Backtracking analysis method and device for risk control model strategy dynamic correction

Abstract

The invention relates to the field of financial science and technology, and discloses a backtracking analysis method and device for risk control model strategy dynamic correction, and the key point of the technical scheme is that the method comprises the following steps: S1, collecting historical loan complete cycle data, and extracting historical risk features; s2, a backtracking model is constructed, training is carried out through historical risk features, an applicable backtracking model is obtained, and the backtracking model is used for analyzing feature categories and influence degrees related to overdue risks; s3, in response to the backtracking analysis instruction, obtaining real-time loan complete cycle data, and extracting real-time risk features; s4, inputting the real-time risk features into a backtracking model to obtain a backtracking analysis result; s5, adjusting the risk control strategy of the risk control model according to a preset strategy improvement mechanism based on the backtracking analysis result; and S6, after the backtracking model is applied, performing performance monitoring on the backtracking model, and performing incremental training on the backtracking model according to a monitoring result.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and more specifically, to a backtracking analysis method and apparatus for dynamic correction of risk control model strategies. Background Technology

[0002] In the current credit sector, online risk control largely relies on models for decision-making. However, in actual business operations, it has been found that loans approved now still experience delinquencies in the future, indicating that existing models and strategies have certain limitations.

[0003] In the lending sector, the development of online risk control technology has evolved from traditional rule-driven approaches to the application of machine learning models. While machine learning models have been gradually applied to risk control scenarios and have improved the accuracy of risk assessment to some extent, significant shortcomings remain: 1. Existing models focus primarily on pre-loan and mid-loan data, while underutilizing post-loan performance data, making it difficult to comprehensively capture risks; 2. The strategies are mostly statically set and cannot adapt to dynamic changes in the market environment and customer behavior, which leads to frequent overdue payments of loans approved now in the future, highlighting the limitations of existing models and strategies.

[0004] To address the shortcomings of existing machine learning models, some banks have proposed strategies for reanalyzing post-loan data. However, current backtracking techniques still have certain limitations, including: lack of a closed-loop mechanism, insufficient dynamic adaptability, shallow utilization of post-loan data, low incremental learning efficiency, and coarse-grained strategy optimization. Therefore, it is necessary to redesign the backtracking analysis scheme to avoid these problems. Summary of the Invention

[0005] The purpose of this invention is to provide a retrospective analysis method and apparatus for dynamic correction of risk control model strategies, which can realize full-process control from risk identification to non-performing loans, including accurate filtering in pre-loan approval, real-time early warning in loan monitoring, and strategy optimization in post-loan management.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a backtracking analysis method for dynamic correction of risk control model strategies, comprising the following steps: S1. Collect historical loan data throughout the entire loan cycle and extract historical risk characteristics; S2. Construct a retrospective model and train it using historical risk features to obtain an applicable retrospective model. The retrospective model is used to analyze the feature categories and impact degree related to overdue risk. S3. Responding to the retrospective analysis command, obtain real-time loan full-cycle data and extract real-time risk characteristics; S4. Input the real-time risk characteristics into the backtracking model to obtain the backtracking analysis results; S5. Based on the backtracking analysis results, adjust the risk control strategy of the risk control model according to the preset strategy improvement mechanism; S6. After the backtracking model is applied, its performance is monitored, and incremental training is performed on the backtracking model based on the monitoring results.

[0007] As a preferred technical solution of the present invention, the loan lifecycle data includes: pre-loan application data, mid-loan approval decision data, and post-loan performance data.

[0008] As a preferred technical solution of the present invention, the backtracking model is a machine learning model that employs logistic regression, decision tree, random forest, XGBoost and a deep learning neural network model.

[0009] As a preferred technical solution of the present invention, after extracting historical risk features, a training set and a test set are established, and overdue risk features are screened out through a feature selection algorithm; the overdue risk features are used as labels, and the backtracking model is trained with the training set; during the training process, the model parameters are adjusted through cross-validation. After training, the backtracking model is evaluated using a test set; at the same time, the stability and adaptability of the backtracking model are verified using a time series verification method. When the backtracking model passes the performance evaluation, stability and adaptability verification, an applicable backtracking model is obtained.

[0010] As a preferred embodiment of the present invention, the performance indicators corresponding to the performance evaluation include: accuracy, recall, F1 score, and AUC score.

