Repayment ability dynamic prediction and overdue early warning method based on multi-source time series data

Abstract

This invention relates to the field of financial risk early warning technology, specifically a method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data. The method includes: acquiring multi-source time-series data of a target object over a continuous time period; generating a clean time-series data stream through time axis normalization and data quality restoration; constructing a feature set of income stability, expenditure volatility, and debt pressure based on this data stream, forming a multi-dimensional risk feature space; inputting this data into a probabilistic graphical model based on state transitions, outputting a probability distribution of repayment ability states; and calling an inversion inference engine to trace the key characteristic variables and change trajectories of abnormal states through state inversion technology, comparing them with preset risk thresholds to determine the risk critical point and level. This method achieves dynamic prediction of repayment ability and overdue early warning, improving the accuracy and targeting of risk identification.

Description

Technical Field

[0001] This invention relates to the field of financial risk early warning technology, and in particular to a method for dynamic prediction of repayment ability and overdue early warning based on multi-source time series data. Background Technology

[0002] In the field of financial credit risk prevention and control, dynamic prediction of repayment ability and overdue warning are important technical means to ensure the safety of credit assets. Currently, existing technologies mostly acquire multi-source time series data of the target object, and after simple data cleaning and normalization, construct basic income, expenditure and debt-related characteristics. Then, traditional regression models, classification models and other algorithms are used to output a single repayment ability score or overdue probability, thereby achieving risk warning.

[0003] Existing technical solutions have shortcomings. After data processing, they can only output a single-dimensional evaluation result, failing to comprehensively reflect the dynamic changes in the repayment ability of the target entity and making it difficult to capture subtle fluctuations in repayment ability at different time points. Furthermore, for identified abnormal repayment ability states, they can only provide anomaly alerts, without identifying the core cause of the anomaly or tracing the key characteristic variables leading to the anomaly and their changes. This makes it difficult to take targeted preventative measures after an alert, failing to meet the refined and dynamic risk control needs of financial scenarios. Even when a multi-dimensional feature space is constructed, existing technologies do not incorporate state transition logic to achieve a probabilistic presentation of repayment ability states and lack an effective mechanism for tracing the root causes of anomalies, making it difficult to solve the aforementioned problems of insufficient accuracy and lack of source tracing capabilities. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data, comprising: Acquire multi-source time-series data of the target object within a continuous time period, perform time axis normalization and data quality repair operations on the multi-source time-series data, and generate a set of clean time-series data streams with unified timestamps; Based on the clean time-series data stream, an income stability feature set, an expenditure volatility feature set, and a debt pressure feature set are constructed respectively. The income stability feature set, the expenditure volatility feature set, and the debt pressure feature set together constitute a multidimensional risk feature space. The multidimensional risk feature space is input into the state transition-based probabilistic graphical model, which, through inference calculation, outputs the probability distribution of the target object's repayment ability status at the current time point. The inversion and inference engine is invoked to perform root cause analysis on the abnormal states that are in the low probability range of the repayment ability state probability distribution. Through state inversion technology, the key feature variables that caused the abnormal state and their change trajectory are traced in reverse. The key feature variables and their change trajectories obtained from the inversion are compared with the pre-set risk thresholds to determine the critical point for risk triggering and the risk level.

[0006] As a further aspect of the present invention, time axis normalization and data quality repair operations are performed on the multi-source time-series data to generate a set of clean time-series data streams with unified timestamps, including: The multi-source time-series data includes at least bank transaction data, consumer behavior data, credit history data, and occupational change data; The transaction records in the bank statement data are summarized by three time granularities: daily, weekly, and monthly. The net inflow and net outflow of funds at each time granularity are calculated to form the time series sub-data of fund flow. The transaction location and transaction type in the aforementioned consumer behavior data are desensitized and encoded. Combined with the transaction amount, the proportion of non-essential consumption per unit time is calculated to form a time-series sub-data of the consumption structure. The entry date, exit date, and job salary information in the occupational change data are analyzed to construct a time-series state sequence reflecting occupational stability, forming occupational state time-series sub-data. The repayment and query records in the credit history data are statistically analyzed using a time window sliding window to calculate the recent overdue frequency and hard query count, forming a time series sub-data of credit performance. The generated time-series sub-data of fund flow, consumption structure, occupation status, and credit performance are interpolated using time interpolation to fill in missing time points, achieving strict alignment of the entire timeline and forming the clean time-series data stream.

