Credit dynamic assessment risk control method and system based on propensity score matching

Abstract

The application discloses a credit dynamic evaluation risk control method and system based on propensity score matching, relates to the technical field of credit evaluation, and comprises the following steps: S1, periodically collecting credit behavior data, and performing pretreatment on the collected data; S2, constructing a field priority order table, quantifying the joint fluctuation degree of transaction behavior and payment behavior, and obtaining a credit behavior evaluation value; S3, generating an effective difference value sequence, performing a sliding window aggregation operation on the effective difference value sequence, and identifying a loss of stability window sequence; and S4, reading the loss of stability window sequence to construct a loss of stability segment feature sequence, evaluating the credit fluctuation risk level of each loss of stability segment, performing credit risk grade determination, and generating a graded risk report. The application solves the problem that the existing credit risk evaluation technology cannot identify credit behavior loss of stability segments by using the time sequence joint fluctuation of transaction behavior and payment behavior, and thus the dynamic credit evaluation result is prone to false risk fluctuation.

Description

A propensity score matching-based dynamic credit assessment and risk control method and system Technical Field

[0001] This invention relates to the field of credit assessment technology, specifically to a dynamic credit assessment and risk control method and system based on propensity score matching. Background Technology

[0002] With the widespread application of big data, cloud computing, and machine learning technologies in the financial sector, credit risk assessment has gradually evolved from an analysis model relying on static reports and human experience to a comprehensive assessment system based on multi-source heterogeneous data and intelligent algorithms. In enterprise and SME credit management scenarios, the industry has proposed various intelligent risk control solutions that combine historical financial data, operational data, credit data, and industry environment data to support credit approval, limit management, and risk monitoring, providing a technological foundation for financial institutions to establish a relatively complete enterprise credit assessment framework.

[0003] For example, application CN120912315A discloses a corporate credit assessment method based on an intelligent risk control model, which involves the field of corporate credit assessment. It includes basic data acquisition steps, data preprocessing steps, preliminary assessment steps, and comprehensive assessment steps. By acquiring the historical financial data, historical operating data, and historical credit data of the target company for each year, and conducting data analysis on various sub-data corresponding to the historical financial data, historical operating data, and historical credit data, and by supplementing the analysis with industry risk data and development potential data, the accuracy of corporate credit assessment is effectively improved.

[0004] For example, application CN119444397A discloses a method for constructing a credit risk assessment model for SMEs based on big data from industry and finance. This method includes acquiring multi-source data for SME credit risk assessment; standardizing, normalizing, and reducing the dimensionality of the multi-source data to obtain a second feature set; training and integrating three credit risk assessment models using the second feature set to obtain a comprehensive credit scoring model for multi-dimensional and accurate assessment of enterprise credit risk; and training a dynamic monitoring and early warning model using real-time risk assessment data to classify enterprise credit risk in real time. This application comprehensively solves the problems of single data source, lagging analysis, low accuracy, and lack of early warning mechanisms in traditional SME credit risk assessment models by introducing a dynamic risk monitoring mechanism based on multi-source heterogeneous data.

[0005] However, the existing corporate credit assessment and SME credit risk assessment schemes mentioned above mostly focus on integrating multi-source features at the model level and outputting a comprehensive credit score. They are difficult to structurally characterize the fine-grained fluctuations in transaction and repayment behavior within a continuous sampling period. They have not yet identified unstable segments in the credit behavior sequence from the perspective of tendency score matching and behavioral stability. They also have not finely controlled the impact of technical factors such as changes in field priority and perturbations in difference structure on dynamic risk control results. This makes it easy to generate false risk increases or decreases in high-concurrency or complex behavioral scenarios.

[0006] Therefore, in order to address the above issues, there is an urgent need for a credit dynamic assessment and risk control method and system based on propensity score matching. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a credit dynamic assessment and risk control method and system based on propensity score matching. This solves the problem that existing credit risk assessment technologies cannot utilize the temporal joint fluctuations of transaction behavior and repayment behavior to identify unstable segments of credit behavior, which leads to false risk fluctuations in dynamic credit assessment results.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a credit dynamic assessment and risk control method based on propensity score matching, comprising: S1, periodically collecting credit behavior data, and performing multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization on the collected credit behavior data to generate preprocessed credit behavior data; S2, constructing a field priority sorting table based on the preprocessed credit behavior data, and performing feature operations based on the differences between credit behavior data in adjacent sampling periods to obtain a credit behavior assessment value, inputting the credit behavior assessment value and the corresponding credit behavior data into a propensity score estimation model to output a propensity score; S3, generating matching period pairs based on the credit behavior assessment value and the corresponding propensity score to obtain an assessment difference sequence, and performing a sliding window aggregation operation on the effective difference sequence to obtain a propensity stability assessment value, and identifying an unstable window sequence based on the propensity stability assessment value; S4, reading the unstable window sequence to construct an unstable segment feature sequence, assessing the credit volatility risk level of each unstable segment, determining the credit risk level based on the assessment results, and generating a graded risk report.

[0011] Furthermore, credit behavior data is periodically collected, and the collected credit behavior data undergoes multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization processing to generate preprocessed credit behavior data. The specific steps are as follows: A fixed-width time window is defined as a sampling period, and user credit behavior data is periodically collected. Credit behavior data includes transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days. For the collected credit behavior data, the recording time from different collection links is uniformly corrected using a multi-source timestamp alignment algorithm, and an exponentially weighted moving average algorithm is used to suppress sudden jitter and short-period fluctuations in the credit behavior data. An outlier detection algorithm based on the median absolute deviation method is used to automatically detect and remove extreme offset records in the credit behavior data, and a missing segment reconstruction method based on neighborhood imputation is used to complete short-term missing segments. The credit behavior data is scaled and normalized using an interval scaling standardization algorithm.

[0012] Furthermore, the specific steps for constructing a field priority ranking table based on the preprocessed credit behavior data are as follows: Read the preprocessed credit behavior data, use a field access frequency statistics algorithm to count the cumulative number of references for each credit behavior data field within N consecutive sampling periods, and sort the cumulative number of references in descending order; in the ranking results, define the top M fields with the highest cumulative number of references as the priority processing field set, and define the remaining fields as the regular processing field set; and generate a field priority ranking table in the order of priority processing field set first, followed by regular processing field set.

