Enterprise credit overdue early warning analysis method, system, device and storage medium

Abstract

The present application relates to a kind of enterprise credit overdue early warning analysis method, system, equipment and storage medium, belong to credit analysis technical field, wherein, the method includes: collecting the enterprise credit data of at least two industries, and constructs sample data set;The sample data set is divided into internal verification set, time external verification set and regional external verification set;Early warning analysis model is trained using the internal verification set, and cross-validation is used to generate out-of-fold prediction probability, and the final classification threshold is determined based on the out-of-fold prediction probability, to obtain the trained early warning analysis model;The credit data of target enterprise is input into the trained early warning analysis model, to obtain the credit overdue probability of the target enterprise.This scheme solves the problems of insufficient external generalization ability, limited adaptability to class imbalance and table features, and weak interpretability in cross-scene migration of existing methods.

Description

Technical Field

[0001] This invention relates to the field of credit analysis technology, and in particular to a method, system, device and storage medium for early warning analysis of corporate credit delinquency. Background Technology

[0002] Faced with the continuous accumulation of credit risks for SMEs and the data distribution drift caused by changes in industry structure and regional distribution, how to achieve robust identification and accurate early warning of overdue risks has become a core issue of common concern to academia and industry. Existing methods generally suffer from insufficient external generalization ability, limited adaptability to class imbalance and tabular features, and weak interpretability when transferring across scenarios.

[0003] Therefore, there is an urgent need for a method, system, equipment, and storage medium for early warning analysis of corporate credit delinquency. Summary of the Invention

[0004] In view of this, it is necessary to provide a method, system, device and storage medium for early warning analysis of corporate credit delinquency, in order to solve the problems that existing methods generally have inadequate external generalization ability, limited adaptability to class imbalance and table features, and weak interpretability when migrating across scenarios.

[0005] To address the aforementioned problems, this invention provides a method for early warning analysis of corporate credit delinquency. This method utilizes the SAFE robust early warning framework, which includes: Input layer: Used to collect corporate credit data from at least two industries and construct a sample dataset; Drift perception layer: used to divide the sample dataset into an internal validation set, a time-external validation set, and a region-external validation set; Leakage prevention layer: used to train the early warning analysis model using the internal validation set, generate the out-of-bounds prediction probability using cross-validation, and determine the final classification threshold based on the out-of-bounds prediction probability to obtain the trained early warning analysis model; The trained early warning analysis model is then validated using the time-based external validation set and the region-based external validation set to examine its generalization ability in external scenarios. Output layer: Used to input the credit data of the target enterprise into the trained early warning analysis model to obtain the credit delinquency probability of the target enterprise.

[0006] The beneficial effects of this invention are: For sample datasets from different industries, this invention constructs internal validation sets, time-based external validation sets, and region-based external validation sets to evaluate the generalization performance of the early warning analysis model across time and regions. Experimental results show that the TabICLv2 model achieves superior overall performance on sample datasets from each industry, outperforming existing linear models, ensemble learning models, and deep table models.

[0007] Interpretable analysis shows that variables related to judicial enforcement are common core high-risk drivers across different industries, while factors such as business registration changes, financing collateral, tax credit, and innovation qualifications exhibit differentiated effects across industries. The results indicate that the SAFE robust early warning framework can improve the reliability, extrapolation ability, and business usability of early warning systems for SME loan delinquency risks under data drift conditions. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the SAFE robust early warning framework used in an enterprise credit delinquency early warning analysis method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the risk stratification results for an industrial enterprise dataset, as provided in an embodiment of the present invention. Figure 3 A cumulative capture rate curve for an industrial enterprise dataset provided in this embodiment of the invention; Figure 4 This invention provides a KS curve diagram for an industrial enterprise dataset. Figure 5 This is a simulation diagram of a strategy for calculating the change in precision as a function of disposal ratio for an industrial enterprise dataset, provided as an embodiment of the present invention. Figure 6 This invention provides a simulation diagram of a TPR (Total Return) strategy for industrial enterprise datasets as a function of disposal ratio. Figure 7 This invention provides a simulation diagram of the FPR (Fault Ratio) variation strategy for an industrial enterprise dataset, as shown in this embodiment. Figure 8 This invention provides a SHAP-like beeswarm global interpretation graph for industrial enterprise datasets. Figure 9 This invention provides a schematic diagram of the Top 10 positive features for an industrial enterprise dataset. Figure 10 This invention provides a schematic diagram of the Top 10 negative features of an industrial enterprise dataset. Figure 11This invention provides a path diagram for increasing default risk in a dataset of industrial enterprises. Figure 12 This invention provides a path diagram for reducing default risk in industrial enterprise datasets. Figure 13 This is an overview diagram of all feature directions provided in an embodiment of the present invention; Figure 14 This is an example diagram illustrating a low-risk sample local interpretation for an industrial enterprise dataset, provided as an embodiment of the present invention. Figure 15 This is an example diagram illustrating a partial interpretation of high-risk samples in an industrial enterprise dataset, provided as an embodiment of the present invention. Detailed Implementation

[0009] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0010] Small and medium-sized enterprises (SMEs) are among the most dynamic business entities in my country's real economy, playing a vital role in absorbing employment, stabilizing industrial chains, and promoting regional economic development. However, influenced by multiple factors such as macroeconomic fluctuations, changes in industry conditions, tightening financing environments, and differences in corporate governance, SME credit risk exhibits significant dynamic, heterogeneous, and phased characteristics. Especially against the backdrop of industry restructuring, customer base shifts, regional business migration, and changes in the external business environment, the sample distribution used in model development often struggles to remain stable over the long term. This data drift problem has gradually become a significant factor restricting the external generalization ability of credit early warning models.

[0011] The data drift referred to in this embodiment of the invention refers to a change in the data generation mechanism between the machine learning model development environment and the actual application environment, leading to a shift in the mapping relationship between sample distribution, risk exposure structure, or features and default outcomes. Specifically, on the one hand, the statistical distribution of a company's characteristics such as years of establishment, regional distribution, judicial enforcement records, business registration changes, and operational stability may change with time and context; on the other hand, the risk meaning of the same feature may also change in different periods, regions, and customer groups, making it difficult for the discrimination boundary learned by the machine learning model in the development samples to be directly transferred to the external environment. The former is closer to feature distribution drift, while the latter reflects a more general concept drift. When both work together, they can easily lead to problems such as decreased recognition ability, threshold failure, and increased false alarms in real business deployments of machine learning models.

[0012] Furthermore, this paper focuses not only on whether the early warning analysis model is "accurate in its predictions," but also on whether the model is "clearly interpretable and usable." Early studies on the interpretability of financial (credit default) prediction models mainly relied on statistical tools such as regression coefficients, variable signs, and group default rates to explain the model results. With the introduction of machine learning methods such as random forests, gradient boosting trees, and pre-trained tabular models, the predictive performance of early warning analysis models has been significantly improved, but their internal decision-making logic has also become more complex, making traditional parametric interpretation methods difficult to apply.

[0013] At the interpretability analysis level, this paper combines SHAP-like global attribution, risk path mapping, and feature directionality analysis to reveal in layers "what types of features drive up or down default risk." At the business interpretation level, the interpretation results are combined with risk stratification, Top-k capture capabilities, and strategy simulation, thereby transforming the machine learning model output into an actionable basis for pre-loan screening, credit review, and post-loan early warning. SHAP is a method for interpreting the prediction results of machine learning models.

[0014] While existing research provides a solid foundation for corporate credit risk identification, several aspects warrant further improvement. First, many studies still rely on single datasets or random partitions for validation, neglecting data drift scenarios such as time migration, regional migration, and industry migration. However, related research has pointed out that distribution shifts significantly affect the model's reliability in target scenarios; therefore, relying solely on internal results from identically distributed samples can easily overestimate the model's true extrapolation ability. Second, in credit risk control tasks, some studies have not strictly constrained the risk of information leakage during model development. For example, information backflow from external test sets in feature selection, threshold selection, encoding fitting, or parameter adjustment may lead to overly optimistic evaluation results and weaken the credibility of the conclusions. Third, existing research still has limitations in model evaluation and interpretation. Many studies primarily focus on comparing single metrics such as AUC and accuracy, or only ranking the importance of methods like SHAP, without systematically integrating the interpretation results with business scenarios such as risk stratification, high-risk priority identification, handling ratios, false alarm costs, and strategy simulation. This limits the transformation of model conclusions into actual risk control decisions.

[0015] To address the aforementioned shortcomings, this paper focuses on data drift scenarios and constructs a credit delinquency early warning research framework that balances robustness, rigorous evaluation, and business interpretation, based on two large industry datasets: industrial enterprises and wholesale and retail enterprises. Specifically, this paper proposes the SAFE robust early warning framework, using the TabICLv2 pre-trained table model as the core early warning analysis model. Here, S stands for Shift-aware (drift perception layer); A stands for Anti-leakage (leakage prevention layer); F stands for Fair-evaluation (fair evaluation layer); and E stands for Explainable (explanation layer).

[0016] In terms of sample design, one internal validation set and two external validation sets are set for each industry dataset. The discrimination threshold is strictly limited to the prediction determined by the internal cross-validation OOF (the prediction value generated for each sample in K-fold cross-validation by the "model that has not seen the sample", often used to evaluate generalization performance or build ensemble models). The external test set does not participate in the selection of threshold. In terms of result presentation, in addition to indicators such as AUC, KS, and ACC, the business usability of the model is further systematically evaluated by combining risk stratification, Gains curve, KS curve, and changes in Precision, TPR (true positive rate), and FPR (false positive rate) under different treatment ratios.