[0011] As a preferred technical solution of the present invention, after the backtracking model is applied, the performance indicators of the backtracking model are monitored in real time, and an early warning is triggered when the performance indicators exceed the preset performance threshold. After an alert is triggered, the corresponding risk characteristics are broken down dimensionally to pinpoint the failure scenario of the backtracking model.

[0012] As a preferred technical solution of the present invention, when applying the backtracking model, the difference in feature distribution between real-time data and training set data is calculated, and when the group stability index is greater than the preset stability threshold, the backtracking model is retrained. When the importance of features fluctuates, the feature weights in the risk model are adjusted according to the ranking of feature importance.

[0013] As a preferred technical solution of the present invention, when the backtracking model triggers retraining, an incremental learning framework is used to update the tree structure corresponding to the newly added data before retraining.

[0014] As a preferred technical solution of the present invention, the strategy improvement mechanism includes a risk indicator adjustment scheme, a threshold optimization scheme, and a rule supplementation and update scheme.

[0015] The risk indicator adjustment scheme includes risk assessment indicator adjustment strategies and indicator weight adjustment strategies corresponding to the backtesting analysis results. The threshold optimization scheme includes a decision threshold adjustment strategy corresponding to the backtracking analysis results. The rule supplementation and update scheme includes a rule change strategy corresponding to the backtracking analysis results.

[0016] A backtracking analysis method and apparatus for dynamic correction of risk control model strategies includes: a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor implements the above method when executing the computer program.

[0017] In summary, this invention offers the following advantages: it improves model accuracy, enabling precise capture of risk characteristics from data backtracking; it allows for full-cycle data feature mining, enhances model generalization ability, and allows for dynamic adjustment of feature weights.

[0018] It can achieve full-process control from risk identification to non-performing loans, including accurate filtering in pre-loan approval, real-time early warning in loan monitoring, and strategy optimization in post-loan management.

[0019] It can enhance market competitiveness by improving both customer experience and business efficiency, including: rapid approval of high-quality customers, differentiated product pricing strategies, and agile support for business innovation. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the A / B testing mechanism of the present invention; Figure 3 This is a schematic diagram of the monitoring scheme of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings.

[0022] This invention provides a backtracking analysis method for dynamic correction of risk control model strategies, comprising the following steps: S1. Collect historical loan data throughout the entire loan cycle and extract historical risk characteristics; Loan lifecycle data includes: pre-loan application data, mid-loan approval decision data, and post-loan performance data.

[0023] Pre-loan application data includes borrower basic information, credit history, and application behavior. Basic information includes structured data such as borrower identity (name, ID number, age), occupation (employer, industry, income), and assets (real estate, vehicles), covering household registration and residential address down to the house number. Credit history includes data on overdue payments, credit card usage, and loan balances over the past 24 months from the central bank's credit reporting system, as well as synchronized data on e-commerce consumption, mobile phone calls, and public utility payments from third-party platforms. Application behavior includes recording application channels (APP / offline), time, and application frequency over the past 6 months, clearly defining the data authorization scope (credit reporting, mobile phone operators, etc.).

[0024] Loan approval decision data includes risk control model scores, approval decisions, and contract terms; risk control model scores include: saving the total risk score output by each risk control model and sub-dimensional scores such as repayment ability and willingness; approval decisions include: recording the approver, time, opinion (approval / rejection), and suggested amount and interest rate; contract terms include: retaining core terms such as loan amount, term, repayment method, and interest rate.

[0025] Post-loan performance data includes borrower repayment behavior, risk events, and external dynamics. Repayment behavior includes: tracking monthly repayment date deviations and full payment status, calculating overdue days (M1 / M2+), and early repayment records; risk events include: recording the first overdue date, the distribution of overdue periods, and status changes such as judicial intervention; external dynamics include: capturing borrower occupation, residential address changes, and litigation information through social security and telecom operator data.

[0026] Extract historical risk characteristics from the entire lifecycle data of historical loans.

[0027] Specifically, features related to delinquency risk are extracted from the integrated data. In addition to traditional features such as credit scores and debt-to-income ratios, the focus is on identifying features related to actual performance after loan disbursement, such as the stability of repayment behavior (whether repayments are made on time and the fluctuation of repayment amounts) and the timing of delinquencies (how many months after the loan date the first delinquency occurs). After extracting risk features from historical loan lifecycle data, training and testing sets are established. Furthermore, feature selection algorithms are used to screen features that significantly impact delinquency risk, improving the model's predictive ability.