[0013] Furthermore, by performing feature operations on the differences in credit behavior data within adjacent sampling periods, a credit behavior assessment value is obtained. The credit behavior assessment value and the corresponding credit behavior data are then input into the propensity score estimation model to output the propensity score. The specific steps are as follows: Based on the field priority sorting table, credit behavior data is read according to the sorted field order. Differences are calculated for transaction amount, repayment amount, transaction time interval, bill generation interval, and repayment action trigger interval for adjacent sampling periods. The absolute value of the transaction amount difference is incremented by one and the natural logarithm is taken to obtain the amount fluctuation term. The absolute value of the repayment amount difference is taken as the square root to obtain the repayment change term. The algebraic sum of the transaction time interval difference and the bill generation interval difference is used as the exponent, and an exponential function value with the natural constant e as the base is calculated. The exponential function value is incremented by one and the reciprocal is taken to obtain the behavior rhythm term. The absolute value of the repayment action trigger interval difference is incremented by one and the reciprocal is taken as the base. The reciprocal is used to obtain the repayment frequency term. The squares of the transaction amount difference and the repayment amount difference are summed, and then divided by the absolute value of the product of the transaction time interval difference, the bill generation interval difference, and the repayment action triggering interval difference, plus one, to obtain the fluctuation coupling term. The amount fluctuation term, repayment change term, behavior rhythm term, repayment frequency term, and fluctuation coupling term are added sequentially to obtain the credit behavior assessment value. After obtaining the credit behavior assessment value, a propensity score estimation model is constructed. The credit behavior data of continuous sampling periods are arranged in the field order of the field priority sorting table to form a feature vector, and the corresponding credit behavior assessment value constitutes the training sample. The support vector machine algorithm is used to perform propensity score learning on the training sample. After training, the feature vector of each sampling period is input into the propensity score estimation model, and the numerical results output by the model are processed by interval mapping to obtain the propensity score representing the tendency of credit behavior changes within the sampling period.

[0014] Further, the specific steps for generating matching period pairs based on credit behavior assessment values ​​and corresponding propensity scores to obtain the assessment difference sequence are as follows: Read credit behavior assessment values ​​for N consecutive sampling periods and construct a matching sequence together with the corresponding propensity scores; employ a one-to-one propensity score nearest neighbor matching strategy in the propensity score space to determine the closest matching period pair for each sampling period, and calculate the algebraic difference for the credit behavior assessment values ​​in the matching period pair to generate the assessment difference sequence after matching calibration; for two adjacent algebraic differences in the assessment difference sequence, calculate the absolute value of the algebraic difference between them and check whether their signs are consistent; when the sign consistency result is consistent, write the corresponding absolute value of the algebraic difference into the valid difference set; when the sign consistency result is inconsistent, discard the corresponding absolute value of the algebraic difference; arrange the absolute values ​​of each algebraic difference in the valid difference set in chronological order to generate the valid difference sequence for window aggregation.

[0015] Furthermore, the specific steps for performing sliding window aggregation on the effective difference sequence to obtain the propensity stability evaluation value are as follows: Within a sliding window of length W and sliding step size L, perform sliding window traversal on the effective difference sequence; at the starting position j of each window, expand the window range [j, j+W]... The absolute values ​​of all algebraic differences within [1] are written into the window difference set, and the arithmetic mean of all credit behavior assessment differences within the window is calculated; the absolute value of each algebraic difference in the window difference set is added by one and the reciprocal is taken to obtain the periodic volatility consistency term; the absolute value of the algebraic difference between each credit behavior assessment difference within the window and the arithmetic mean of all credit behavior assessment differences within the window is added by one and the reciprocal is taken to obtain the window trend consistency term; the volatility consistency term and the window trend consistency term within the window are added together and averaged according to the window size to obtain the current window's tendency stability assessment value.

[0016] Furthermore, the specific steps for identifying the unstable window sequence based on the tendency stability assessment value are as follows: compare the tendency stability assessment value of all windows with the stability threshold; when the stability assessment value is lower than the stability threshold, write the starting position of the window into the unstable window index set, and write the absolute value sequence of algebraic differences within the window coverage area into the unstable difference segment set; when the tendency stability assessment value is not lower than the stability threshold, write the starting position of the window into the stable window index set; after traversing all windows, arrange the unstable window index set in order of the starting position of the window to generate the unstable window sequence.

[0017] Further, the specific steps for constructing a feature sequence of unstable segments from the unstable window sequence and assessing the credit volatility risk level of each unstable segment are as follows: Read the unstable window sequence, extract the algebraic difference absolute value sequence segments corresponding to each unstable window in order of the window's starting position, calculate the maximum, minimum, and arithmetic mean of the algebraic difference absolute values ​​within the segment, and construct the unstable segment feature sequence; add one to the maximum value of the algebraic difference absolute values ​​within the unstable segment and take its reciprocal to obtain the maximum value suppression term; add one to the minimum value of the algebraic difference absolute values ​​within the unstable segment and take its square root to obtain the minimum value sensitivity term; add one to the algebraic difference absolute value between the arithmetic mean and the minimum value of the algebraic difference absolute values ​​within the unstable segment and take its reciprocal to obtain the mean shift term; calculate the ratio of the maximum value of the algebraic difference absolute values ​​to the arithmetic mean plus one, add one to the obtained ratio and take its natural logarithm to obtain the skewed structure term; add the maximum value suppression term, the minimum value sensitivity term, the mean shift term, and the skewed structure term sequentially to obtain the credit instability risk assessment value.

[0018] Furthermore, the specific steps for determining the credit risk level based on the assessment results and generating a tiered risk report are as follows: Assess the credit instability risk value of all instability segments. With multi-level risk thresholds and Compare: When ≤ When a user is determined to be at Level 1 credit risk, the corresponding credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value are extracted to generate a credit behavior steady-state characteristic report; when < < When a user is deemed to be at level two credit risk, the trend of changes in the credit instability risk assessment value is continuously monitored, and the number of consecutive login days of the user is counted. Combined with the corresponding credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value, a credit behavior deviation characteristic report is generated. ≥ When a user is identified as being at level three credit risk, a risk alarm is generated, and corresponding credit behavior data, credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value are extracted to generate a credit behavior instability characteristic report; and the multi-level risk report is sent to the superior risk management interface.

[0019] The second aspect of this invention provides a credit dynamic assessment and risk control system based on propensity score matching, comprising: a credit behavior data preprocessing module, a credit feature assessment value construction module, a propensity stability matching analysis module, and a credit instability risk response module. The credit behavior data preprocessing module periodically collects credit behavior data and performs multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization on the collected credit behavior data to generate preprocessed credit behavior data. The credit feature assessment value construction module constructs a field priority ranking table based on the preprocessed credit behavior data and combines the differences between credit behavior data in adjacent sampling periods. The system performs feature operations to obtain credit behavior assessment values. These assessment values ​​and corresponding credit behavior data are then input into a propensity score estimation model, which outputs a propensity score. A propensity stability matching analysis module generates matching period pairs based on the credit behavior assessment values ​​and corresponding propensity scores, producing an assessment difference sequence. A sliding window aggregation operation is then performed on the effective difference sequence to obtain a propensity stability assessment value. Based on this assessment value, an unstable window sequence is identified. A credit instability risk response module reads the unstable window sequence to construct an unstable segment feature sequence. The system assesses the credit volatility risk level of each unstable segment, determines the credit risk level based on the assessment results, and generates a tiered risk report.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) A credit dynamic assessment risk control method and system based on propensity score matching, by performing difference calculation, nonlinear function transformation and coupling calculation on multiple key fields such as transaction amount, repayment behavior and billing rhythm, the small changes of complex credit behavior in continuous sampling period are transformed into quantifiable credit behavior assessment values, so that the time series fluctuations present observable structured characteristics.

[0023] (2) A credit dynamic assessment risk control method and system based on propensity score matching. By training propensity scores based on support vector machines and performing nearest neighbor matching in the propensity score space, comparable period pairs with similar behavioral characteristics are formed between sampling periods, reducing the inconsistency caused by cross-period behavioral differences, thereby improving the stability and comparability of credit assessment results.