[0017] Compared to existing approaches that rely solely on a single random partitioning, the above design is closer to the real-world application scenarios of models deployed by financial institutions.

[0018] It should be noted that AUC (Area Under Curve) is defined as the area under the ROC curve and the coordinate axes. KS (Kolmogorov-Smirnov) is the discrimination metric for binary classification models. ACC is the accuracy metric, the most basic evaluation metric for classification tasks, referring to the proportion of correctly predicted samples out of the total number of samples.

[0019] Figure 1 This is a schematic diagram of the SAFE robust early warning framework used in an enterprise credit delinquency early warning analysis method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, to address the common problems in early warning tasks for overdue loans to SMEs, such as data drift, information leakage, overly optimistic assessments, and difficulty in applying interpretation results, this paper proposes the SAFE robust early warning framework. This framework includes an input layer, a drift perception layer, a leakage prevention layer, a fair assessment layer, an interpretation layer, and an output layer. Specifically, the drift perception layer is used for sample partitioning and external extrapolation design oriented towards data drift; the leakage prevention layer is used for leakage prevention coding and threshold determination mechanisms; the fair assessment layer is used for multiple external verifications, risk stratification, and strategy simulation; and the interpretation layer is used for multi-dimensional interpretation and analysis of drift perception.

[0020] The SAFE robust early warning framework uses TabICLv2 as its core prediction model and constructs a complete methodology from four levels: sample design, leakage prevention control, robustness assessment, and interpretability analysis. Its goal is to improve the model's generalization ability, assessment reliability, and business usability across time, region, and industry scenarios. Unlike traditional research that often uses a single dataset, random partitioning, and static interpretation, the SAFE robust early warning framework emphasizes explicitly incorporating data drift into the experimental design and uses a combination of internal and multiple external validations to verify the model's stable performance under real-world deployment conditions.

[0021] Credit delinquency early warning for SMEs is a typical tabular binary classification task, but it is characterized by the coexistence of multi-source heterogeneous features, strong nonlinearity in variable relationships, a large number of high-cardinality category features, a low proportion of default samples, and significant differences in risk mechanisms across different industries. To more effectively characterize the complex risk structure, this paper selects the TabICLv2 model as the core prediction model in the SAFE robust early warning framework.

[0022] TabICLv2 belongs to the basic tabular model paradigm, and its core idea is: given a training set and the sample to be predicted The TabICLv2 model no longer explicitly performs the traditional two-stage "retraining-repredicting" process, but instead directly outputs the conditional prediction distribution through a single Transformer-based forward propagation. Unlike the traditional supervised learning process of "train first, then predict," the TabICLv2 model essentially performs an in-context learning mechanism within a table context. That is, given the context of the training samples, the TabICLv2 model directly outputs prediction results for the test samples. The pre-training objective of the TabICLv2 model can be written as: (1) In the formula, This represents the overall loss function when the TabICLv2 model has parameters θ; θ represents the learnable parameters of the TabICLv2 model. Indicates the distribution of training tasks The above describes the expectation for different table tasks; D represents a table learning task sampled from the task distribution; This represents the sample dataset used for contextual training in this task; This represents the feature vector of the sample to be predicted. This represents the true label of the sample to be predicted; Indicates that in a given training context and test sample features Under these conditions, the TabICLv2 model supports labels. The given conditional probability distribution; This represents the negative log-likelihood loss term, used to measure the deviation between the model's predicted distribution and the true label.

[0023] Equation (1) means that the TabICLv2 model can learn to directly predict new samples given the context of training samples by minimizing the negative log-likelihood of the test labels on a large number of tabular tasks.

[0024] From an architectural perspective, the TabICLv2 model can be summarized into three stages: column-level embedding, row-level interaction, and dataset-level ICL. For the input table... First, the column representation is constructed through the column direction encoder (Equation-2). Then, each row is compressed into a fixed-dimensional row representation through the row interaction module (Equation-3). Finally, the row representations of the training samples and the row representations of the test samples are fed into the dataset-level Transformer, where the test samples perform attention aggregation on the training samples, and the predicted probability is output (Equation-4).

[0025] (2) (3) (4) In equation (2), This represents the input table matrix, where n represents the number of samples and m represents the number of features; This indicates a column-direction encoder or a column-level Transformer module; This represents the column representation result obtained after column-level modeling. Equation (2) shows that the TabICLv2 model first encodes the original table in the column dimension to learn the semantic representation and statistical structure between different fields.

[0026] In equation (3), The column-level representation obtained by expression (2); This indicates a row-direction interaction module or a row-level Transformer encoder. Equation (3) represents the fixed-dimensional row representation obtained for each sample. Equation (3) illustrates that the model further models the interaction relationships between features within the same sample based on column-level encoding, thereby obtaining a sample-level representation that can be used for classification decisions.

[0027] In equation (4), This represents the predicted probability of a company defaulting on its credit in this paper, as shown by the TabICLv2 model output. The row representation of the training samples; This represents the true label corresponding to the training sample; The row representation of the sample to be predicted; This represents a dataset-level Transformer, whose function is to enable the test samples to perform attention aggregation on the training samples within the context; Let θ represent the final prediction mapping function determined by the TabICLv2 model parameter θ. Equation (4) shows that the TabICLv2 model does not first train a static classifier independently and then make predictions, but instead directly combines the training sample context and the test sample representation to complete conditional predictions during the inference stage.

[0028] For the credit delinquency task in this paper, the early warning analysis model ultimately outputs the delinquency probability of enterprises. All subsequent threshold divisions, risk stratification, Top-k high-risk identification, and interpretability analysis are based on this probability output. It should be noted that the experiments in this paper call the model through the TabICLClassifier interface in the official implementation of the TabICL model. Since the checkpoint version was not manually specified, the experimental environment loads tabicl-classifier-v2-20260212.ckpt by default, so the core prediction model in this paper corresponds to the TabICLv2 model. The SAFE robust early warning framework does not replace the early warning analysis model, but rather builds a robust training, validation, and interpretation process around the early warning analysis model: the TabICLv2 model is responsible for "how to predict", while the SAFE robust early warning framework is responsible for "how to make the prediction reliable, robust, and usable in drift scenarios".

[0029] SAFE Robust Early Warning Framework's Drift Awareness Layer: Sample Partitioning and Extrapolation Design for Data Drift.

[0030] In real-world lending, the training and deployment environments are often not the same. Credit risk for SMEs can drift due to industry restructuring, stratification by age, changes in credit granting strategies, regional customer group differences, and fluctuations in the external operating environment. If only a randomized partitioning assessment model is used, the distribution of training and testing samples will be too similar, easily overestimating the model's generalization ability in external scenarios.

[0031] Therefore, this paper further constructs an out-of-time proxy (OOT-proxy) drift proxy scenario within the internal validation set to assist in parameter selection and model configuration comparison. Unlike traditional methods that select parameters solely based on internal cross-validation results, this paper does not directly utilize real external sets for parameter tuning. Instead, it simulates a lightweight drift environment closer to queue extrapolation within the internal validation set based on the queue information of enterprise samples, using this as an auxiliary evaluation basis during model development. This design avoids evaluation bias caused by external test information feedback and also allows the parameter selection process to maintain a certain sensitivity to potential time drift, thereby improving the model's adaptability to real external scenarios.

[0032] Instead of using a simple random dataset partitioning, this paper constructs a three-layer validation structure of "internal validation set + two external validation sets" for each industry dataset. The internal validation set is used for model training, cross-validation, and threshold selection; the first external validation set (time external validation set) is mainly used to simulate distribution drift caused by time changes; and the second external validation set (regional external validation set) is used to simulate distribution drift caused by regional migration or differences in external customer groups.

[0033] Let the set of all firms be Then the sets of enterprises corresponding to the internal samples, time-external samples, and region-external samples are respectively , and All three conditions are met. , , The core significance of this design lies in the fact that it no longer only answers the question "Is the TabICLv2 model effective in identically distributed samples?", but further answers "Is the TabICLv2 model still effective in the context of time drift and regional drift?" Furthermore, to improve the adaptability of the parameter selection stage to potential drift scenarios, this paper constructs an OOT-proxy (out-of-time proxy) as a lightweight extrapolation agent in the internal samples.

[0034] Let the candidate parameter space be Then the parameter selection target can be written as: (5) In equation (5), This represents a candidate model configuration, including model parameters, preprocessing methods, and feature selection configurations; ^ represents the parameter space consisting of all candidate configurations. This represents the optimal configuration that was ultimately selected. Indicates configuration Performance in internal validation or internal cross-validation; Indicates configuration Performance in OOT-proxy drift proxy scenarios; represents the trade-off coefficient between internal performance and surrogate extrapolation performance. Equation (5) shows that in the parameter selection process, this paper not only considers the model's fitting effect on internal samples, but also takes into account its robustness in lightweight extrapolation scenarios.

[0035] in, This is a tradeoff coefficient between internal performance and surrogate extrapolation performance. This represents the combination of TabICLv2 model parameters and preprocessing configuration. Thus, the drift-aware layer achieves a shift from static validation to drift-aware validation, providing a more realistic deployment sample basis for subsequent robust evaluation and cross-industry interpretation.

[0036] The leakage prevention layer of the SAFE robust early warning framework: leakage prevention coding and threshold determination mechanism.