[0007] As a further aspect of the present invention, based on the clean time-series data stream, a set of characteristics for income stability, a set of characteristics for expenditure volatility, and a set of characteristics for debt pressure are constructed, including: For the time series sub-data of the cash flow in the clean time series data stream, calculate the slope and variance of its moving average line, extract indicators that reflect the income growth trend and fluctuation range, and form the income stability feature set; For the consumption structure time-series sub-data in the clean time-series data stream, calculate its Shannon entropy and kurtosis, extract indicators reflecting the dispersion of consumption behavior and the concentration of abnormal consumption, and form the expenditure volatility feature set; The cash flow time series sub-data and the credit performance time series sub-data in the clean time series data stream are calculated using two dimensions: debt-to-income ratio and remaining credit line utilization rate. Indicators reflecting short-term debt repayment pressure are extracted and formed into the debt pressure feature set. The constructed income stability feature set, expenditure volatility feature set, and debt pressure feature set are concatenated to form the multidimensional risk feature space, wherein the feature vector at each time point is composed of the corresponding elements of these three feature sets.

[0008] As a further aspect of the present invention, the multidimensional risk feature space is input into a state transition-based probabilistic graphical model. The state transition-based probabilistic graphical model, through inference calculation, outputs the probability distribution of the target object's repayment ability status at the current time point, including: In the probabilistic graphical model based on state transition, several mutually exclusive repayment ability state nodes are defined, and the repayment ability state nodes include at least a robust state, a stressed state, and a deteriorating state. Initialize a state transition matrix, where each element of the state transition matrix represents the conditional probability of transitioning from one repayment ability state to another, and the conditional probability is determined by the feature differences between adjacent time points in the multidimensional risk feature space. Based on the current time series data points, extract their feature vectors and update the values ​​in the state transition matrix using the maximum likelihood estimation method; The forward algorithm is run to traverse all time-series data points in the multidimensional risk feature space, and the posterior probability of the target object being in each repayment ability state at the current time point is calculated cumulatively, and the summaries form the repayment ability state probability distribution.

[0009] As a further aspect of the present invention, the step of invoking the inversion inference engine to perform root cause analysis on the abnormal states in the low probability interval of the repayment ability state probability distribution, and through state inversion technology, tracing back the key characteristic variables and their change trajectories that led to the abnormal states, including: In the probability distribution of repayment ability status, repayment ability statuses with a probability value lower than a preset warning line are selected as abnormal statuses to be analyzed. Activate the inversion and inference engine, set the abnormal state as the termination node, and backtrack along the reverse link of the probabilistic graphical model based on state transition; During the backtracking process, the contribution of each predecessor feature node to the abnormal state is calculated. The contribution is obtained by calculating the partial derivative of the change in the feature variable with the change in the state transition probability. Identify feature variables whose absolute contribution value is greater than a preset threshold and mark them as key feature variables; Along the positive time axis, extract the value sequence of the key feature variable within the historical time period and plot the change trajectory of the key feature variable.