[0028] Before extracting features from the historical loan data throughout its entire lifecycle, a series of optimization processes are performed on the data, including data cleaning and handling of missing values.

[0029] S2. Construct a backtracking model and train it using historical risk characteristics. After training, an applicable backtracking model is obtained. The backtracking model is used to analyze the characteristic categories and impact levels related to overdue risk. The backtracking model employs machine learning models such as logistic regression, decision trees, random forests, and XGBoost, as well as deep learning neural network models.

[0030] The backtracking model uses overdue risk features from historical loan data as labels and trains on a training set. During training, model parameters are adjusted through cross-validation to improve the model's generalization ability. For example, when using a random forest model, parameters such as the number of trees, maximum depth, and minimum number of samples for node splits can be adjusted to optimize model performance.

[0031] After training, the backtracking model's performance was evaluated using a test set. The performance metrics included accuracy, recall, F1 score, and AUC. This ensured the backtracking model could accurately identify loan cases with delinquency risk and performed well across different loan types and risk levels. Furthermore, time-series validation was used to verify the model's stability and adaptability. A valid backtracking model was deemed applicable when it passed all performance metrics evaluations, stability tests, and adaptability tests.

[0032] S3. Responding to the retrospective analysis command, obtain real-time full-cycle loan data; based on the real-time full-cycle loan data, extract real-time risk characteristics; S4. Input the real-time risk characteristics into the backtracking model to obtain the backtracking analysis results; S5. Based on the backtracking analysis results, adjust the risk control strategy of the risk control model according to the preset strategy improvement mechanism.

[0033] The strategy improvement mechanism includes risk indicator adjustment schemes, threshold optimization schemes, and rule supplementation and update schemes.

[0034] The risk indicator adjustment scheme includes risk assessment indicator adjustment strategies and indicator weight adjustment strategies corresponding to the backtesting analysis results. For example, if the backtesting model shows that in a certain type of loan, the borrower's occupational stability has a significant impact on delinquency risk, but its weight is low in the existing risk control model, then the weight of this indicator in the risk assessment system will be increased accordingly. By dynamically adjusting the risk indicator weights, the risk control model can better reflect the actual risk situation.

[0035] The threshold optimization scheme includes a decision threshold adjustment strategy corresponding to the backtesting analysis results. Specifically, through the backtesting model, the threshold settings in the existing risk control model are re-examined, such as credit scoring thresholds and debt-to-income ratio thresholds. By analyzing the delinquency rate and approval rate under different thresholds, the optimal threshold combination is found. This controls risk while improving the approval rate of high-quality customers, avoiding the loss of high-quality customers or the misjudgment and approval of high-risk customers due to unreasonable thresholds.

[0036] The rule supplementation and update plan includes rule change strategies corresponding to the results of backtracking analysis. Specifically, based on new risk patterns and regularities discovered by the backtracking model, existing risk control rules are supplemented or updated into the risk control model. For example, if some overdue customers are found to frequently change their mobile phone numbers, and the existing rules do not cover this situation, a corresponding rule can be added, such as "Borrowers who have changed their mobile phone numbers more than twice in the past six months need further manual review."

[0037] S6. After the backtracking model is applied, its effectiveness is monitored, and incremental training is performed on the backtracking model based on the monitoring results.

[0038] Specifically, after the backtracking model is applied, its performance metrics are monitored in real time. When a performance metric exceeds a preset performance threshold, an alert is triggered. For example, real-time tracking of model performance metrics such as AUC, KS value, precision, and recall is performed. An alert is triggered when the AUC drops by more than 5% from the baseline or the KS value drops by more than 0.05. For instance, a consumer finance company sets an AUC baseline of 0.89. When the test set AUC drops to 0.84, the system automatically sends an alert to the model management device for relevant administrators to view and handle.

[0039] After an alert is triggered, the corresponding risk characteristics are broken down dimensionally to pinpoint the failure scenarios of the backtracking model. For example, indicators are broken down by business line (consumer loans / business loans), customer group (age / industry), and risk level (high / medium / low) to locate model failure scenarios. If the KS value of the "25-30 year old young customers" group drops from 0.32 to 0.21, it indicates that the characteristics of this group need to be optimized in a targeted manner.