[0037] In corporate credit delinquency early warning tasks, information leakage is one of the important reasons for overly optimistic model evaluation results. Especially in business data with many derived variables, high cardinality of categorical features, and complex preprocessing procedures, if there are no strict constraints on sample partitioning, feature processing, parameter selection, and threshold optimization, test information can easily flow back to the training stage, thus overestimating the model's recognition ability in real deployment scenarios. To address this, this paper introduces an anti-leakage mechanism into the SAFE robust early warning framework to control the risk of information leakage from three levels: sample partitioning, feature processing, and threshold determination.

[0038] First, in the sample partitioning stage, this paper adopts the enterprise-level mutual exclusion principle to ensure that there is no overlap between internal samples, time-external samples, and region-external samples, thus avoiding information leakage caused by the same enterprise appearing across sets from the source. Second, in the feature processing, all numerical mapping, missing value imputation, and feature selection steps are only fitted within the internal training range before being applied to the validation set and external test set, and the mapping parameters are not allowed to be re-estimated using external data, thereby preventing test information from flowing back through the preprocessing process. Finally, in determining the classification threshold, this paper does not directly adjust the threshold based on the performance of the external test set, but strictly determines it based on the OOF (Out-Of-Fold) prediction results of internal cross-validation.

[0039] Let the true labels of the internal samples and their corresponding OOF predicted probabilities be respectively... Therefore, this paper uses the Youden exponent-based method to determine the optimal threshold, as expressed in the following expression: (6) In equation (6), Indicates the candidate classification threshold; This indicates the optimal threshold that was ultimately selected; Indicates at the threshold The TabICLv2 model's recognition rate of real overdue samples, i.e., the true recognition rate; Indicates at the threshold The false positive rate is the proportion of normal samples that the TabICLv2 model misclassifies as overdue samples. Equation (6) indicates that this paper uses the Youden exponent maximization principle to determine the final classification threshold, and this threshold is obtained only based on the internal OOF results without using any external test set information.

[0040] in, This represents the final classification threshold. Equation (6) shows that the selection of the threshold depends entirely on the cross-validation results of the internal samples, without using any external test set information. The optimal threshold is obtained... Subsequently, the predicted labels for external samples are uniformly given by a fixed threshold, i.e.: (7) In equation (7), Represents external samples The predicted probability; This represents the final threshold determined by the internal OOF result; Represents external samples The final predicted category; 1(·) represents the indicator function, which takes the value of 1 when the condition in parentheses is true, and takes the value of 0 otherwise. Equation (7) shows that, whether it is the time external validation set or the region external validation set, it is only used at a fixed threshold. The verification is performed once and then no longer involved in threshold correction or parameter tuning.

[0041] in, Represents external samples The predicted probability, This indicates the final classification result. Equation (7) shows that both the time-based external validation set and the region-based external validation set are validated only once at a fixed threshold and are no longer involved in threshold correction or parameter tuning.

[0042] Through the above design, this paper achieves a rigorous process of "internal threshold setting and external one-time verification." The core value of this mechanism lies in its strict separation of the model development phase and the model verification phase, making external verification more closely resemble the deployment scenario in real business operations. For research on credit early warning for SMEs, this leakage prevention control is not merely a technical detail, but a crucial prerequisite for ensuring that the model results have extrapolation significance and decision-making reference value.

[0043] SAFE Robust Early Warning Framework's Fair Assessment Layer: Multiple External Validation, Risk Stratification, and Strategy Simulation.

[0044] For the task of early warning of overdue loans to small and micro enterprises, relying solely on a single classification indicator is insufficient to fully reflect the practical application value of the model. This paper constructs a comprehensive evaluation system within the SAFE robust early warning framework, consisting of multiple external validations, risk stratification, and strategy simulation, to systematically evaluate the model from three aspects: model discrimination ability, risk ranking ability, and business application value.

[0045] In terms of basic performance evaluation, this paper uses indicators such as AUC, KS, ACC, TPR, and FPR to examine the model's recognition effect from multiple dimensions. Among them, the AUC indicator reflects the model's overall ability to distinguish between positive and negative samples. The AUC indicator is not affected by sample imbalance and is an important indicator for evaluating credit default prediction models. The KS indicator measures the maximum degree of separation between high-risk and low-risk samples. In terms of risk ranking evaluation, this paper further introduces risk stratification analysis. Specifically, according to the predicted probabilities output by the model, the samples are divided into several risk levels or decimal intervals, and the distribution of actual delinquency rates in each level is statistically analyzed.

[0046] (8) In equation (8), This represents the b-th risk group or tenths group obtained after sorting according to the predicted probabilities. Indicates the number of samples in this group; Indicates sample The actual label is 1, where 1 indicates overdue and 0 indicates not overdue; Let g represent the actual delinquency rate of group g. Equation (8) is used to measure the observed default rate level in different risk levels to test whether the risk score output by the model has good stratification ability and monotonicity.

[0047] To evaluate the TabICLv2 model's ability to prioritize high-risk clients, this paper employs Gains curve analysis and Top-k analysis. Given a disposal ratio... Take the one with the highest predicted probability. For each sample, the cumulative bad customer accuracy rate is defined as: (9) Equation (9) is used to measure the effect of treatment only before the procedure. When dealing with a high proportion of high-risk customers, what percentage of the total actual overdue samples can the TabICLv2 model cover? If a high capture rate can be achieved with a relatively small disposal ratio, it indicates that the TabICLv2 model has stronger front-end screening efficiency.

[0048] In equation (9), This represents the set of the top k high-risk samples selected after sorting them from highest to lowest predicted probability; k can also correspond to the number of samples under a given treatment ratio r; N represents the total number of samples; This represents the true label of sample i; This represents the proportion of all real overdue samples that can be covered when only the top k high-risk samples are processed. Equation (9) is used to evaluate the TabICLv2 model's ability to prioritize and capture key high-risk enterprises.

[0049] In terms of strategy simulation, this paper further examines the accuracy of the high-risk list under different disposal ratios to evaluate the list quality of the model in actual business. Given a disposal ratio... The number of true positives and false positives were respectively and Then its accuracy is defined as: (10) In equation (10), Indicates the proportion of cases handled; This represents the number of samples that were identified as high-risk by the model and actually became overdue under this treatment ratio, i.e., the number of true positives; This represents the number of samples identified as high-risk by the model under this treatment ratio but which were not actually overdue, i.e., the number of false positives; This represents the proportion of truly overdue enterprises in the high-risk list. Equation (10) reflects the nominal simplicity of the model under different treatment intensities and can be used to evaluate the strategic usability of the model in actual business.

[0050] Equation (10) reflects the proportion of genuine overdue enterprises among the samples identified as high-risk enterprises. By combining the changes in capture rate, precision, and false alarm level under different disposal ratios, this paper can further evaluate the strategy benefits and business availability of the model under resource constraints.

[0051] In summary, this paper's fairness assessment layer not only focuses on whether the TabICLv2 model "accurately identifies" risks, but also on whether it "sorts clearly and is usable." By combining multiple external validations, risk stratification, and strategy simulation, the application potential of the TabICLv2 model in early warning scenarios for overdue loans to SMEs can be more realistically reflected.

[0052] The interpretable layer of the SAFE robust early warning framework: multidimensional interpretive analysis of drift perception.

[0053] In research on corporate credit delinquency early warning, interpretability is a crucial bridge connecting model predictions with actual risk control decisions. While the TabICLv2 model used in this paper demonstrates strong predictive capabilities in complex tabular scenarios, the lack of effective interpretive tools makes it difficult for business personnel to understand why the model identifies certain companies as high-risk. Therefore, this paper constructs a drift-aware multi-dimensional interpretive analysis system within the SAFE robust early warning framework to enhance the interpretability and business usability of the model results.

[0054] Specifically, this paper first approximates the local contribution of each feature to the prediction result based on the standardized feature matrix and a linear surrogate model. Let the... In the nth sample The standardized values ​​of each feature are The proxy model coefficient is Then its local contribution is defined as: (11) In equation (11), Indicates the first The feature is related to the first The local contribution value of the prediction results for each sample. Indicates the first The sample at the th Standardized values ​​for each feature In the linear proxy model, the first... The regression coefficients corresponding to each feature , All are positive integers.

[0055] like A value greater than 0 indicates that this feature increases the risk of default in this sample; if... A value less than 0 indicates that the feature has a mitigating effect on default risk in this sample; the larger the absolute value, the stronger the influence. Equation (11) is used to measure the magnitude and direction of the local influence of a feature at the level of a single sample.

[0056] Equation (11) is used to measure the first The feature in the first The magnitude and direction of the local influence on the prediction results on each sample.

[0057] Furthermore, to measure the average influence of a feature across the entire sample, this paper defines global importance as: (12) In equation (12), Indicates the first The global importance of each feature in the sample dataset, where n represents the total number of samples. Indicates the first The feature is related to the first The local contribution value of the prediction results for each sample. , All are positive integers. Equation (12) measures the overall influence of a feature on the model output by calculating the average absolute contribution of the feature across all samples. The larger the value, the more important the feature is in explaining the company's delinquency risk.

[0058] in, Indicates the first The average absolute contribution of each feature across the entire sample. This indicator can be used to identify the key variables that best explain the overall corporate delinquency risk and to provide a basis for subsequent risk path mapping.

[0059] Importance ranking alone is insufficient to meet business interpretation needs. Therefore, this paper further introduces directional statistics to determine whether a feature "increases default rates," "decreases default rates," or is a "mixed directional trend." Its definition is: (13) In equation (13), Indicates the first The directional statistics of a feature; n represents the total number of samples; Indicates the first The sample at the th Standardized values ​​for each feature; Indicates the first The feature in the first Local contribution on each sample; sign(·) represents the sign function, a positive result indicates "high value increases risk", and a negative result indicates "high value decreases risk".