[0010] As a further aspect of the present invention, the step of comparing the key feature variables and their change trajectories obtained from the inversion with a pre-set risk threshold to determine the critical point for risk triggering and the risk level includes: Retrieve a pre-configured single-dimensional risk threshold table for each key feature variable, wherein the risk threshold table defines normal range, attention range and danger range; Map each data point in the change trajectory of the key feature variables to the corresponding interval in the risk threshold table; Find the point in time when the change trajectory of the key characteristic variable first falls into the danger zone, and mark that point as the risk triggering critical point; The number of key characteristic variables falling into the danger zone within the time window approaching the risk trigger threshold is counted, and the risk level is divided according to the number of variables: one is considered low risk, two are considered medium risk, and three or more are considered high risk.

[0011] As a further aspect of the present invention, it also includes: Based on the determined risk level, activate the corresponding early warning signal and generate a risk diagnosis report containing the risk triggers and key characteristic fluctuations. Sensitivity analysis was performed on the key feature variables in the risk diagnosis briefing to quantify the impact coefficient of each key feature variable on the transition of repayment ability status. Using the influence coefficient, the parameters of the probabilistic graphical model based on state transition are fine-tuned to correct the model's bias in predicting the state at future time points. The state transition-based probabilistic graphical model, after parameter fine-tuning, is applied to the data of the next time window to begin a new cycle of dynamic prediction of repayment ability and overdue warning. Based on the determined risk level, the corresponding early warning signal is activated, and a risk diagnosis report containing the risk triggers and key characteristic fluctuations is generated, specifically including: Based on the determined risk level, a corresponding early warning signal template is matched from the early warning strategy library. The early warning signal template specifies the color code and notification frequency for the early warning. Fill in the placeholders in the warning signal template to generate specific warning instructions; Extract the original data snapshots of the key feature variables within two time windows before and after the risk triggering critical point from the multidimensional risk feature space; The original data snapshots of the key feature variables and their corresponding change trajectory charts are integrated together with the textual descriptions of the risk triggers to form the risk diagnosis brief.

[0012] As a further aspect of the present invention, sensitivity analysis is performed on the key characteristic variables in the risk diagnosis report to quantify the impact coefficient of each key characteristic variable on the transition of repayment ability status, including: In the state transition-based probabilistic graphical model, all other feature variables except for one key feature variable are frozen. By artificially adding a small perturbation to the value of one of the key characteristic variables, the magnitude of the change in the probability distribution of the repayment ability status is observed. Calculate the ratio of the variation amplitude of the probability distribution of the repayment ability status to the amount of small disturbance, and define the ratio as the univariate influence coefficient of a certain key characteristic variable; By iterating through all the key feature variables and repeating the perturbation and calculation process, the influence coefficient of each key feature variable is obtained.

[0013] As a further aspect of the present invention, the influence coefficient is used to fine-tune the parameters of the probabilistic graphical model based on state transition, in order to correct the model's deviation in predicting the state at future time points, including: The absolute values ​​of the influence coefficients are sorted, and the top-ranked key feature variables are selected as the dominant variables. Based on the sign of the influence coefficient of the dominant variable, adjust the bias parameter of the state transition matrix in the probabilistic graphical model based on state transition; The gradient descent method is used to iteratively optimize the adjusted bias parameter and other relevant parameters with the goal of minimizing the prediction error of the model on historical data. When the model's prediction error converges to an acceptable range, the iteration stops, and the model parameters at this point are saved as the fine-tuned parameter configuration.