[0040] When applying the backtracking model, the difference in feature distribution between real-time data and training set data is calculated. When the group stability index exceeds a preset stability threshold (e.g., 0.2), the backtracking model is retrained. For example, if the PSI of the feature "Number of Online Loan Applications in the Past Six Months" rises from 0.09 to 0.25, it indicates a change in customer application behavior patterns, requiring re-validation of the feature's effectiveness. When feature importance fluctuates, the feature weights in the risk model are adjusted accordingly. For instance, changes in feature contribution are monitored using the SHAP value, and an alert is issued when the ranking of a feature's importance changes by more than 20%. For example, if the importance of the feature "Occupational Stability" drops from 3rd to 10th, it suggests the need to analyze the impact of industry changes on risk assessment.

[0041] When the backtracking model triggers retraining, an incremental learning framework is used to update the tree structure corresponding to the newly added data before retraining.

[0042] After the backtracking model was applied, a continuous optimization mechanism was designed for the backtracking model, including a periodic retraining plan, an incremental learning optimization scheme, and a dynamic strategy adjustment scheme. The periodic retraining program includes: Regular retraining is performed quarterly based on the latest post-loan data (≥100,000 new data entries), incorporating new risk features. For example, during the retraining in Q2 2024, due to changes in consumption habits, the feature of "frequency of local life consumption" was added, which improved the AUC by 0.03.

[0043] Emergency retraining is initiated when market conditions change drastically (such as policy adjustments or economic crises) or monitoring indicators trigger warnings. For example, if mortgage policies in a certain region tighten, the regional mortgage model will be retrained within two weeks, adding a "down payment ratio" feature to control delinquency risk.

[0044] Incremental learning optimization schemes include: By employing the H2O.ai incremental learning framework and using the updater_additive mechanism on the XGBoost model, only the tree structure corresponding to newly added data is updated, resulting in a 60% improvement in training efficiency. For example, when adding 50,000 new loan data entries each month, the incremental training time is reduced from 4 hours for full retraining to 1.5 hours.

[0045] For the logistic regression model, SGDClassifier is used to implement stochastic gradient descent incremental learning, dynamically updating the weight parameters to ensure that the model adapts to changes in customer behavior in a timely manner.

[0046] Dynamic strategy adjustment plan: Threshold optimization involves reviewing historical approval data quarterly to analyze delinquency rates and approval rates under different credit score thresholds and optimizing threshold combinations. For example, if a threshold of 650 points is found to have a delinquency rate of 5% and an approval rate of 35%, which better aligns with risk appetite than a threshold of 600 points (delinquency rate of 8% and approval rate of 45%), the approval strategy is adjusted accordingly.

[0047] Rule iteration involves adding or removing risk control rules based on model backtesting results. For example, if the model shows that the delinquency rate of customers with "credit card overdraft rate > 80% in the past 3 months" is 2.5 times that of other customers, this rule is added; if the delinquency differentiation of a certain rule (such as "additional review for customers with non-local household registration") decreases, it is removed.

[0048] After the backtracking model was applied, an optimization effect verification mechanism was designed for the backtracking model, including A / B testing mechanism, cross-time window verification mechanism, and business acceptance mechanism. A / B testing mechanism: After each model optimization, 5% of the application samples are selected for parallel testing of the old and new models for comparison. Risk indicators: Whether the new model's AUC and KS are superior to the old model; Business metrics: Increase in the success rate of high-quality customers, and changes in the rejection rate of high-risk customers.

[0049] Case Study: After optimizing its model, a bank found that A / B testing showed the new model's AUC improved from 0.87 to 0.89, and the pass rate for high-quality customers increased by 7%, leading to its full deployment.

[0050] Cross-time window validation mechanism: Historical data is divided into 5 time windows (e.g., 2022Q1-2023Q4). The model is trained in each window and tested in subsequent windows to evaluate model stability. The AUC fluctuation is required to be <5% and the KS fluctuation <0.03 for each window to ensure consistent model performance across different time periods.

[0051] Business Acceptance Mechanism: The optimized model must pass acceptance testing by risk control, business, and IT departments. Risk control side: AUC≥0.85, KS≥0.3, PSI<0.1; On the business side: approval time ≤ 3 minutes, and a decrease of ≥ 10% in the churn rate of high-quality customers; IT side: Model interface call success rate ≥99.9%, resource utilization rate <80%.