[0060] like A value greater than 0 indicates that a high value for this characteristic generally tends to increase the risk of default; if... A value less than 0 indicates that a high value for this characteristic generally suggests a lower risk of default; if... If the value is close to 0, it indicates that the directionality is weak or that the influence is mixed in different samples. Equation (13) is used to characterize the stability of the direction of the feature action from an overall perspective.

[0061] like This indicates that a high value for this characteristic generally tends to increase the risk of default; if This indicates that its high values ​​generally tend to reduce default risk; if If the value is close to 0 and has both positive and negative contributions, it indicates that the feature may exhibit mixed directional characteristics in different samples or different industries.

[0062] Based on equations (11) to (13), this paper forms a three-layer explanatory framework of "local contribution - global importance - directional analysis". On this basis, it summarizes thematic modules such as judicial enforcement risk, business registration change risk, financing and collateral risk, operational stability, tax credit and innovation qualification in conjunction with risk path diagram. Unlike traditional static explanation, this paper further compares the differences in risk drivers between industrial enterprises and wholesale and retail enterprises, thereby revealing the path of different characteristics on overdue risk in the context of data drift and industry heterogeneity.

[0063] Overall, the explanatory layer in this paper not only explains "which variables are important," but also explains "how these variables affect risk, whether they are stable in different scenarios, and how they map to business strategies." Therefore, this module not only improves model transparency but also enhances the application value of the SAFE robust early warning framework in practical risk control decision-making.

[0064] Specifically, the input layer is used to collect corporate credit data from at least two industries and construct a sample dataset; This paper uses two industry datasets as examples for illustration: a credit default dataset for industrial enterprises and a credit default dataset for wholesale and retail enterprises. Both datasets use enterprises as the basic observation unit and whether or not overdue payments have occurred as the prediction label, where a label of 1 indicates overdue payment and a label of 0 indicates no overdue payment.

[0065] Unlike studies that focus on a single industry, this paper includes both industrial and wholesale / retail companies. The aim is to examine the robustness of the model under different business structures and risk formation mechanisms from a cross-industry perspective, and to provide a basis for subsequent comparisons of risk-driven paths across different industries.

[0066] In terms of feature construction, this paper focuses on the credit risk control scenario of SMEs and constructs an early warning feature system from multiple dimensions, including judicial enforcement risk, business registration change risk, financing collateral risk, operational stability, tax credit, innovation qualifications, and some basic operational and financial information. Specific variables include judicial enforcement indicators such as the number of times listed under travel restrictions, the number of times a defaulter has been subject to enforcement, the number of court announcements, the number of times judicial freezes have occurred, and the number of court hearing announcements in recent months (e.g., 12, 36, 60), as well as business registration change indicators such as name changes, address changes, business scope changes, and increases or decreases in registered capital. Variables reflecting the enterprise's operational quality and credit constraints are also included, such as the total amount of movable property collateral, the total amount of loan collateral, the growth rate of operating revenue, total profit, the growth rate of net profit, the number of times the tax credit rating is A, and the number of patents. Compared to research designs that only use financial indicators or static credit indicators, this feature system emphasizes the integration of multi-source heterogeneous information, enabling a more comprehensive characterization of the formation mechanism of overdue risk for SMEs.

[0067] In terms of data preprocessing, this paper first performs unified cleaning and field standardization on the original variables, and handles outliers and missing values; secondly, it performs encoding conversion on categorical fields and performs necessary alignment and standardization on numerical variables; finally, it organizes the two industry datasets into a unified table format that can be directly input into TabICLv2.

[0068] Throughout the preprocessing process, this paper consistently adheres to the Anti-leakage principle in the SAFE robust early warning framework, which means that all fitting-type processing based on data distribution is only performed within the internal training scope before being applied to the validation set and external test set to avoid information leakage.

[0069] Drift perception layer: used to divide the sample dataset into an internal validation set, a time-external validation set, and a region-external validation set; To evaluate the generalization performance of the model under different data distribution conditions, this study adopts an enterprise-level mutually exclusive three-part partitioning strategy for both industrial enterprise datasets and wholesale and retail enterprise datasets, dividing the samples into an internal validation set, a time-based external validation set, and a regional external validation set.

[0070] The segmentation process uses enterprises as the smallest unit to ensure that the same enterprise appears in only one subset, thus avoiding the leakage of enterprise-level information. For enterprises with multiple records, default tags are aggregated at the enterprise level, and the maximum value of whether a default has occurred is used as the enterprise-level tag.

[0071] The overall target split ratio is approximately 80%:10%:10%. The internal validation set is used for model training, cross-validation, and threshold determination. The temporal external validation set and the regional external validation set are both used as one-time independent external validation sets and do not participate in model training, parameter adjustment, or threshold correction.

[0072] In terms of specific construction methods, the time-based external validation set is constructed based on variables with time-series characteristics such as the number of years the enterprise has been established, to simulate data distribution drift caused by the advancement of time or changes in the customer queue; the regional external validation set is constructed using a regional block hold-out strategy to simulate the model's external generalization ability under cross-regional or cross-customer group migration conditions.

[0073] To reduce the impact of extreme class imbalance on the evaluation results, random search is further used during the partitioning process to maintain a relative balance in terms of sample size and enterprise-level default rate among the internal validation set, time-based external validation set, and regional external validation set.

[0074] For example, the industrial enterprise dataset contains 11,088 samples, with 8,826 (79.60%) for the internal validation set, 1,109 (10.00%) for the time-based external validation set, and 1,153 (10.40%) for the regional external validation set, corresponding to enterprise-level default rates of 6.56%, 5.95%, and 6.94%, respectively. The wholesale and retail enterprise dataset follows the same partitioning principle, containing 19,333 samples, with 15,423 (79.77%) for the internal validation set, 1,933 (10.00%) for the time-based external validation set, and 1,977 (10.23%) for the regional external validation set. This setup ensures consistency in experimental design and comparability of results between the two datasets.

[0075] To realistically characterize the model's extrapolation capability under data drift scenarios, this paper does not employ a standard random partitioning method. Instead, it constructs a three-layer validation structure for each industry dataset: an internal validation set, a time-based external validation set, and a regional external validation set. The internal validation set is used for model training, internal cross-validation, parameter selection, and threshold determination. The time-based external validation set simulates distribution drift caused by time progression, changes in sample queues, or differences in enterprise lifecycles. The regional external validation set simulates distribution drift caused by different regions and customer groups. Through this design, the model not only needs to achieve good results on internal samples but also maintain stable performance on external samples with extrapolation properties.

[0076] In sample partitioning, this paper uses mutually exclusive partitioning by enterprise to ensure that the same enterprise will not appear simultaneously in both the internal validation set and the two external validation sets, thereby reducing the risk of information leakage caused by repeated exposure at the enterprise level. Furthermore, in the internal validation set, this paper uses cross-validation to generate out-of-flight (OOF) probabilities and determines the final classification threshold based on these probabilities; the external validation set is used only for one-time testing and does not participate in any parameter learning or threshold adjustment. Compared with the common approach of "random partitioning + repeated threshold testing on the external set" in traditional research, this design better reflects the model's performance in real-world business scenarios.

[0077] Furthermore, this paper conducts all validation work within a dual-industry parallel experimental framework, meaning that industrial enterprises and wholesale and retail enterprises each execute the same validation process. Table 1 shows the enterprise-level mutually exclusive partitioning and dual external validation design for the industrial and wholesale / retail enterprise datasets in this embodiment of the invention. As shown in Table 1, this design not only examines the robustness of the model when facing different distribution samples within the same industry, but also further compares the differences in model performance, explanatory structure, and strategy returns between different industries, thereby enhancing the universality and persuasiveness of the research conclusions.

[0078] Table 1

[0079] Leakage prevention layer: used to train the early warning analysis model using the internal validation set, generate the out-of-bounds prediction probability using cross-validation, and determine the final classification threshold based on the out-of-bounds prediction probability to obtain the trained early warning analysis model; The trained early warning analysis model is then validated using the time-based external validation set and the region-based external validation set to examine its generalization ability in external scenarios. This paper uses the TabICLv2 model as the core prediction model in the SAFE robust early warning framework and focuses on model development for data drift scenarios. Specifically, firstly, the model is trained based on the internal validation sets of each industry; secondly, out-of-flight (OOF) prediction probabilities are generated through internal cross-validation, and thresholds are determined accordingly; finally, the trained model and fixed thresholds are directly applied to the time-based and region-based external validation sets to test its generalization ability in external scenarios.

[0080] Considering the characteristics of SME credit data, such as high-dimensional heterogeneity, complex nonlinear relationships, class imbalance, and significant industry differences, this paper does not use the native TabICLv2 model to directly model all features. Instead, it further introduces a robust feature selection and drift-aware development mechanism under the SAFE robust early warning framework.

[0081] On the one hand, this paper implements a three-step robust feature selection process within the internal validation set to reduce the interference of redundant and noisy variables on model stability. On the other hand, a time extrapolation surrogate scenario is constructed within the internal validation set to help determine the performance stability of different model configurations under potential time or queue drift conditions. In this way, the model development process not only focuses on the results of internal cross-validation but also takes into account the adaptability to potential extrapolation scenarios.

[0082] Throughout the training process of the TabICLv2 model, this paper adheres to the principle of "internal development, external validation." Model parameter selection, feature selection, preprocessing fitting, and threshold determination are all completed within the scope of the internal validation set. The temporal and regional external validation sets do not participate in any form of parameter learning or threshold adjustment. In particular, the classification threshold is determined solely based on the internal out-of-flight (OOF) prediction results, and the external validation set is used only for one-time validation and does not interfere with the model development process in the reverse direction.