[0024] (3) Credit dynamic assessment risk control method and system based on propensity score matching: By performing sliding window aggregation on the assessment difference sequence after matching and calibration, and constructing fluctuation consistency term and trend consistency term within the window, this invention can accurately characterize the structural changes in credit behavior during consecutive weeks, so that hidden instability signals such as short-cycle anomalies and continuous slight disturbances can be identified in a timely manner.

[0025] (4) A credit dynamic assessment risk control method and system based on propensity score matching constructs a multi-dimensional risk term by performing nonlinear transformation on the maximum, minimum and arithmetic mean of the unstable segment. The final credit instability risk assessment value can reflect the comprehensive structural characteristics of fluctuation amplitude, skewness and mean deviation, so that credit risks of different levels have separable quantitative boundaries. Attached Figure Description

[0026] Figure 1 is a flowchart of the credit dynamic assessment and risk control method based on propensity score matching;

[0027] Figure 2 is a structural diagram of a credit dynamic assessment and risk control system based on propensity score matching;

[0028] Figure 3 is a bar chart showing the credit risk level determination based on the credit instability risk assessment value;

[0029] Figure 4 is a flowchart of a credit dynamic assessment and risk control system based on propensity score matching. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please refer to Figures 1-4. This invention provides a technical solution: a credit dynamic assessment and risk control method based on propensity score matching, comprising: S1, periodically collecting credit behavior data and performing multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization on the collected credit behavior data to generate preprocessed credit behavior data; S2, constructing a field priority sorting table based on the preprocessed credit behavior data, and performing feature operations based on the differences between credit behavior data within adjacent sampling periods to obtain a credit behavior assessment value, inputting the credit behavior assessment value and the corresponding credit behavior data into a propensity score estimation model to output a propensity score; S3, generating matching period pairs based on the credit behavior assessment value and the corresponding propensity score to obtain an assessment difference sequence, and performing sliding window aggregation operations on the effective difference sequence to obtain a propensity stability assessment value, identifying unstable window sequences based on the propensity stability assessment value; S4, reading the unstable window sequences to construct an unstable segment feature sequence, assessing the credit volatility risk level of each unstable segment, determining the credit risk level based on the assessment results, and generating a graded risk report.

[0032] Specifically, credit behavior data is collected periodically, and the collected credit behavior data undergoes multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization processing to generate preprocessed credit behavior data. The specific steps are as follows: A 24-hour time window is defined as a sampling period. User credit behavior data is collected periodically, including transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days. Among them, transaction amount is collected through the bank's transaction record interface, repayment amount is collected through the financial institution's bill feedback interface, account balance is obtained periodically through the account status query interface, transaction time interval is calculated from continuous transaction timestamps, bill generation interval is calculated from the timestamps in the bill generation record, repayment action trigger interval is calculated by reading the user's repayment request trigger log, and consecutive login days are calculated from the first login timestamp of the user's login record each day. During the collection process, a sampling timestamp is attached to each record to ensure the traceability of the time series. For the collected credit behavior data, a multi-source timestamp alignment algorithm is used to uniformly correct the recording times from different collection links. During the correction process, the offset of the recording time of each link is calculated based on the sampling timestamp. Records with offsets greater than the millisecond-level offset threshold are marked as records to be corrected. The recording time of the records to be corrected is gradually rolled back or moved forward according to a fixed adjustment step size. After each adjustment, the offset between the record to be corrected and the benchmark sampling timestamp is recalculated. When the offset is not greater than the millisecond-level offset threshold, the recursive time adjustment is terminated, and the recording time after termination is used as the uniformly corrected recording time to ensure that the cross-channel data can be arranged in the order of actual occurrence. An exponentially weighted moving average algorithm is used to suppress sudden jitters and short-cycle fluctuations in the credit behavior data. During smoothing, a weighted update is performed on the continuous sampling points of transaction amount, repayment amount, account balance, and transaction time interval based on a fixed attenuation factor. The value of the fixed attenuation factor is limited to 0.1 to 0.The values ​​between 9 and 9 remain constant within the same sampling period to prevent local noise from affecting subsequent difference calculations. An outlier detection algorithm based on the median absolute deviation method automatically detects and removes extreme offset records from credit behavior data. During the detection process, skewness indices and median absolute deviations are calculated for transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days, respectively. The skewness index is a skewness measure obtained by calculating the ratio of the third central moment to the second central moment of the field samples within the current sampling period, and is used to characterize the left-right skewness of the field sample distribution. The degree of asymmetry is determined by using a skewness index to determine the offset direction of field samples and by using the median absolute deviation to determine the robust offset magnitude of field samples relative to the median. Unreasonable jump entries are identified based on a combination of offset direction and offset magnitude. Furthermore, a missing data reconstruction method based on neighborhood imputation is used to complete short-term missing segments. During the completion process, feature distances are compared between multiple reliable records before and after the missing position, and the complete value is determined based on the principle of minimum distance, ensuring that the imputed content maintains a consistent trend with surrounding records. Finally, a range scaling standardization algorithm is used to perform scale normalization and numerical normalization on the credit behavior data. During the normalization process, the maximum and minimum values ​​of transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days are calculated within the continuous sampling period. This maps the field content to a unified scale interval, allowing each field to participate in subsequent feature calculations and propensity score learning under a dimensionless representation, while maintaining a consistent numerical reference structure within the unified scale system. Furthermore, when the subsequent field access frequency statistics algorithm calculates the cumulative number of references for each credit behavior data field, it calculates the maximum and minimum values ​​of transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days within each sampling period. The reading and calculation operations for the transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days are counted as one reference. When multiple readings and calculations are performed on the same field within the same sampling period, the actual number of readings and calculations is accumulated. Subsequent preference score matching introduces time constraint rules. These rules use the sampling period index difference to represent the range of the matching period difference, limiting the sampling period index difference of matching period pairs to no greater than a time constraint threshold and ensuring that matching period pairs do not cross non-continuous sampling segments. This maintains the temporal proximity of matching period pairs when performing nearest neighbor matching in the preference score space.

[0033] This implementation plan constructs a unified time benchmark, noise suppression mechanism, anomaly removal rules, missing data reconstruction strategy, and dimensionless expression method to form a stable and coherent data structure for transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days within a continuous sampling period. This improves the comparability of cross-link records at different time scales, and the features used for biased score learning have balanced data sensitivity. It effectively suppresses the interference caused by time offset, local jumps, and missing segments, and makes the overall credit behavior sequence present a structured, regular, and continuous processing characteristic, providing a robust and reliable data foundation for subsequent risk identification processes.

[0034] Specifically, the steps for constructing a field priority ranking table based on preprocessed credit behavior data are as follows: Read the preprocessed credit behavior data. During the reading process, establish independent field access counters for transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days. Use a field access frequency statistics algorithm to count the cumulative number of references for each credit behavior data field over N consecutive sampling periods, where N is the sliding observation period length used to evaluate field stability, and N ranges from 10 to 200. During the statistical process, accumulate field reference behaviors one by one according to the sampling order, so that the cumulative result can reflect the field's role in feature operations. The system calculates the actual participation level of each field and sorts them in descending order of cumulative citation count. In the sorted results, the fields with the highest cumulative citation count (ranked M) are defined as the priority processing field set, and the remaining fields are defined as the regular processing field set. The value of M ranges from one-half to two-thirds of the total number of fields. The field types and their original meanings are kept consistent during set partitioning to ensure the stability of the field order upon which subsequent difference calculations depend. A field priority sorting table is generated with the priority processing field set first, followed by the regular processing field set. Uniqueness checks are performed on the field order during generation to ensure the sorting table maintains a reproducible data reading structure throughout continuous sampling periods.