[0052] Corresponding to the above method, the present invention also provides a backtracking analysis method and apparatus for dynamic correction of risk control model strategy, comprising: a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor implements the above-described methods S1-S7 when executing the computer program.

[0053] The technical solution of this invention provides a fully adaptive feedback system, realizing closed-loop management. This includes real-time monitoring, model retraining, strategy calibration, and effect verification.

[0054] Real-time monitoring: Track metrics such as AUC and PSI in real time using Prometheus+Grafana. When the AUC drops by more than 5% from the baseline or the KS value drops by more than 0.05, an alert is automatically triggered (e.g., if a consumer finance company sets the AUC baseline at 0.89, the system will automatically start retraining when it drops to 0.84).

[0055] For model retraining, the H2O.ai incremental learning framework was used, which only updated the tree structure corresponding to the newly added data in the XGBoost model, improving training efficiency by 60%. During quarterly retraining, new features such as "frequency of local life consumption" were incorporated, which improved AUC by 0.03.

[0056] Strategy calibration involves dynamically optimizing thresholds by reviewing historical approval data (e.g., increasing the credit score threshold from 600 to 650) and adjusting feature weights based on SHAP values ​​(e.g., increasing the weight of "number of times mobile phone number has been changed" from 5% to 15%).

[0057] Performance verification is conducted through A / B testing and cross-time window verification to ensure that AUC ≥ 0.85 and KS ≥ 0.3 after the new strategy goes live. For example, after optimization, a bank's AUC increased from 0.87 to 0.89, and the pass rate of high-quality customers increased by 7%.

[0058] The technical solution of this invention has multi-dimensional strategy deduction capabilities, including: intelligent threshold optimization, automated rule iteration, and dynamic calibration of indicator weights.

[0059] Intelligent Threshold Optimization: Analyzes delinquency rates and approval rates under different thresholds. For example, it finds that a delinquency rate of 5% and an approval rate of 35% at a threshold of 650 points are more in line with risk preferences, reducing the loss of high-quality customers compared to the traditional static threshold (600 points).

[0060] Automated rule iteration: rules are added or removed based on model backtracking results. For example, if it is found that the delinquency rate of a customer with a credit card overdraft rate of >80% in the past 3 months is 2.5 times that of other customers, the rule is automatically added; if the distinguishability of a rule decreases, it is removed.

[0061] Dynamic calibration of indicator weights: SHAP values ​​are used to analyze changes in feature contribution. For example, when the importance of the "career stability" feature drops from 3rd to 10th place, the industry risk coefficient is adjusted to ensure that the model is aligned with business scenarios.

[0062] The technical solution of this invention breaks through the efficiency bottleneck of traditional full retraining, realizes the model's "learning on demand" of new data, and has XGBoost incremental training and logistic regression dynamic optimization capabilities; XGBoost Incremental Training: Using the updater_additive mechanism, when adding 50,000 new post-loan data entries per month, the training time is reduced from 4 hours to 1.5 hours. For example, an internet bank used this mechanism to quickly capture the feature of "e-commerce installment overdue times", which improved the accuracy of risk assessment for customers aged 25-35 by 28%.

[0063] Logistic Regression Dynamic Optimization: SGDClassifier is used to implement stochastic gradient descent incremental learning, update weight parameters in real time, adapt to changes in customer behavior (such as the high-frequency small-amount borrowing characteristics of the new generation of customers), and ensure that the PSI of the model is <0.1 in different time windows.

[0064] Example 1: Consumer loan scenario: Application Background: In Q4 of 2023, a leading consumer finance company discovered that the delinquency rate of young customers aged 25-35 was 18% higher than that of other age groups. Traditional models, which did not incorporate e-commerce consumption behavior data, led to biases in risk identification.

[0065] Application of technical solutions: Data collection: Retrospectively examine the installment records of this customer group on e-commerce platforms over the past 12 months (covering third-party data such as Taobao and JD.com) and extract post-loan characteristics such as "number of overdue e-commerce installment payments in the past 3 months" and "high-frequency return rate".

[0066] Model optimization: "≥2 times of overdue e-commerce installment payments" was added as a new risk feature and incorporated into the XGBoost model training. The SHAP value analysis determined that its feature importance ranked 4th (weighting 12%).