[0083] Therefore, the model development process in this paper is not a direct application of the native TabICLv2 model, but rather incorporates drift-aware sample design, robust feature selection, strict leakage prevention verification, and multi-external scenario testing mechanisms within the SAFE robust early warning framework, thereby improving the model's stability, credibility, and interpretability in real business environments.

[0084] Output layer: Used to input the credit data of the target enterprise into the trained early warning analysis model to obtain the credit delinquency probability of the target enterprise.

[0085] By collecting credit data from the target company and inputting it into the trained early warning analysis model, the probability of credit delinquency for that target company can be obtained. This target company can be considered as the enterprise requiring credit delinquency early warning, i.e., the test sample mentioned earlier.

[0086] Furthermore, to comprehensively evaluate the effectiveness of the proposed SAFE robust early warning framework, this paper sets up several classic table learning models as comparison methods, including linear models such as Logistic Regression, SGD, and GaussianNB; ensemble learning models such as Random Forest, Extra Trees, AdaBoost, HistGradientBoosting, and LightGBM; and deep table models such as TabTransformer and FT-Transformer. Since SAFE is the overall robust early warning framework proposed in this paper, and TabICLv2 is its core prediction model, the focus of the comparative experiments is to examine the overall performance of SAFE+TabICLv2 relative to different types of baseline methods under the same data partitioning, preprocessing procedures, and evaluation metric system.

[0087] In the comparative experiments, all models were trained and tested using the same industry data segmentation, the same preprocessing procedures, and the same evaluation metric system to ensure consistency in the comparison environment. For models requiring threshold determination, the internal OOF results were uniformly used to determine the classification threshold, and this threshold was fixed and applied to the external validation set. For ranking metrics that do not rely on threshold optimization, evaluation was performed directly based on the predicted probability. Through horizontal comparison with the aforementioned baseline models, this paper can more objectively examine the relative advantages of SAFE+TabICLv2 in dual-industry, multi-external validation, and data drift scenarios.

[0088] In addition to conventional model comparisons, this paper also focuses on the question of "whether a verification system is explicitly built to address data drift." In other words, the comparisons in this paper are not only reflected in the static performance differences on the same test set, but also in the differences in dynamic business indicators such as external migration, ranking stratification, key high-risk capture, and strategy simulation, thereby avoiding the limitations of evaluating the merits of models based solely on a single ACC or AUC.

[0089] In some embodiments, to comprehensively evaluate the model's performance in the task of early warning of overdue loans to small and micro enterprises, this paper constructs an evaluation index system from three levels: differentiation ability, ranking ability, and strategy value.

[0090] In terms of basic classification and ranking capabilities, this paper uses indicators such as AUC, KS, ACC, TPR, and FPR. Among them, AUC reflects the model's overall ability to distinguish between positive and negative samples. The AUC value is not affected by sample imbalance and is an important indicator for evaluating credit default prediction models. KS measures the maximum separation between high-risk and low-risk samples. Accuracy reflects the overall classification accuracy. TPR and FPR measure the ability to identify true overdue samples and the degree of false positives for normal samples, respectively.

[0091] In terms of risk ranking and evaluation, this paper further introduces risk stratification analysis. Specifically, the sample is divided into several risk levels or decimal groups according to the predicted probabilities output by the model, and the changes in the actual delinquency rate within each level are statistically analyzed. If the model has good ranking ability, the observed delinquency rate within a group should show a significant upward trend as the risk group changes from low to high. This analysis can more intuitively demonstrate whether the model truly has the ability to effectively rank enterprises according to their risk level.

[0092] In terms of business strategy value, this paper employs high-risk sample priority capture capability and strategy simulation analysis. The former measures the proportion of real overdue samples the model can cover when intervention is only applied to the top percentage of high-risk enterprises; the latter further analyzes the changing trends of Precision, TPR, and FPR under different disposal ratios, thereby evaluating the model's nominal simplicity, capture benefits, and false alarm costs under limited risk control resources. Through these indicators, this paper examines not only the model's "accuracy of prediction" but also its "clear ranking and value for money."

[0093] In some embodiments, this paper incorporates the Explainable module into the experimental design, rather than as an adjunct analysis after the experiment, in terms of model interpretation. Specifically, this paper first identifies the most important risk-driving features overall through SHAP-like global attribution; secondly, it constructs a risk path graph, classifying highly important features according to thematic modules such as judicial enforcement risk, business registration change risk, financing and collateral risk, operational stability, tax credit, and innovation qualifications, thereby transforming the importance results of scattered features into a risk path structure that is more in line with business understanding; thirdly, through feature directionality analysis, it distinguishes between three types of variables: "high values ​​drive up default," "high values ​​reduce default," and "mixed directions," to enhance the readability of the interpretation results. It should be noted that the interpretation analysis in this paper is a post-hoc approximate interpretation of the model output results, mainly used to identify key risk drivers and their direction of action, and is not equivalent to a precise analysis of the internal inference mechanism of TabICLv2.

[0094] Furthermore, this paper pays special attention to cross-industry explanatory differences. By conducting explanatory analyses of industrial enterprises and wholesale and retail enterprises respectively, it compares the differences in the importance, directionality, and path structure of key risk characteristics in the two industries, thereby revealing whether the risk formation mechanisms of different industries are consistent in the context of data drift and industry heterogeneity. In this way, model interpretation not only serves to "understand the model" but also to "develop industry-specific, level-specific, and executable risk control strategies."

[0095] The experimental results and analysis are as follows: To verify the effectiveness and robustness of the SAFE robust early warning framework in data drift scenarios, this paper conducts experiments on industrial enterprise datasets and wholesale and retail enterprise datasets, and evaluates the model in multiple dimensions based on internal validation sets, time-based external validation sets, and regional external validation sets. The experimental results analysis mainly includes three aspects: first, examining the overall early warning performance of the SAFE robust early warning framework under different validation scenarios; second, comparing the SAFE robust early warning framework with several representative baseline models to verify the relative advantages of the proposed method; and third, discussing the model's business usability and key risk drivers by combining risk stratification, the ability to prioritize high-risk samples, strategy simulation, and interpretable results. Based on a parallel research design across two industries, this paper not only focuses on the model's performance on a single dataset but also emphasizes the differences in model robustness and risk formation mechanisms across different industry scenarios.

[0096] The experimental results and analysis for the industrial enterprise dataset are as follows: First, overall early warning performance analysis Table 2 compares the results of various models for the industrial dataset in this embodiment of the invention, and Table 3 compares the results of various models for the retail dataset in this embodiment of the invention. As shown in Tables 2 and 3, on the industrial enterprise dataset, this paper first analyzes the overall performance of the SAFE prediction framework on the internal validation set, the time-based external validation set, and the regional external validation set. The results show that the model exhibits good discrimination ability, ranking ability, and policy value on all three data subsets, indicating that the SAFE robust early warning framework can not only stably identify overdue risks in internal samples, but also has strong generalization ability in external extrapolation scenarios.

[0097] In terms of basic discriminative ability, the SAFE robust early warning framework achieves an AUC of 0.8737 and a KS of 0.5765 on the internal validation set for industrial enterprises; an AUC of 0.8013 and a KS of 0.4931 on the external time validation set; and an AUC of 0.8968 and a KS of 0.6632 on the external regional validation set. Overall, the model maintains a high level of discriminative ability on both external validation sets, with performance on the external regional validation set even outperforming that of the internal validation set. However, the performance on the external time validation set shows a decline, indicating that time queue drift has a more significant impact on the model's extrapolation ability.

[0098] Based on the classification results, under the premise of uniformly using internal OOF results to determine the threshold, the SAFE robust early warning framework achieved accuracy of 0.7438, 0.7755, and 0.8222 on the internal validation set, time-based external validation set, and regional external validation set for industrial enterprises, respectively; TPR of 0.8394, 0.6515, and 0.8375; and FPR of 0.2629, 0.2167, and 0.1789. This indicates that while maintaining a high positive class identification ability, the model did not exhibit an unacceptable level of false positives, and its overall classification performance was relatively balanced. Particularly noteworthy is the model's achievement of both a high TPR and a low FPR on the regional external validation set, demonstrating a clearer separation boundary between high-risk and normal enterprises in this scenario.

[0099] In summary, the overall results on the industrial enterprise dataset demonstrate that the SAFE robust early warning framework not only performs well on internal samples but also maintains strong early warning capabilities under both temporal and regional extrapolation conditions. This indicates that the drift awareness, leakage prevention, and fair assessment designs proposed in this paper effectively enhance the model's transferability and robustness in real-world business environments.

[0100] Table 2

[0101] Table 3

[0102] Second, performance comparison with the comparative model. To further validate the effectiveness of our proposed method in industrial enterprise scenarios, this paper compares the SAFE prediction framework with several representative models, including linear models (Logisti regression, SGD, GaussianNB), ensemble learning models (Random Forest, Extra Trees, AdaBoost, HistGradientBoosting, LightGBM), and deep table models (TabTransformer, FT-Transformer). The comparison results show that TabICLv2 achieves strong overall performance on both the internal validation set and the two external validation sets, particularly demonstrating stable advantages in core ranking metrics such as AUC and KS.

