data, including using the KNN algorithm to fill in missing values, replacing outliers based on box plots, using the SMOTEENN algorithm to balance the classes, and performing min-max normalization. S3. Screening prediction indicators based on the significance test process of Shapiro-Wilk, T-test and Mann-Whitney U-test; S4. Construct an XGBoost prediction model and optimize its hyperparameters using the Red-billed Blue Magpie Optimization Algorithm. S5. Predict the financial distress of listed companies based on the optimized model; S6. Use SHAP value theory to interpret the prediction results and analyze the direction and intensity of the influence of each feature on the prediction results; S7. Output a report that includes a probability forecast of financial distress and an analysis of key risk factors.

Images