[0035] In this implementation scheme, by constructing a field access frequency structure within a continuous sampling period and establishing a field priority ranking table based on the cumulative number of references, the credit behavior data has a stable and controllable field arrangement foundation for subsequent feature operations. This ensures a consistent field participation sequence in stages such as difference calculation, behavioral feature extraction, and propensity score learning. This ranking system plays a stable role in unifying the field structure, reducing the impact of random fluctuations, and ensuring the continuity of time-series analysis, providing a clear and reliable data foundation for dynamic credit risk assessment.

[0036] Specifically, feature operations are performed by combining the differences in credit behavior data within adjacent sampling periods to obtain credit behavior assessment values. The credit behavior assessment values ​​and corresponding credit behavior data are then input into the propensity score estimation model to output the propensity score. The specific steps are as follows: Based on the field priority sorting table, credit behavior data is read according to the sorted field order. The differences in transaction amount, repayment amount, transaction time interval, bill generation interval, and repayment action trigger interval for adjacent sampling periods are calculated separately. During the difference calculation process, the field sequence and priority are kept consistent to ensure that the participation order of different fields in the subsequent feature construction process is traceable. The absolute value of the transaction amount difference is incremented by one, and the natural logarithm is taken to obtain the amount fluctuation term. The non-linear compression property of the logarithmic function is utilized when taking the natural logarithm to smoothly represent larger changes in the transaction amount difference across the value range, thus constructing a quantitative structure insensitive to large-scale amount jumps. The square root of the absolute value of the repayment amount difference is taken to obtain the repayment change term. The growth slowdown property of the power function is utilized in the square root operation to present the change in repayment amount in a continuously differentiable manner, ensuring that the difference structure has higher resolution within a small range of change. The algebraic sum of the transaction time interval difference and the bill generation interval difference is taken as the negative of the sum. The exponent is calculated by multiplying the exponential function value with the natural constant e as the base. Adding one to this exponential function value and taking its reciprocal yields the behavioral rhythm term. The exponential construction utilizes the sensitivity of the exponential function to positive and negative input values, allowing changes in transaction rhythm to exhibit asymmetric response characteristics in the exponential space, thus leading to a temporal expression that distinguishes between fast and slow-paced behaviors. Adding one to the absolute value of the difference in repayment trigger intervals and taking its reciprocal yields the repayment frequency term. In the reciprocal structure, the nonlinear decay characteristic of a 1 / x-type function is utilized, ensuring that shorter repayment trigger intervals significantly contribute to the final repayment frequency term. The difference between the transaction amount and the repayment amount are then flattened. The summation of the squares, followed by division by the absolute value of the product of the transaction time interval difference, the bill generation interval difference, and the repayment action triggering interval difference, plus one, yields the fluctuation coupling term. In the square summation structure, the amplification characteristic of the square term for large changes is utilized to enhance the influence of amount fluctuations on the coupling term. In the denominator product absolute value plus one structure, nonlinear scaling is used to strengthen the modulation effect of the behavior rhythm-related fields on the mapping relationship of amount fluctuations. The amount fluctuation term, repayment change term, behavior rhythm term, repayment frequency term, and fluctuation coupling term are added sequentially to obtain the credit behavior assessment value, so that the nonlinear expressions of different structures can form a synthesizable feature quantity within the same value domain.After obtaining the credit behavior assessment values, a propensity score estimation model is constructed. Credit behavior data from consecutive sampling periods are arranged according to the field priority sorting table to form a feature vector. During the formation process, the field order of transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days is maintained. These vectors, along with the corresponding credit behavior assessment values, constitute training samples. During the formation of training samples, the execution order of feature standardization is fixed, ensuring that standardized parameters are calculated only based on the training samples. Standardization parameters are applied only to validation and test samples in a read-only manner to prevent cross-interval transmission of feature parameters from the training phase to the validation phase. A support vector machine algorithm is used to learn propensity scores from the training samples. During the process, the numerical structure of the feature vectors is scaled at fixed intervals, with the fixed interval being limited to one to five sampling periods. Matching constraint parameters are set during the propensity score learning process, defining a caliper radius for the propensity score and limiting it to between 5% and 20% of the propensity score distance, ensuring a fixed score distance constraint for all sampling periods during the matching process. After training, the feature vectors of each sampling period are input into the propensity score estimation model, and interval mapping is performed on the model's output. During interval mapping, upper and lower bounds are set on the output to maintain a consistent numerical range for the mapping results across consecutive sampling periods, resulting in a propensity score representing the tendency of credit behavior changes within the sampling period.

[0037] The specific formula for calculating the credit behavior assessment value is as follows:

[0038] ;

[0039] In the formula, Indicates the credit behavior assessment value, This represents the difference in transaction amounts. This indicates the difference in repayment amount. Indicates the difference in transaction time intervals. This represents the difference in the interval between bill generation. This indicates the difference in the interval between repayment actions.

[0040] In this implementation scheme, distinct nonlinear transformations are introduced for the differences in transaction amounts, repayment amounts, transaction time intervals, bill generation intervals, and repayment trigger intervals. This transforms credit behavior data into multiple feature expressions with differentiated and sensitive structures, while maintaining aggregatability under a unified dimension. The response patterns established by each nonlinear function within different magnitude ranges ensure that credit behavior assessment values ​​exhibit a stable trajectory over continuous sampling periods. This provides a structurally stable and clearly discriminative feature foundation for subsequent propensity score learning, enabling the high-dimensional mapping relationships formed during support vector machine training to possess a clear behavioral distribution pattern. Consequently, an overall framework for quantitatively representing credit behavior can be constructed that continuously characterizes the temporal trends of behavioral changes.

[0041] Specifically, the steps for generating matching period pairs based on credit behavior assessment values ​​and corresponding propensity scores to obtain the assessment difference sequence are as follows: Read the credit behavior assessment values ​​of N consecutive sampling periods and construct a matching sequence together with the corresponding propensity scores. During the construction process, the credit behavior assessment values ​​and propensity scores of each sampling period are indexed and bound in chronological order to ensure that the correspondence between sampling periods remains consistent in subsequent matching processes. In the propensity score space, a one-to-one propensity score nearest neighbor matching strategy is adopted to determine the matching period pair with the closest propensity scores for each sampling period. During the matching process, L1 distance based on absolute difference and L2 distance based on squared difference are introduced as distance measurement criteria. After distance calculation, a caliper threshold is set to limit the maximum allowable range of matching distance. When performing nearest neighbor search, candidate periods exceeding the caliper threshold are excluded to control abnormal long-distance matching. Based on this, self-matching cases are excluded, and matching objects are determined sequentially. A time constraint rule is further introduced during the matching process to specify the acceptable range of matching cycle differences. This ensures that the time span during the matching cycle remains within the interpretable interval of the continuous sampling cycle structure, avoiding behavioral drift interference caused by crossing long time periods. Algebraic differences are calculated for the credit behavior assessment values ​​in the matching cycle pairs, generating a matching-calibrated assessment difference sequence. For two adjacent algebraic differences in the assessment difference sequence, their absolute algebraic differences are calculated separately. During the calculation, the sign information of the differences is recorded synchronously, and the consistency of their signs is checked. When the sign consistency result is consistent, the corresponding absolute algebraic difference is written into the valid difference set, and its original time position index is retained to maintain sequence continuity. When the sign consistency result is inconsistent, the corresponding absolute algebraic difference is discarded, and the remaining differences are subjected to sequential compression to eliminate index gaps. The absolute algebraic differences in the valid difference set are arranged in chronological order. During the arrangement process, priority coverage is performed on multiple valid differences generated in the same time period to generate a valid difference sequence for window aggregation.