[0103] On the internal validation set, the SAFE prediction framework achieved an AUC of 0.8737, higher than Extra Trees (0.8503), HistGradientBoosting (0.8501), Random Forest (0.8467), and LightGBM (0.8405), and significantly better than TabTransformer (0.8029) and FT-Transformer (0.8085); its KS was 0.5765, also the highest among all models. It should be noted that the TabICLv2 model's ACC in this scenario was 0.7438, not the highest value, indicating that the proposed method focuses more on improving the ranking ability and recognition sensitivity of overdue samples, rather than simply pursuing overall classification accuracy.

[0104] On the time-extra validation set, the TabICLv2 model shows a more significant advantage. Its AUC reaches 0.8013, higher than FT-Transformer (0.7631), HistGradientBoosting (0.7618), Random Forest (0.7571), and ExtraTrees (0.7536); its KS is 0.4931, also among the highest of all models. This indicates that even in time-queue drift scenarios, the proposed method can still maintain the model's discriminative ability well, demonstrating stronger time extrapolation robustness.

[0105] On the external validation set, the TabICLv2 model achieved an AUC of 0.8968, the highest among all models. Its KS was 0.6632, slightly lower than the FT-Transformer's 0.6670, but the difference was small, maintaining an overall optimal or near-optimal level. Considering AUC, KS, TPR, and FPR, the TabICLv2 model demonstrates a more balanced risk identification capability in external scenarios for industrial enterprises.

[0106] Overall, the comparison results on the industrial enterprise dataset show that: traditional linear models are insufficient to fully characterize the complex nonlinear relationships in the credit data of these enterprises; while tree models have some competitiveness, their stability in external drift scenarios is still weaker than the TabICLv2 model; among deep table models, FT-Transformer outperforms TabTransformer, but overall it still does not surpass the proposed method in terms of comprehensive metrics. Therefore, a more conservative conclusion is that the TabICLv2 model achieves the best or near-best overall performance under the condition of dual external validation for industrial enterprises.

[0107] Third, risk stratification and strategy simulation analysis Figure 2 This is a schematic diagram illustrating the risk stratification results for an industrial enterprise dataset, as provided in an embodiment of the present invention. Figure 2 As shown in the figure, internal_cv represents the internal validation set, external_time represents the time-based external validation set, and external_2 represents the regional external validation set. The horizontal axis, Risk Decile, represents the risk level, divided into ten risk levels from D1 to D10. The vertical axis, Observed Overdue Rate, represents the proportion of samples with actual overdue payments to the total sample. From the risk stratification diagram of the industrial enterprise dataset, it can be seen that the risk score output by the model has good monotonicity and hierarchical discrimination ability. Regardless of whether it's the internal cross-validation sample, the time-based external validation sample, or the regional external validation sample, as the risk decimal places increase from D1 to D10, the observed default rate generally shows an upward trend, especially with a significant jump at the highest risk group, D10. This indicates that the risk score output by the SAFE robust early warning framework can not only effectively rank defaulting enterprises but also maintain a relatively stable high-risk clustering effect under different distribution scenarios.

[0108] Figure 3 A cumulative capture rate curve for an industrial enterprise dataset is provided as an embodiment of the present invention, such as... Figure 3As shown, the horizontal axis, Top fraction treated as high-risk, represents the proportion of companies classified as high-risk, while the vertical axis, Bad recall (captured overdue / total overdue), represents the recall rate of bad samples or overdue samples (total number of real overdue samples / number of captured overdue samples). The Gains curve indicates that the model has a strong ability to prioritize the identification of high-risk companies. Even when only the top 10%–20% of companies are considered high-risk intervention targets, the model is already able to identify a significant proportion of defaulting samples; as the intervention rate increases to 30%–50%, the cumulative identification rate further increases significantly. In particular, the external validation set curve is located at a higher position overall, indicating that the model can still effectively concentrate defaulting samples into the high-risk segment in cross-regional scenarios.

[0109] Figure 4 This invention provides a KS curve chart for an industrial enterprise dataset, as shown in the embodiment of the invention. Figure 4 As shown, the horizontal axis, Population fraction (sorted by risk), represents the percentage of the population (sorted by risk) / the percentage of the total sample. The vertical axis, TPR-FPR, represents the difference between TPR and FPR. The KS curve further validates the model's effectiveness in ranking and identification. The peak KS values ​​for the internal validation set, the time-external validation set, and the regional-external validation set remain at high levels, with the highest peak value observed in the regional-external validation set. This indicates that the model maintains a strong ability to separate positive and negative samples even under cross-regional drift conditions. From a business perspective, this means that TabICLv2 under the SAFE robust early warning framework is not only effective for identically distributed samples but also remains usable under more stringent migration scenarios.

[0110] Figure 5 This is a simulation diagram of a strategy for calculating the change in precision as a function of disposal ratio for an industrial enterprise dataset, provided in an embodiment of the present invention. Figure 6 This is a simulation diagram of a TPR (Total Return) strategy as a function of disposal ratio for an industrial enterprise dataset, provided in an embodiment of the present invention. Figure 7 This invention provides a simulation diagram of the FPR (Frequency Rate) variation strategy with disposal ratio for an industrial enterprise dataset, as shown in the embodiment of the invention. Figures 5 to 7As shown, the horizontal axis represents the proportion of samples that are treated, and the vertical axis represents the accuracy rate. TPR (Treatment Percentage) represents the proportion of all truly bad samples that are judged as bad by the model, and FPR (Fault-Precision Percentage) represents the proportion of all truly good samples that are misjudged as bad by the model. Further analysis of the strategy simulation results reveals that the three curves—Precision, TPR, and FPR—together depict the business benefit-cost balance of the model under different treatment intensities. As the treat fraction increases, the TPR continues to rise, indicating that more defaulting companies are identified earlier. Simultaneously, the Precision gradually decreases, suggesting that the proportion of non-defaulting companies in the high-risk pool increases as the intervention scope expands. The FPR rises with the treatment rate, but its growth trend is relatively smooth.

[0111] In summary, if banks wish to balance identification benefits with the cost of misreporting, they can prioritize monitoring the top 10%-20% of enterprises; if they emphasize including as many high-risk samples as possible in the monitoring process in advance, they can extend the disposal threshold to the 20%-30% range.

[0112] Fourth, analysis of key risk drivers. In summary, the default risk in the industrial enterprise sample is not determined by a single financial indicator, but is influenced by a combination of factors, including judicial enforcement, credit penalties, financing collateral, and operational status. Therefore, the value of the model in industrial scenarios lies not only in its predictive accuracy, but also in its ability to integrate the behavior of discrete enterprises into interpretable and actionable risk warning clues.

[0113] Secondly, global interpretability analysis.

[0114] Figure 8 This invention provides a SHAP-like beeswarm global interpretation graph for an industrial enterprise dataset. In the graph, the SHAP-like value (impact on model output) represents the influence of the SHAP-like value on the model output, and the feature value represents the specific value of the sample on that feature, such as... Figure 8As shown, the SHAP-like beeswarm plot of the industrial enterprise dataset illustrates the overall direction and strength of the influence of each feature on the model output across the entire sample. Overall, variables related to judicial enforcement, the number of times a company has been restricted from high-end travel, the number of times it has received court notices or announcements as a plaintiff or defendant, and judicial assistance-related asset freezes are generally among the most important, indicating that a company's probability of default increases significantly once it has frequent records in the judicial enforcement chain. Meanwhile, indicators such as the number of times a company has achieved an A-level tax credit rating, total profit, and revenue growth rate are more likely to be risk mitigation variables, suggesting that operational stability and credit compliance can, to some extent, offset the default pressure brought about by external negative events.

[0115] It should be noted that the SHAP-like beeswarm global interpretability graph is a visualization of the model's global interpretability, drawn in the form of bee colony scatter plots using SHAP-like feature attribution values. It is used to understand, as a whole, which features are important, how feature values ​​affect the model output, and the distribution of contributions among samples.

[0116] Figure 9 This is a schematic diagram of the Top 10 positive features for an industrial enterprise dataset provided in an embodiment of the present invention. Figure 10 This invention provides a schematic diagram of the Top 10 negative features of an industrial enterprise dataset, as shown in the embodiment of the invention. Figure 9 and Figure 10 As shown in the Top 10 positive and negative key feature graphs, it can be further seen that the features driving up default risk are mainly concentrated in two dimensions: judicial enforcement and financing collateral, while the features reducing default risk come more from dimensions such as tax credit, business growth, and innovation qualifications. This indicates that risk identification for industrial enterprises has a clear "two-sided mechanism": on the one hand, judicial enforcement and negative events can rapidly increase risk; on the other hand, stable operations and high credit quality can form a certain risk buffer.

[0117] Furthermore, risk path analysis.

[0118] Figure 11 This invention provides a path diagram for increasing default risk in a dataset of industrial enterprises. Figure 12 This invention provides a path diagram for reducing default risk in industrial enterprise datasets. Figure 13 An overview diagram of all feature directions provided in an embodiment of the present invention, such as Figures 11 to 13As shown, the risk path diagram further reveals the hierarchical relationship between "characteristics - meso-level risk themes - overall default risk". From the positive path, judicial enforcement risk is the core driver of default, with the number of times restrictions on high-end consumption, court announcements, judicial assistance freezes, and records of being a dishonest debtor constituting the main path to increased default risk. From the negative path, operational stability, tax credit, innovation qualifications, and profit performance constitute the main sources of risk mitigation. This path structure illustrates that default risk in industrial enterprises is not randomly generated, but gradually forms along a chain of "deeper judicial punishment - damaged operational credit - accumulated financing and operational pressure". Further, from the overview of characteristic directions, the risk drivers for industrial enterprises are not simply "the higher the value, the higher the risk". Some judicial enforcement and negative event variables show a clear characteristic of high values ​​pushing up default risk. For example, indicators related to enforcement, restrictions on high-end consumption, court announcements, and judicial assistance freezes usually indicate that the enterprise has entered a state of stronger external constraints or damaged credit, thus making it easier to increase the probability of default. Meanwhile, variables such as the number of times a taxpayer has achieved an A-level credit rating, profit performance, and revenue growth rate show that higher values ​​primarily reduce default risk, indicating that better operational quality, credit compliance, and growth capacity have a risk mitigation effect. Other variables exhibit mixed directional characteristics, suggesting that their impact on default risk may be influenced by the combined effects of the company's operational stage, industry background, and other characteristics. Therefore, the formation mechanism of default risk in industrial enterprises exhibits significant directional heterogeneity, encompassing both high-value risk exposure and high-value risk mitigation, as well as a mixed action path dependent on specific contexts.