[0067] Strategy adjustment: For customers whose characteristics trigger this feature, the credit score threshold will be automatically lowered from 650 to 600, and a manual review process will be added. Implementation effect

[0068] The model's AUC improved from 0.82 to 0.87, increasing the accuracy of identifying high-risk customers in this customer group by 22%. Within three months of loan disbursement, the M1+ delinquency rate decreased from 3.7% to 2.3%, and the non-performing loan generation rate decreased by 1.4 percentage points. While the acceptance rate for high-quality customers remained unchanged, the rejection rate for high-risk customers increased by 15%.

[0069] Example 2: Business Loan Scenario: Application Background: When a city commercial bank launched its micro and small enterprise operating loan business in Q2 of 2024, the traditional model had an overdue prediction accuracy of only 68% for "first-time borrowers" and could not capture the operational volatility risk of the catering industry.

[0070] Application of technical solutions: Data integration: Retrospectively analyze the post-loan transaction data (including Alipay / WeChat payment records) of borrowers in the catering industry and extract dynamic features such as "monthly average transaction volume decreased by more than 30% year-on-year for three consecutive months" and "weekend transaction volume accounted for less than 20%".

[0071] Model building: The LightGBM algorithm was used to train a specialized model, a new "operating flow stability index" was added, and the parameters were optimized through 5-fold cross-validation (e.g., setting num_leaves=31, learning_rate=0.05).

[0072] Strategy iteration: Treat "a decrease in cash flow of more than 30%" as a high-risk signal, triggering loan monitoring rules such as "requiring supplementary operating financial statements" or "reducing credit limits".

[0073] Implementation results: The model's AUC improved from 0.75 to 0.86, enhancing its ability to differentiate risks for catering industry clients by 28%. The M3+ delinquency rate of newly issued business loans decreased from 5.1% to 2.7%, and the annual amount of non-performing loan write-offs decreased by 80 million yuan; The approval time has been shortened from 24 hours to 4 hours (instant approval for low-risk customers), and customer satisfaction has increased by 12 percentage points.

[0074] Through the two examples above, it can be seen that in the consumer loan scenario, the model's AUC value steadily increased from 0.72 in Q1 2022 to 0.89 in Q1 2024, showing an overall upward trend, with fluctuations of less than 5% between quarters, indicating that the model has good stability and adaptability across different time periods. In the business loan scenario, the model's AUC value increased from 0.68 in Q1 2022 to 0.86 in Q1 2024, also showing a steady upward trend, with fluctuations of less than 5% between quarters, verifying the model's cross-time stability in the business loan scenario.

[0075] The advantages of the technical solution of this invention include: (a) Improve model accuracy: from data backtracking to the precise capture of risk characteristics.

[0076] 1. Full-cycle data-driven feature mining: By integrating and analyzing data across the entire loan lifecycle—pre-loan, during-loan, and post-loan—backtracking models can uncover risk characteristics overlooked by traditional risk control models. For example, by utilizing the characteristic of "first delinquency occurring in the 3rd month after loan repayment" in post-loan repayment records, combined with survival analysis algorithms, early default signals can be identified. Furthermore, by analyzing the correlation between borrowers' post-loan occupational changes (changes in social security payment units) and delinquency, "number of occupational changes in the past 6 months" can be incorporated into the risk assessment system. A consumer finance company's practical test showed that after introducing post-loan features, the risk control model's accuracy in identifying high-risk customers increased by 25%, and the AUC value increased from 0.72 to 0.89.

[0077] 2. Improved model generalization ability driven by algorithm optimization: An ensemble learning strategy combining XGBoost and neural networks is employed to dynamically adjust model parameters by backtracking historical data. For example, during the training phase, grid search is used to optimize the number of trees in the random forest (increasing from 100 to 200) and the split threshold, combined with 5-fold cross-validation to reduce overfitting. For instance, when the PSI (Population Stability Index) of the backtracking model decreased from 0.18 to 0.09 in cross-quarter testing, it indicates a significant improvement in the predictive stability of the backtracking model across different time windows, effectively avoiding model failure caused by changes in the market environment.

[0078] 3. Dynamic calibration of feature importance: The decision-making logic of the risk control model is analyzed using SHAP values, and feature weights are adjusted in real time. For example, when backtesting reveals that borrowers who have changed their mobile phone numbers more than twice in the past six months have a delinquency rate three times higher than those who haven't changed their numbers, the weight of this feature is increased from 5% to 15%, and L1 regularization is used to ensure model sparsity. A practice at an internet bank shows that this adjustment improved the risk control model's ability to identify fraudulent customers by 30%, while reducing the false positive rate by 40%.