[0042] In this implementation scheme, a one-to-one matching relationship is introduced between credit behavior assessment values ​​and propensity scores over consecutive sampling periods. After matching, an effective difference sequence is constructed using sign consistency constraints, significantly improving the comparability of credit behavior changes across different time segments. This process generates a difference expression with a stable scale within a unified time series, ensuring that the credit behavior assessment values, after calibration, accurately reflect the true fluctuation structure during the period. The effective difference sequence generated by matching maintains a continuous index relationship, enabling subsequent window aggregation processing to complete trend analysis without time discontinuities. This provides a structurally consistent input sequence for subsequent credit fluctuation identification and improves the overall reliability of dynamic credit assessment.

[0043] Specifically, the steps for performing sliding window aggregation on the effective difference sequence to obtain the tendency stability assessment value are as follows: Within a sliding window of length W and sliding step size L, the effective difference sequence is traversed using a sliding window. During the traversal, the window width is fixed to ensure the continuity of the periodic difference structure. The value of W is limited to three to fifteen consecutive sampling periods, and the value of L is limited to one to three, so that the window length can capture short-term internal fluctuations while maintaining the representativeness of cross-period trends. At the starting position j of each window, the window range [j, j+W] is set. The absolute values ​​of all algebraic differences within [1] are written into the window difference set. During the writing process, j is scanned incrementally with a step size L. When j+W 1. Stop window construction when the window exceeds the end of the effective difference sequence, ensuring that the window boundary always remains within the readable range of the effective difference sequence. For successfully constructed windows, calculate the arithmetic mean of all credit behavior assessment differences within the window. During the calculation, read the credit behavior assessment differences at each position in a fixed order to ensure that the mean within the window accurately represents the baseline change level for that time period. Add one to the absolute value of each algebraic difference in the window difference set and take the reciprocal to obtain the periodic volatility consistency term. The addition operation avoids unstable output caused by extremely small differences, and the reciprocal function is used to compress extreme abrupt changes, ensuring that the periodic volatility within the window is consistent. It exhibits amplitude convergence characteristics under nonlinear structures; the absolute value of the algebraic difference between each credit behavior assessment difference and the arithmetic mean of all credit behavior assessment differences within the window is increased by one and the reciprocal is taken to obtain the window trend consistency term. This term enables the local fluctuations at each position within the window to exhibit trend stability relative to the mean under nonlinear compression; the sum of all fluctuation consistency terms and the window trend consistency term within the window is averaged according to the window size to obtain the current window's tendency stability assessment value, so that the fluctuation structure within the window can form a quantitative expression of the stability of continuous sampling periods after mean-averaging.

[0044] The specific formula for calculating the propensity stability assessment value is as follows:

[0045] ;

[0046] In the formula, This indicates a preference for stability assessment values. This represents the difference in credit behavior assessment during the i-th sampling period. Indicates the window length. This represents the difference in credit behavior assessment during the (i-1)th sampling period. This represents the arithmetic mean of the differences in credit behavior assessments within the window. This represents the absolute value of the algebraic difference between the credit behavior assessment differences of two adjacent sampling periods within the window. It represents the absolute value of the algebraic difference between the credit behavior assessment difference in the i-th sampling period and the arithmetic mean of all credit behavior assessment differences within the window.

[0047] In this implementation, by introducing a fixed value range and nonlinear transformation into the window structure, the effective difference sequence can form a time-series expression with stable scaling characteristics within a local interval of continuous sampling periods. This makes the relationship between the absolute value of the algebraic difference within the window and the benchmark mean comparable under a uniform scale. This mechanism enables the propensity stability assessment value to exhibit consistent temporal sensitivity in the context of multi-period fluctuations, allowing stability changes across windows to be presented in a continuous manner, and providing a range-representative stability quantification basis for the subsequent identification of unstable window sequences.

[0048] Specifically, the steps for identifying unstable window sequences based on propensity stability assessment values ​​are as follows: The propensity stability assessment values ​​of all windows are compared with the stability threshold. During the comparison, the propensity stability assessment values ​​of each window are read sequentially according to the window generation order, ensuring consistency in the assessment process within the time series structure of continuous sampling periods. When the stability assessment value is lower than the stability threshold, the window's starting position is written into the unstable window index set. Before writing, the window's starting position is timestamped to ensure the index accurately locates the corresponding algebraic difference absolute value sequence range. The algebraic difference absolute value sequence fragments within the window's coverage area are written into the unstable difference fragment set. During the writing process, the original arrangement order of the algebraic difference absolute values ​​is maintained, ensuring the fragment content fully reflects the continuous change pattern of the unstable stage. When the propensity stability assessment value is not lower than the stability threshold, the window's starting position is written into the stable window index set. During the process, the index increment order is kept consistent with the window sliding relationship to distinguish continuous window segments in a stable state. After traversing all windows, the unstable window index set is arranged in order of the window's starting position. During the arrangement process, consistency is checked according to the time sequence to ensure that the ordered indexes can form a complete temporal structure and generate an unstable window sequence. After generating the unstable window sequence, an interval detection is performed on the starting positions of adjacent unstable windows in the sequence. When the difference between adjacent starting positions is not greater than the interval threshold, the corresponding windows are regarded as belonging to the same continuous unstable segment. The starting index and ending index of the segment are merged according to the continuity rule. During the merging process, the starting position of the earliest window in the continuous unstable window sequence is used as the segment starting index, and the ending position of the latest covered area in the continuous unstable window sequence is used as the segment ending index, forming an unstable segment start and end index record for subsequent risk calculation.

[0049] In this implementation, a temporal identification mechanism based on propensity stability assessment values ​​is introduced during window traversal. This enables the unstable window index set to form a stable positioning pattern within the temporal structure of continuous sampling periods, thereby ensuring that the unstable difference segment set has a clear temporal correlation when recording algebraic difference absolute value sequence segments. This mechanism maintains the continuous expression of the window's starting position when identifying unstable window sequences, making the final unstable window sequence traceable and interpretable, and providing a clear temporal boundary foundation for subsequent unstable segment feature construction and risk identification.