[0119] Finally, a local interpretation of typical samples.

[0120] Figure 14 This is an example diagram illustrating a low-risk sample local interpretation for an industrial enterprise dataset, provided as an embodiment of the present invention. Figure 15 An example diagram illustrating a local interpretation of high-risk samples in an industrial enterprise dataset, provided as an embodiment of the present invention, is shown below. Figure 14 and Figure 15 As shown in the partial explanation results of a single sample, the TabICLv2 model's assessment of default risk for industrial enterprises does not rely on a single variable, but rather on the combined effect of multidimensional risk signals. For low-risk samples, the model gives a default probability close to 0, primarily because variables such as the number of times a tax credit rating of A has been obtained in the past 60 months and the past 12 months significantly suppress the default probability. Simultaneously, the zero values ​​for multiple indicators of judicial enforcement, business registration changes, and negative events further reinforce the low-risk assessment. Although the length of establishment contributes somewhat positively to this sample, the overall impact is weak and does not alter its low-risk attribute.

[0121] For high-risk samples, the model yields a higher probability of default. The main driving factors include variables such as the number of times judicial assistance-assisted asset freezes occurred in the past 60 months, the number of times a company was listed as a restricted traveler in the past 12 months, the number of times administrative penalties were imposed in the past 60 months, the total amount of credit collateral in the past 12 months, and the number of times a company was listed as a dishonest debtor in the past 12 months. Among these, judicial assistance-assisted asset freezes and listing as a restricted traveler have the most significant positive contributions, indicating that the company's negative exposure in the judicial enforcement chain is the core basis for the increased probability of default. At the same time, some variables show a certain negative contribution, indicating that the local explanatory results are not a simple linear summation, but are jointly influenced by the combination of feature relationships and the contextual structure.

[0122] Overall, interpretable analysis of the industrial enterprise dataset reveals the following: First, judicial enforcement risk is the most significant driver of default; second, financing collateral, operational fluctuations, and business registration changes further amplify risk; and third, tax credit, profit performance, and innovation qualifications have a mitigating effect on default risk. These results suggest that default warnings for SMEs in industrial settings should focus on both external judicial penalties and internal operational quality.

[0123] The experimental results and analysis for the dataset of wholesale and retail enterprises are as follows: First, an analysis of the overall early warning performance.

[0124] On the wholesale and retail enterprise dataset, this paper also analyzes the overall performance of the TabICLv2 core prediction model under the SAFE robust early warning framework from three levels: internal validation set, time-based external validation set, and regional external validation set. Overall, the model also demonstrates good discrimination ability, ranking ability, and business application potential in the wholesale and retail enterprise scenario, indicating that the SAFE robust early warning framework is not limited to industrial enterprise samples, but has good applicability to different industry customer groups.

[0125] Compared to industrial enterprises, wholesale and retail enterprises typically exhibit higher operational volatility, more frequent changes in ownership, and more pronounced regional customer segment differences, thus placing higher demands on the robustness of the model's extrapolation. Experimental results show that TabICLv2 achieves AUCs of 0.8838, 0.8202, and 0.8420 on the internal validation set, the time-based external validation set, and the regional external validation set for wholesale and retail enterprises, respectively, and KS values ​​of 0.6013, 0.4935, and 0.5229, respectively, demonstrating strong overall discriminative and ranking capabilities.

[0126] Based on the classification results, under the premise of uniformly using internal OOF results to determine the threshold, TabICLv2's ACC on the internal validation set, time-based external validation set, and regional external validation set for wholesale and retail enterprises are 0.8730, 0.9234, and 0.8366, respectively; TPR is 0.7224, 0.5417, and 0.6364, respectively; and FPR is 0.1211, 0.0668, and 0.1553, respectively. Overall, the AUC and KS on the time-based external validation set are somewhat lower than those on the internal validation set, indicating that the wholesale and retail industry is more sensitive to time or queue drift; however, the regional external validation set still maintains a high level, indicating that the model still has good applicability under cross-regional migration conditions.

[0127] Second, a performance comparison with the comparative model.

[0128] To examine the effectiveness of the SAFE robust early warning framework in wholesale and retail enterprise scenarios, this paper also compares it with several representative models, including linear models (Logistic Regression, SGD, GaussianNB), ensemble learning models (Random Forest, Extra Trees, AdaBoost, HistGradientBoosting, LightGBM), and deep tabular models (TabTransformer, FT-Transformer). By maintaining the same comparison criteria as the industrial enterprise section, the robustness and relative advantages of the SAFE robust early warning framework in different industry scenarios can be examined more objectively.

[0129] On the internal validation set, TabICLv2's AUC is 0.8838, significantly higher than Random Forest (0.8657), HistGradientBoosting (0.8623), Extra Trees (0.8523), and FT-Transformer (0.8475); its KS is 0.6013, also the highest among all models, indicating that TabICLv2 has a strong risk ranking ability on samples from wholesale and retail enterprises.

[0130] On the two external validation sets, TabICLv2 maintained a strong advantage: on the time-based external validation set, its AUC was 0.8202 and KS was 0.4935, both higher than other comparative models; on the regional external validation set, its AUC was 0.8420 and KS was 0.5229, also among the best of all models. It should be noted that some baseline models showed extreme results on this dataset. For example, Logistic Regression exhibited a TPR close to 1 but also an FPR close to 1 and an extremely low ACC on all three subsets, indicating significant classification bias under class imbalance and fixed threshold conditions. GaussianNB, on the other hand, nearly "classified all as non-default" in multiple scenarios, with a high ACC but a TPR of 0, lacking practical risk control significance. In contrast, TabICLv2 demonstrated a more balanced overall advantage in ranking ability, recognition benefits, and false positive control.

[0131] This invention also includes results analysis and interpretability analysis of the dataset for wholesale and retail enterprises. A comparison and comprehensive discussion of the results from the two industries are as follows: Experimental results on both industrial and wholesale / retail enterprise datasets show that the SAFE robust early warning framework demonstrates good discriminative, ranking, and policy value across both industry scenarios. However, significant differences remain between the two industries in terms of external drift sensitivity and key risk drivers. Overall, our proposed method achieves optimal or near-optimal results on most core metrics, but it does not necessarily lead in every single metric.

[0132] From an overall early warning performance perspective, the internal validation results on both industry datasets are good, indicating that TabICLv2 can effectively learn the overdue risk patterns of SMEs in heterogeneous table feature spaces. However, in the external validation stage, the two industries do not show completely consistent sensitivity to different types of drift. Generally speaking, temporal external validation is usually more challenging than regional external validation, indicating that temporal queue drift is a type of distribution shift problem that deserves more attention in corporate credit overdue early warning. Among industrial enterprises, model performance is more affected by changes in operational stability, financing collateral, and capital structure; while wholesale and retail enterprises are more susceptible to changes in business entity, adjustments to business scope, and regional differences in operations.

[0133] From the interpretable results, judicial enforcement risk is the most core common high-risk source in both industries, indicating that variables such as court announcements, enforcement of judgments against defaulters, restrictions on high-end consumption, and judicial freezes have strong cross-industry stability. However, the importance of variables such as financing and collateral risk, profitability, and operational stability is more prominent in industrial enterprises, indicating that the risks of industrial enterprises are more easily reflected through capital occupation, asset burden, and operational fluctuations. In contrast, in wholesale and retail enterprises, business registration change risk and entity stability indicators are more sensitive, indicating that the risks in this industry are more manifested through behaviors such as name changes, address changes, changes in business scope, and capital adjustments.

[0134] From a business strategy perspective, the results from both industries jointly indicate that early warning systems for overdue loans to SMEs cannot simply rely on uniform static rules. Instead, they should, while retaining common signals such as judicial enforcement, develop differentiated early warning strategies for different industries. For industrial enterprises, greater attention should be paid to financing collateral, operational quality, and profit fluctuations; for wholesale and retail enterprises, dynamic monitoring of entity stability, business registration changes, and abnormal operational behavior should be strengthened. The value of the SAFE robust early warning framework lies in its ability to maintain strong robustness across different industry scenarios and to provide a basis for industry-specific risk control through interpretive analysis.

[0135] The results of this paper indicate that in the task of early warning of overdue loans to SMEs, model evaluation should not be limited to the internal accuracy under random partitioning conditions, but should be examined in the context of data drift. A multi-layered validation system consisting of an internal validation set, a temporal external validation set, and a regional external validation set shows that the TabICLv2 model maintains good discriminative and ranking abilities in both industry datasets, indicating that the tabular-based model has strong extrapolation potential in complex and heterogeneous credit data. Meanwhile, temporal external validation is generally more challenging than regional external validation, indicating that distribution shifts caused by the progression of time, changes in customer queues, and differences in sample entry times are more likely to weaken the discriminative boundaries learned by the model during the development phase.