[0079] (II) Optimize asset quality: manage the entire process from risk identification to non-performing loans.

[0080] 1. Precise filtering in pre-loan approval: Based on the risk score output by the backtracking model, the approval threshold is dynamically optimized. For example, when backtracking analysis shows that the actual delinquency rate for customers with credit scores above 650 is only 3%, while the delinquency rate for customers with scores between 600 and 650 reaches 8%, the approval threshold is raised from 600 to 650, and a manual review process is added for customers with scores between 600 and 650. After implementing this system, a city commercial bank saw its M3+ delinquency rate for newly issued loans decrease from 5.2% to 2.1%, and its non-performing loan generation rate decrease by 59.6%.

[0081] 2. Real-time alerts for loan monitoring: A real-time early warning mechanism is established by mining post-loan risk characteristics through backtracking models (such as repayment amounts fluctuating by more than 30% for two consecutive periods). When a borrower triggers an early warning rule, the system automatically initiates post-loan checks, such as requiring the customer to provide supplementary income documentation or adjusting the credit limit. After deploying this mechanism, a consumer finance company reduced its response time for mid-loan risk events from 3 days to 4 hours, identified and handled 42% of potential default cases in advance, and saved more than 50 million yuan in economic losses per year.

[0082] 3. Optimization of post-loan management strategies Based on the analysis of historical collection data using a backtracking model, differentiated collection strategies were developed. For example, backtracking revealed that among customers who made their first payment within 30 days of delinquency, those with a telephone call connection rate >70% achieved an 85% repayment success rate, while those with a connection rate <30% had only a 20% success rate. Accordingly, collection resources were prioritized for customers with high connection rates. After implementing this strategy, a joint-stock bank saw a 35% increase in collection efficiency, a 22% reduction in collection costs, and an increase in overdue loan recovery rate from 68% to 81%.

[0083] (III) Enhance market competitiveness: improve both customer experience and business efficiency.

[0084] 1. Fast approval for high-quality clients: By identifying low-risk customer characteristics (such as civil servant occupation or no overdue records for five consecutive years) through a backtracking model, a "whitelist" fast-track approval channel was established. For example, one bank eliminated manual review for eligible customers, achieving instant system approval and reducing the approval time from 24 hours to 3 minutes. Data shows that this strategy increased the conversion rate of high-quality customers by 40%, raised customer satisfaction from 72 to 91 points, and drove an annual growth of 25% in loan balance.

[0085] 2. Product pricing differentiation strategy: Based on a backtesting model for accurate assessment of customer risk, a risk pricing mechanism is implemented. For example, an annualized interest rate of 7.2% is applied to low-risk customers, 10.8% to medium-risk customers, and credit is denied to high-risk customers. After implementation, a private bank saw the proportion of high-risk customers decrease from 18% to 9%, while the average loan yield increased from 8.5% to 9.8%, achieving yield optimization while controlling risk.

[0086] 3. Agile support for business innovation: The rapid iteration capability of the retrospective model supports risk assessment for new business scenarios. For example, when launching a new "home renovation loan" product, a specialized scoring model can be quickly built by retrospectively analyzing the transaction characteristics of "home renovation-related merchants" in historical consumer loan data. This allows the new product to go from project initiation to launch in just 45 days, shortening the time by 60% compared to the traditional process and seizing a market advantage.

[0087] (iv) Adapting to market changes: dynamic response from data iteration to strategy adjustment.

[0088] 1. Rapid adaptation to macroeconomic fluctuations: By continuously analyzing the correlation between economic cycles and delinquency rates using backtesting models, for example, the backtesting revealed that the delinquency rate of customers in the catering industry was 27% higher than that of other industries. Subsequently, the industry risk coefficient was adjusted, raising the entry threshold for catering industry customers by 15%. As a result, a certain bank controlled its new non-performing loan ratio to 1.8%, 2.5 percentage points lower than the industry average.

[0089] 2. Compliance response to policy and regulatory changes: When regulatory policies require stronger data minimization compliance, the retrospective model uses feature importance analysis to screen out core fields (such as ID number, income, and historical overdue number) that contribute more than 1% to risk assessment, reducing unnecessary data access by 60% while ensuring that the model's AUC value decreases by only 1.2%, thus meeting compliance requirements while maintaining risk identification capabilities.