[0050] Specifically, the steps for constructing a feature sequence of unstable segments from the unstable window sequence and assessing the credit volatility risk level of each unstable segment are as follows: Read the unstable window sequence, maintaining consistency in the temporal order of the window's starting position during the reading process, so that each unstable segment can be arranged according to the actual volatility order; extract the algebraic difference absolute value sequence segment corresponding to each unstable window sequentially according to the window's starting position order, fixing the recording order of the sequence segments during extraction, so that the algebraic difference absolute value within the segment can maintain a temporal structure consistent with the original sampling period; calculate the algebraic difference within the segment. The maximum, minimum, and arithmetic mean of the absolute values ​​are calculated as follows: The maximum value is obtained by sequentially scanning the absolute values ​​of the algebraic differences at each position within the segment to obtain a benchmark value representing the maximum fluctuation amplitude within the segment; the minimum value is determined based on rigorous point-by-point comparison to determine the lowest fluctuation intensity within the segment; and the arithmetic mean is calculated by summing all the absolute values ​​of the algebraic differences based on the segment length to form a central value characterizing the overall fluctuation level of the segment, thus constructing a characteristic sequence of unstable segments. The maximum absolute value of the algebraic differences within the unstable segments is incremented by one, and the reciprocal is taken to obtain a maximum value suppression term. This incrementing operation is used to eliminate the tendency of the denominator to extreme values ​​caused by extreme maximum values. Zero-risk, the reciprocal transformation is used to compress abnormally high amplitude fluctuations to enhance the comparability between segments; the minimum absolute value of the algebraic difference within unstable segments is incremented by one and the square root is taken to obtain the minimum-sensitive term. The increment operation is used to maintain the numerical stability of the minimum value under the square root operation, and the square root transformation is used to enhance the ability to identify small fluctuations in the segment; the absolute value of the algebraic difference between the arithmetic mean and the minimum value of the algebraic difference within unstable segments is incremented by one and the reciprocal is taken to obtain the mean offset term. The increment operation is used to mitigate the risk of numerical amplification caused by the small difference between the mean and the minimum value, and the reciprocal transformation is used to measure the average fluctuation within the segment. The degree of deviation relative to the lowest fluctuation; calculate the ratio of the maximum absolute value of the algebraic difference to the arithmetic mean plus one, add one to the obtained ratio and take the natural logarithm to obtain the skewed structure term. The addition operation is used to stabilize the output of the ratio under extreme conditions. The natural logarithmic transformation is used to map the structural difference between the maximum fluctuation and the average fluctuation within a segment to a continuous and controllable range of change; add the maximum value suppression term, minimum value sensitivity term, mean deviation term and skewed structure term in sequence to obtain the credit instability risk assessment value. Maintain a fixed order in the summation process to ensure that the contribution structure of each indicator to the credit instability risk assessment value is consistent.

[0051] The specific formula for calculating the credit instability risk assessment value is as follows:

[0052] ;

[0053] In the formula, This represents the credit instability risk assessment value of the k-th unstable segment. This represents the maximum absolute value of the algebraic difference. This represents the minimum absolute value of the algebraic difference. It represents the arithmetic mean of the absolute values ​​of algebraic differences.

[0054] In this embodiment, Table 1 is a data table of credit instability risk assessment values, listing all characteristic parameters used to assess the intensity of credit behavior fluctuations in the five instability segments during this risk quantification process. Specifically: Instability segment 1: The maximum absolute value of the algebraic difference is 4.2, the minimum absolute value is 0.8, the arithmetic mean of the absolute values ​​is 3.5, and the credit instability risk assessment value is approximately 2.69; Instability segment 2: The maximum absolute value of the algebraic difference is 4.5, the minimum absolute value is 1.3, the arithmetic mean is 2.2, and the corresponding credit instability risk assessment value is approximately 3.10; Instability segment 3: The maximum absolute value of the algebraic difference is 5.1, the minimum absolute value is 0. 0.6, the arithmetic mean of the absolute algebraic difference is 3.4, and the credit instability risk assessment value is approximately 2.46; Instability segment 4: the maximum absolute algebraic difference is 2.8, the minimum absolute algebraic difference is 0.9, the arithmetic mean is 1.9, and the calculated credit instability risk assessment value is approximately 2.82; Instability segment 5: the maximum absolute algebraic difference is 3.7, the minimum absolute algebraic difference is 0.3, the arithmetic mean of the absolute algebraic difference is 3.0, and the corresponding credit instability risk assessment value is approximately 2.28.

[0055] Table 1. Data Table of Credit Instability Risk Assessment Values

[0056]

[0057] Figure 3 shows the credit instability risk assessment values ​​calculated based on five instability segments, along with the corresponding risk level determinations. The bar chart uses different colors to distinguish the risk levels of each instability segment: green bars represent Level 1 credit risk; orange bars represent Level 2 credit risk; and red bars represent Level 3 credit risk. Gray and black dashed lines in the figure mark the Level 1 and Level 2 risk thresholds, respectively, used to delineate the risk assessment values ​​for different instability segments. As can be seen from the figure, the credit instability risk assessment value of instability segment 2 exceeds the Level 2 risk threshold and is classified as Level 3 credit risk, indicating that the absolute value of the algebraic difference within this segment exhibits a significant fluctuation structure. The credit instability risk assessment values ​​of instability segments 1 and 4 are between the Level 1 and Level 2 risk thresholds and are marked as Level 2 credit risk, suggesting that their volatility is more active than that of regular segments. The credit instability risk assessment values ​​of instability segments 3 and 5 are below the Level 1 risk threshold, indicating that their credit volatility is relatively stable. Figure 3 visually illustrates the credit risk classification and identification mechanism based on the credit instability risk assessment value and multi-level risk thresholds. By jointly quantifying the absolute value of the algebraic difference of the instability segment in the dimensions of maximum, minimum and mean, it realizes the risk stratification judgment of continuous credit behavior fluctuations and provides visual support for dynamic credit risk control.

[0058] In this implementation scheme, a feature sequence of unstable segments with continuity, comparability, and structural integrity is constructed based on the instability window index. Combined with a multi-dimensional feature structure formed by the maximum, minimum, and arithmetic mean of the absolute values ​​of algebraic differences, the internal fluctuation patterns of segments can be uniformly characterized from three directions: local peaks, bottom fluctuations, and central levels. Through the combined construction of maximum suppression terms, minimum sensitivity terms, mean offset terms, and skewed structure terms, the differences in amplitude, fluctuation, central offset, and structural tilt of unstable segments can be stably quantified within the same expression system. This ensures that the credit instability risk assessment value maintains a consistent evaluation scale when facing different segment lengths, fluctuation patterns, and spans. This feature construction method provides a structured expression of the overall risk characteristics of unstable segments, offering a continuous, stable, and identifiable quantitative basis for subsequent credit volatility risk level classification.