[0136] The comparison between the two industries further reveals that judicial enforcement risk is a core high-risk signal shared by both industrial and wholesale / retail enterprises, demonstrating strong cross-industry stability. However, factors such as financing collateral, operational stability, business registration changes, and entity stability exhibit significant industry heterogeneity. This indicates that credit early warning for SMEs possesses both transferable common risk structures and differentiated risk paths dependent on industry characteristics. Therefore, risk governance cannot simply rely on uniform static rules but should develop industry-specific strategies based on commonalities identified.

[0137] From a methodological perspective, this paper is significant not only for validating the predictive performance of TabICLv2, but also for illustrating the importance of the SAFE robust early warning framework for data drift in financial risk control research. The introduction of enterprise-level mutual exclusion partitioning, internal threshold setting, external one-time validation, and internal time extrapolation proxy scenarios helps reduce information leakage and the risk of overly optimistic assessments, making the experimental results closer to real-world deployment conditions.

[0138] It should be noted that the characterization of data drift in this paper is still mainly based on structured features and sample partitioning design. The explanatory analysis is also a post-hoc approximation of the model output results, and the research conclusions still need to be further tested in more industries and more complex business scenarios. In the future, we can continue to expand in areas such as online drift monitoring, dynamic threshold updates, and strategy feedback loops, thereby promoting the shift of credit risk early warning for SMEs from static prediction to continuous governance.

[0139] This paper addresses the issue of credit delinquency early warning for SMEs in data-drift scenarios. Based on two industry datasets—industrial and wholesale / retail enterprises—a robust SAFE early warning framework is constructed, consisting of an internal validation set, a time-based external validation set, and a regional external validation set. The model's discriminative ability, ranking ability, business strategy value, and interpretability are systematically tested using TabICLv2. The results show that multiple external validations and a leak-proof development process can realistically reflect the model's performance in external deployment scenarios, and also indicate that a single internal validation may overestimate the model's robustness and extrapolation ability.

[0140] Empirical results show that TabICLv2 achieves superior overall performance on both industry datasets, particularly demonstrating strong advantages in AUC, KS, risk stratification, and high-risk sample priority identification. Further analysis reveals that external validation sets are generally more challenging than external validation sets, indicating that time drift is a type of distribution shift that deserves greater attention in SME credit early warning. Interpretable analysis also shows that judicial enforcement risk is a core high-risk source for cross-industry stability, and there are significant differences between industrial and wholesale / retail enterprises in risk paths such as financing collateral, operational stability, and business registration changes.

[0141] This invention provides a corporate credit delinquency early warning and analysis system, which includes: The sample module is used to collect corporate credit data from at least two industries and construct a sample dataset; The partitioning module is used to divide the sample dataset into an internal validation set, a time-based external validation set, and a region-based external validation set. The training module is used to train the early warning analysis model using the internal validation set, generate the out-of-bounds prediction probability using cross-validation, and determine the final classification threshold based on the out-of-bounds prediction probability to obtain the trained early warning analysis model. The trained early warning analysis model is then validated using the time-based external validation set and the region-based external validation set to examine its generalization ability in external scenarios. The prediction module is used to input the credit data of the target enterprise into the trained early warning analysis model to obtain the credit delinquency probability of the target enterprise.

[0142] This embodiment is a system embodiment corresponding to the above method embodiment. Its specific implementation is the same as that of the above method embodiment. For details, please refer to the above method embodiment. This system embodiment will not repeat the details.

[0143] Each module in the aforementioned corporate credit overdue early warning analysis system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0144] This invention provides a computer device, which may be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes a computer storage medium and internal memory. The computer storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the computer storage medium. The database stores data generated or acquired during the execution of a corporate credit delinquency early warning analysis method, such as corporate credit data and sample datasets. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a corporate credit delinquency early warning analysis method.

[0145] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the enterprise credit overdue early warning analysis method described in the above embodiment. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the embodiment of an enterprise credit overdue early warning analysis system.

[0146] In one embodiment, a computer storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the steps of the enterprise credit delinquency early warning analysis method described in the above embodiment. Alternatively, when executed by a processor, the computer program implements the functions of each module / unit in the above embodiment of the enterprise credit delinquency early warning analysis system.

[0147] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0148] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning analysis of corporate credit delinquency, characterized in that, The method uses the SAFE robust early warning framework, which includes: Input layer: Used to collect corporate credit data from at least two industries and construct a sample dataset; Drift perception layer: used to divide the sample dataset into an internal validation set, a time-external validation set, and a region-external validation set; Leakage prevention layer: used to train the early warning analysis model using the internal validation set, generate the out-of-bounds prediction probability using cross-validation, and determine the final classification threshold based on the out-of-bounds prediction probability to obtain the trained early warning analysis model; The trained early warning analysis model is then validated using the time-based external validation set and the region-based external validation set to examine its generalization ability in external scenarios. Output layer: Used to input the credit data of the target enterprise into the trained early warning analysis model to obtain the credit delinquency probability of the target enterprise.

2. The enterprise credit overdue early warning analysis method according to claim 1, characterized in that, The SAFE robust early warning framework also includes: Fairness assessment layer: used to examine the recognition effect of the early warning analysis model from multiple dimensions using AUC, KS, accuracy, true positive rate, and false positive rate indicators; The AUC index is used to reflect the overall ability of the early warning analysis model to distinguish between positive and negative samples. The AUC index is not affected by sample imbalance and is an important indicator for evaluating the early warning analysis model. The KS index is used to measure the maximum separation degree of the early warning analysis model for samples with different risk levels.

3. The enterprise credit overdue early warning analysis method according to claim 1, characterized in that, The SAFE robust early warning framework also includes: Interpretation layer: Used to calculate the local contribution of each feature to the prediction result output by the early warning analysis model based on the standardized feature matrix and the linear surrogate model. The calculation formula is as follows: ； in, Indicates the first The feature is related to the first The local contribution value of the prediction results for each sample. Indicates the first The sample at the th Standardized values ​​for each feature In the linear proxy model, the first... The regression coefficients corresponding to each feature , All are positive integers; like This indicates that the first The feature is related to the first The risk of default for each sample has increased; like This indicates that the first The feature is related to the first The default risk of each sample decreased; The larger the absolute value, the stronger the influence.

4. The enterprise credit overdue early warning analysis method according to claim 3, characterized in that, The interpretation layer is also used to measure the overall influence of each feature on the prediction results output by the early warning analysis model by calculating the average absolute contribution of each feature across all samples. The calculation formula is as follows: ； in, Indicates the first The global importance of each feature in the sample dataset, where n represents the total number of samples. Indicates the first The feature is related to the first The local contribution value of the prediction results for each sample. , All are positive integers.

5. The enterprise credit overdue early warning analysis method according to claim 1, characterized in that, The interpretation layer is also used to introduce directional statistics to determine whether each feature is a directional mixture, and its calculation formula is as follows: ； in, Indicates the first The directional statistics of a feature, where n represents the total number of samples. Indicates the first The sample at the th Standardized values ​​for each feature Indicates the first The feature in the first The local contribution on each sample, where sign(·) represents the sign function. , All are positive integers; like Greater than 0 indicates that the first This characteristic is more likely to increase the risk of default; like Less than 0 indicates that the first These characteristics tend to reduce the risk of default; like If the difference between the absolute value of and 0 is within the preset range, then it indicates that the ... The directionality of a feature is weak, or it may exhibit mixed effects in different samples.

6. The enterprise credit overdue early warning analysis method according to claim 1, characterized in that, The input layer includes: The data acquisition unit is used to collect corporate credit data from at least two industries, using enterprises as the basic observation unit and whether or not overdue payments have occurred as a label to construct the sample dataset. The feature acquisition unit is used to collect multi-dimensional features of the enterprise credit data, including judicial enforcement risk, business registration change risk, financing and collateral risk, operational stability, tax credit, innovation qualifications, and basic operational and financial information. First, the enterprise credit data is uniformly cleaned and its fields are standardized, and outliers and missing values ​​are processed. Second, the categorical fields in the enterprise credit data after the outlier and missing values ​​are processed are encoded and converted, and the numerical variables are aligned and standardized as necessary. Finally, the standardized enterprise credit data is processed into a table format acceptable to the early warning analysis model.

7. The enterprise credit overdue early warning analysis method according to claim 1, characterized in that, The early warning analysis model uses TabICLv2 as the prediction model.

8. A corporate credit overdue early warning and analysis system, characterized in that, The system uses the SAFE robust early warning framework, which includes: Input layer: Used to collect corporate credit data from at least two industries and construct a sample dataset; Drift perception layer: used to divide the sample dataset into an internal validation set, a time-external validation set, and a region-external validation set; Leakage prevention layer: used to train the early warning analysis model using the internal validation set, generate the out-of-bounds prediction probability using cross-validation, and determine the final classification threshold based on the out-of-bounds prediction probability to obtain the trained early warning analysis model; The trained early warning analysis model is then validated using the time-based external validation set and the region-based external validation set to examine its generalization ability in external scenarios. Output layer: Used to input the credit data of the target enterprise into the trained early warning analysis model to obtain the credit delinquency probability of the target enterprise.

9. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory is used to store programs or instructions, and the processor is used to execute the programs or instructions stored in the memory to implement the enterprise credit overdue early warning analysis method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the enterprise credit overdue early warning analysis method according to any one of claims 1 to 7.

Images