[0090] 3. Real-time learning of customer behavior evolution: As younger customers become the main force in credit, the retrospective model analyzes the consumption characteristics of this new generation of customers (such as frequent small-amount borrowing and a high proportion of e-commerce consumption), adding features such as "number of e-commerce installment payments in the past 3 months" and "social media activity." For example, it can improve the accuracy of risk assessment for customers aged 25-35 by 28%, and reduce the non-performing loan rate for this customer group from 3.3% to 1.9%.

[0091] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A backtracking analysis method for dynamic revision of a risk control model strategy, characterized in that: Comprising the following steps: S1, collecting historical loan full-cycle data, and extracting historical risk characteristics; S2, constructing a backtracking model, and training it through historical risk characteristics to obtain an applicable backtracking model, which is used to analyze feature categories and influence degrees related to overdue risk; S3, in response to a backtracking analysis instruction, obtaining real-time loan full-cycle data and extracting real-time risk characteristics; S4, inputting the real-time risk characteristics into the backtracking model to obtain a backtracking analysis result; S5, based on the backtracking analysis result, improving the mechanism according to a preset strategy, and adjusting the risk control strategy of the risk control model; S6, after the application of the backtracking model, the performance of the backtracking model is monitored, and the backtracking model is incrementally trained according to the monitoring result.

2. The backtracking analysis method for dynamic revision of a risk control model strategy according to claim 1, characterized in that: The loan full-cycle data includes pre-loan application data, in-loan approval decision data, and post-loan performance data.

3. The backtracking method for dynamic revision of risk control model strategy according to claim 2, characterized in that: The backtracking model is a machine learning model using logistic regression, decision tree, random forest, XGBoost, and a neural network model using deep learning.

4. The backtracking method for dynamic revision of risk control model strategy according to claim 3, characterized in that: After extracting the historical risk characteristics, a training set and a test set are established, and an overdue risk feature is selected through a feature selection algorithm; The overdue risk feature is used as a label, and the training set is used to train the backtracking model. In the training process, the model parameters are adjusted through cross-validation; After the training is completed, the performance of the backtracking model is evaluated through the test set; at the same time, the stability and adaptability of the backtracking model are verified through a time series verification method. When the backtracking model passes the performance index evaluation, stability and adaptability verification at the same time, an applicable backtracking model is obtained.

5. The backtracking method for dynamic revision of risk control model strategy according to claim 4, characterized in that: The performance index evaluation corresponds to performance indicators including accuracy, recall rate, F1 value, and AUC value.

6. The backtracking method for dynamic revision of risk control model strategy according to claim 5, characterized in that: After the application of the backtracking model, the performance indicators of the backtracking model are monitored in real time. When the performance indicators exceed the preset performance threshold, a warning is triggered; After triggering the warning, the corresponding risk characteristics are dimensionally split to locate the failure scenarios of the backtracking model.

7. The backtracking method for dynamic revision of risk control model strategy according to claim 6, characterized in that: When the backtracking model is applied, the feature distribution difference between the real-time data and the training set data is calculated. When the population stability index is greater than the preset stability threshold, the backtracking model is retrained; When the feature importance fluctuates, the feature weight of the risk model is adjusted according to the importance ranking of the features.

8. The backtracking method for dynamic revision of risk control model strategy according to claim 7, characterized in that: When the backtracking model triggers retraining, an incremental learning framework is used to update the tree structure corresponding to the newly added data, and then training is performed.

9. The backtracking method for dynamic revision of risk control model strategy according to claim 8, characterized in that: The strategy improvement mechanism includes a risk indicator adjustment scheme, a threshold optimization scheme, and a rule supplement and update scheme; The risk indicator adjustment scheme sets risk assessment indicator adjustment strategies and indicator weight adjustment strategies corresponding to the backtracking analysis result; The threshold optimization scheme sets decision threshold adjustment strategies corresponding to the backtracking analysis result; The rule supplement and update scheme sets rule change strategies corresponding to the backtracking analysis result.

10. A backtracking analysis method and device for dynamic revision of a risk control model strategy, characterized in that: Comprising: A processor and a memory, the memory storing a computer program executable by the processor, and the processor implementing the method of any one of claims 1-9 when executing the computer program.

Images