[0059] Specifically, the steps for determining credit risk levels based on the assessment results and generating a tiered risk report are as follows: Assess the credit instability risk value of all instability segments. With multi-level risk thresholds and A comparison is performed, following a fixed order of numerical reading, threshold comparison, and level confirmation. The stability of the boundary determination is ensured through two consecutive numerical checks in the threshold boundary region. ≤ When a user is determined to be at Level 1 credit risk, the corresponding content is written in the following order: first, extracting the credit behavior assessment value; then, extracting the tendency stability assessment value; and finally, extracting the credit instability risk assessment value. The extracted content is then used to generate a steady-state credit behavior characteristic report, arranged with risk parameter fields first and behavior fields last, ensuring a consistent field structure in the steady-state report. < < When a user is determined to be at level two credit risk, the monitoring process is executed in the following order: first, reading the credit instability risk assessment value; then, forming a risk change sequence; and finally, calculating the user's consecutive login days. The risk change sequence and the consecutive login days are combined in the order of trend fields first, followed by behavior fields, and then combined with the corresponding credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value to generate a credit behavior deviation feature report. During the generation process, the time-series expressive power of the report is enhanced by fixing the field order. ≥ When a user is determined to be at level three credit risk, the alarm content is constructed in the following order: first, risk alarm information is generated; then, the alarm trigger time is recorded; and finally, risk parameter fields are completed. Corresponding credit behavior data, credit behavior assessment values, tendency stability assessment values, and credit instability risk assessment values ​​are extracted. These are then arranged in the order of behavior data fields first, assessment value fields second, and risk value fields last to form a credit behavior instability feature report. After generation, a field order verification is performed to ensure consistency of the output content across multiple records. The multi-level risk report is then sent to the higher-level risk management interface. During the transmission process, field integrity verification, format checking, and data transmission are completed sequentially to ensure the stability and traceability of the risk transmission process.

[0060] In this implementation plan, credit instability risk assessment value is used... Multi-level risk thresholds and The comparison process incorporates a fixed-order level confirmation mechanism, enabling credit behavior assessment values, propensity stability assessment values, and credit instability risk assessment values ​​to participate in risk level determination under a unified structure. This ensures that the generation process of tiered risk reports possesses a stable hierarchical expression capability. This structured risk output method enhances the ability to continuously identify risk levels over time, provides clearer boundary representations between different credit behavior states, and forms a more interpretable risk expression framework at the risk management interface, thereby improving the applicability of dynamic credit risk identification in cross-cycle scenarios.

[0061] As shown in Figure 2, the second aspect of this invention provides a credit dynamic assessment and risk control system based on propensity score matching, comprising: a credit behavior data preprocessing module, a credit feature assessment value construction module, a propensity stability matching analysis module, and a credit instability risk response module. The credit behavior data preprocessing module is used to periodically collect credit behavior data and perform multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization on the collected credit behavior data to generate preprocessed credit behavior data. The credit feature assessment value construction module is used to construct a field priority sorting table based on the preprocessed credit behavior data and combine it with credit behavior data from adjacent sampling periods. The system performs feature operations on the differences to obtain credit behavior assessment values. These assessment values ​​and corresponding credit behavior data are then input into a propensity score estimation model, which outputs a propensity score. A propensity stability matching analysis module generates matching period pairs based on the credit behavior assessment values ​​and corresponding propensity scores, resulting in an assessment difference sequence. A sliding window aggregation operation is then performed on the effective difference sequence to obtain a propensity stability assessment value. Based on this value, an unstable window sequence is identified. A credit instability risk response module reads the unstable window sequence to construct an unstable segment feature sequence, assesses the credit volatility risk level of each unstable segment, determines the credit risk level based on the assessment results, and generates a tiered risk report. The overall dynamic credit assessment process of the system is shown in Figure 4.

[0062] This implementation scheme establishes a continuous information transmission relationship between credit behavior data processing, credit behavior assessment value construction, propensity score generation, assessment difference sequence matching, propensity stability assessment value identification, and credit risk level determination. This enables credit behavior data, credit behavior assessment values, propensity scores, assessment difference sequences, propensity stability assessment values, and credit instability risk assessment values ​​to form a hierarchical and progressive correlation structure within a unified calculation framework. This correlation structure allows for stable characterization of credit behavior changes at a fine-grained scale, enables cross-cycle credit fluctuations to have a clear trend expression capability through multi-dimensional feature fusion, maintains interpretable temporal boundaries for instability window sequences in continuous identification, and achieves consistent output of credit risk levels based on quantitative information from the complete link. This results in higher structural stability and judgment accuracy in dynamic credit risk identification.

[0063] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A credit dynamic assessment and risk control method based on propensity score matching, characterized in that, The process includes the following steps: S1, periodically collect credit behavior data, and perform multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation and normalization on the collected credit behavior data to generate preprocessed credit behavior data; S2, construct a field priority sorting table based on the preprocessed credit behavior data, and perform feature operations by combining the differences of credit behavior data in adjacent sampling periods to obtain credit behavior evaluation values. Input the credit behavior evaluation values ​​and corresponding credit behavior data into the propensity score estimation model to output the propensity score. The specific steps for generating matching period pairs based on credit behavior evaluation values ​​and corresponding propensity scores to obtain the evaluation difference sequence are as follows: Read the credit behavior evaluation values ​​of N consecutive sampling periods and construct a matching sequence together with the corresponding propensity scores; adopt a one-to-one propensity score nearest neighbor matching strategy in the propensity score space to determine the matching period pair with the closest propensity score for each sampling period, and calculate the algebraic difference of the credit behavior evaluation values ​​in the matching period pair to generate the evaluation difference sequence after matching calibration; for two algebraic differences at adjacent positions in the evaluation difference sequence, calculate the absolute value of the algebraic difference between them respectively, and check whether the signs of the two are consistent. When the sign consistency result is consistent, the corresponding absolute algebraic difference value is written into the effective difference set; when the sign consistency result is inconsistent, the corresponding absolute algebraic difference value is discarded; the absolute algebraic difference values ​​are arranged in chronological order in the effective difference set to generate an effective difference sequence for window aggregation; S3, based on the credit behavior assessment value and the corresponding propensity score, a matching period pair is generated to obtain the assessment difference sequence, and a sliding window aggregation operation is performed on the effective difference sequence to obtain the propensity stability assessment value, and the unstable window sequence is identified based on the propensity stability assessment value; the specific steps of performing the sliding window aggregation operation on the effective difference sequence to obtain the propensity stability assessment value are as follows: within a sliding window with a window length of W and a sliding step size of L, a sliding window traversal is performed on the effective difference sequence; at the starting position j of each window, all absolute algebraic difference values ​​within the window range [j, j+W-1] are written into the window difference set, and the arithmetic mean of all credit behavior assessment differences within the window is calculated; Add one to the absolute value of each algebraic difference in the window difference set and take the reciprocal to obtain the periodic fluctuation consistency term. The window trend consistency term is obtained by adding one to the absolute value of the algebraic difference between each credit behavior assessment difference in the window and the arithmetic mean of all credit behavior assessment differences in the window. The average value of the current window's tendency stability is obtained by summing all the fluctuation consistency items and window trend consistency items within the window and then averaging them according to the window size. S4, read the instability window sequence to construct the instability segment feature sequence, assess the credit volatility risk level of each instability segment, determine the credit risk level based on the assessment results, and generate a graded risk report; the specific steps of reading the instability window sequence to construct the instability segment feature sequence and assessing the credit volatility risk level of each instability segment are as follows: read the instability window sequence, extract the algebraic difference absolute value sequence segment corresponding to each instability window in sequence according to the window start position, calculate the maximum value, minimum value, and arithmetic mean of the algebraic difference absolute value within the segment, and construct the instability segment feature sequence; add one to the maximum value of the algebraic difference absolute value within the instability segment and take the reciprocal to obtain the maximum value suppression term; add one to the minimum value of the algebraic difference absolute value within the instability segment and take the square root to obtain the minimum value sensitivity term; The mean offset term is obtained by adding one to the absolute value of the algebraic difference between the absolute value of the algebraic difference within the unstable segment and the minimum value, and then taking the reciprocal. Calculate the ratio of the maximum absolute value of the algebraic difference to the arithmetic mean plus one, add one to the obtained ratio and take the natural logarithm to obtain the skewed structure term; add the maximum value suppression term, minimum value sensitivity term, mean offset term and skewed structure term in sequence to obtain the credit instability risk assessment value.

2. The credit dynamic assessment and risk control method based on propensity score matching according to claim 1, characterized in that: The specific steps for periodically collecting credit behavior data and performing multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization on the collected credit behavior data to generate preprocessed credit behavior data are as follows: A fixed-width time window is defined as one sampling period. Credit behavior data is periodically collected, including transaction amount, repayment amount, account balance, transaction time interval, bill generation interval, repayment action trigger interval, and consecutive login days. For the collected credit behavior data, the recording time from different collection links is uniformly corrected using a multi-source timestamp alignment algorithm, and an exponentially weighted moving average algorithm is used to suppress sudden jitter and short-period fluctuations in the credit behavior data. An outlier detection algorithm based on the median absolute deviation method is used to automatically detect and remove extreme offset records in the credit behavior data, and a missing segment reconstruction method based on neighborhood imputation is used to complete short-term missing segments. Finally, the credit behavior data is subjected to scale normalization and numerical normalization processing using an interval scaling normalization algorithm.

3. The credit dynamic assessment and risk control method based on propensity score matching according to claim 1, characterized in that: The specific steps for constructing a field priority sorting table based on preprocessed credit behavior data are as follows: Read the preprocessed credit behavior data, use a field access frequency statistics algorithm to count the cumulative number of references for each credit behavior data field within N consecutive sampling periods, and sort the cumulative number of references in descending order; in the sorting results, define the top M fields with the highest cumulative number of references as the priority processing field set, and define the remaining fields as the regular processing field set; and generate a field priority sorting table in the order of priority processing field set first, followed by regular processing field set.

4. The credit dynamic assessment and risk control method based on propensity score matching according to claim 1, characterized in that: The specific steps for performing feature operations on the differences between credit behavior data within adjacent sampling periods to obtain a credit behavior assessment value, and inputting the credit behavior assessment value and corresponding credit behavior data into the propensity score estimation model to output the propensity score are as follows: Based on the field priority sorting table, read the credit behavior data according to the sorted field order, and calculate the differences for transaction amount, repayment amount, transaction time interval, bill generation interval, and repayment action trigger interval for adjacent sampling periods; add one to the absolute value of the transaction amount difference and take the natural logarithm to obtain the amount fluctuation term; take the square root of the absolute value of the repayment amount difference to obtain the repayment change term; and then... The algebraic sum of the time interval difference and the bill generation interval difference, with the inverse of the sum, is used as the exponent. An exponential function value with the natural constant e as the base is calculated. The exponential function value is then incremented by one and its reciprocal is taken to obtain the behavioral rhythm term. The absolute value of the repayment action trigger interval difference is incremented by one and its reciprocal is taken to obtain the repayment frequency term. The squares of the transaction amount difference and the repayment amount difference are summed, and then divided by the absolute value of the product of the transaction time interval difference, the bill generation interval difference, and the repayment action trigger interval difference, plus one, to obtain the fluctuation coupling term. The amount fluctuation term, repayment change term, behavioral rhythm term, repayment frequency term, and fluctuation coupling term are added sequentially to obtain the credit behavior assessment value. After obtaining the credit behavior assessment value, a propensity score estimation model is constructed. The credit behavior data of continuous sampling periods are arranged in the field order of the field priority sorting table to form a feature vector, and the feature vector is formed with the corresponding credit behavior assessment value to constitute the training sample. The support vector machine algorithm is used to perform propensity score learning on the training sample. After training, the feature vector of each sampling period is input into the propensity score estimation model, and the numerical results output by the model are subjected to interval mapping processing to obtain the propensity score representing the tendency of credit behavior changes within the sampling period.

5. The credit dynamic assessment and risk control method based on propensity score matching according to claim 1, characterized in that: The specific steps for identifying the unstable window sequence based on the tendency stability assessment value are as follows: Compare the tendency stability assessment value of all windows with the stability threshold: When the stability assessment value is lower than the stability threshold, write the starting position of the window into the unstable window index set, and write the absolute value sequence of algebraic differences within the window coverage area into the unstable difference segment set; When the tendency stability assessment value is not lower than the stability threshold, write the starting position of the window into the stable window index set; After traversing all windows, arrange the unstable window index set in order of the starting position of the window to generate the unstable window sequence.

6. The credit dynamic assessment and risk control method based on propensity score matching according to claim 1, characterized in that: The specific steps for determining the credit risk level based on the assessment results and generating a graded risk report are as follows: Assess the credit instability risk value of all instability segments... With multi-level risk thresholds and Compare: When ≤ When a user is determined to be at Level 1 credit risk, the corresponding credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value are extracted to generate a credit behavior steady-state characteristic report. when < < When a user is identified as being at level two credit risk, the system continuously monitors the trend of changes in the credit instability risk assessment value and counts the number of consecutive days the user logs in. Based on the corresponding credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value, a credit behavior deviation feature report is generated. when ≥ When a user is determined to be at level three credit risk, a risk alarm is generated, and the corresponding credit behavior data, credit behavior assessment value, tendency stability assessment value, and credit instability risk assessment value are extracted to generate a credit behavior instability characteristic report. And send multi-level risk reports to the higher-level risk management interface.

7. A credit dynamic assessment and risk control system based on propensity score matching, employing the credit dynamic assessment and risk control method based on propensity score matching as described in any one of claims 1-6, characterized in that, include: The system comprises a credit behavior data preprocessing module, a credit feature evaluation value construction module, a propensity stability matching analysis module, and a credit instability risk response module. Specifically: the credit behavior data preprocessing module periodically collects credit behavior data and performs multi-source timestamp alignment, exponential sliding smoothing, robust outlier removal, neighborhood imputation, and normalization on the collected data to generate preprocessed credit behavior data; the credit feature evaluation value construction module constructs a field priority sorting table based on the preprocessed credit behavior data and performs feature calculations based on the differences between credit behavior data in adjacent sampling periods to obtain the credit behavior evaluation value. The credit behavior assessment value and corresponding credit behavior data are input into the propensity score estimation model, which outputs the propensity score. The propensity stability matching analysis module is used to generate matching period pairs based on the credit behavior assessment value and the corresponding propensity score, obtain an assessment difference sequence, and perform a sliding window aggregation operation on the effective difference sequence to obtain a propensity stability assessment value. Based on the propensity stability assessment value, the unstable window sequence is identified. The credit instability risk response module is used to read the unstable window sequence to construct an unstable segment feature sequence, assess the credit volatility risk level of each unstable segment, perform credit risk level determination based on the assessment results, and generate a graded risk report.

Images