[0014] As a further aspect of the present invention, the construction steps of the inversion inference engine include: Create a mirror model in memory that is exactly the same as the state transition-based probabilistic graphical model structure, the mirror model containing the same state nodes and directed edges for state transitions; In the mirror model, the transition direction of each directed edge in the state transition matrix is ​​physically reversed to construct a reverse link, so that in the reverse link, it is possible to trace back to all the predecessor state nodes from the current state node. A weighted backtracking algorithm is defined for each reversed directed edge in the reverse link. The weighted backtracking algorithm takes the posterior probability calculated by forward inference as input and calculates the probability contribution value of the reverse transition according to Bayes' theorem. At each preceding feature node of the reverse link, a contribution calculator is deployed. The contribution calculator is used to execute the weight backtracking algorithm and output the contribution of each preceding feature node to the target abnormal state. The completed computational framework, which includes the reverse link, the weight backtracking algorithm, and the contribution calculator, is encapsulated as the inversion inference engine. When an abnormal state to be analyzed is input, the reverse tracing process can be automatically started.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By inputting a multidimensional risk feature space into a state transition-based probabilistic graphical model, the model outputs the probability distribution of the repayment ability status of the target object at the current point in time through inference calculation. Compared with conventional technologies that output a single repayment ability score, delinquency probability value, or binary classification result, this model can comprehensively present different states of repayment ability and the probability proportion of each state, clearly reflecting the dynamic fluctuation trend and potential change tendency of repayment ability. Combined with the temporal correlation of multi-source time series data, the probability distribution can reflect the correlation changes of repayment ability at different time points, breaking the limitations of conventional technologies in judging repayment ability from a single dimension, achieving a refined characterization of repayment ability status, and avoiding risk misjudgment or omission due to a single result. At the same time, relying on the state transition logic, the judgment of repayment ability is more in line with its actual situation of dynamic changes over time.

[0016] By invoking the inversion and inference engine, root cause analysis is performed on abnormal states in the low-probability range of the repayment ability status probability distribution. The state inversion technology is used to trace back the key characteristic variables that caused the abnormal state and their change trajectory. Compared with the limitations of conventional technology, which can only identify abnormal states but cannot identify the root cause of the abnormality, it can accurately locate the core factors that cause repayment ability abnormalities, clearly present the process, pattern and magnitude of the change of key characteristic variables over time, and clarify the time node and evolution path of the abnormality. The source of risk can be locked without relying on experience judgment, realizing a closed loop from abnormality identification to root cause location. At the same time, it can clearly present the correlation logic between key characteristic variables and abnormal states, providing clear guidance for the formulation of subsequent risk prevention and control measures. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for dynamic prediction of repayment ability and overdue early warning based on multi-source time series data as described in this invention; Figure 2 A flowchart for timeline normalization and data quality repair; Figure 3 A flowchart for inferring repayment ability status using a probabilistic graphical model; Figure 4 A comparison chart of core features before and after the risk threshold; Figure 5 This is a time series analysis diagram of multidimensional risk characteristics. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 The system acquires multi-source time-series data of the target object over a continuous time period and performs timeline normalization and data quality repair operations. This operation standardizes the data and imputes missing values, generating a clean time-series data stream with a unified timestamp. Based on this clean time-series data stream, it constructs three feature sets: income stability, expenditure volatility, and debt pressure. These three feature sets together constitute a multi-dimensional risk feature space to comprehensively characterize the target object's dynamic financial situation. This multi-dimensional risk feature space is input into a probabilistic graphical model based on state transitions. Through inference calculations, the model outputs the probability distribution of the target object's repayment ability at the current time point, quantifying the likelihood of it being in different risk states. The system calls an inversion inference engine to perform root cause analysis on abnormal states in the low-probability range of the repayment ability probability distribution. This engine uses state inversion technology to trace back the key feature variables and their change trajectories that led to the abnormal states. The inverted key feature variables and their change trajectories are compared with pre-set risk thresholds to determine the risk trigger point and risk level, completing the closed loop from data to risk decision-making.

[0021] In one embodiment of the present invention, see [reference] Figure 2This involves acquiring multi-source time-series data of the target object over a continuous time period. This multi-source time-series data includes at least bank transaction records, consumption behavior data, credit history data, and occupational change data. In the example scenario, the target object is an individual user. Bank transaction data includes daily deposit, withdrawal, and transfer records; consumption behavior data includes online and offline transaction details; credit history data includes credit card repayment and loan inquiry logs; and occupational change data includes employment contract start and end dates and salary adjustment records. The system processes transaction records in bank statements, summarizing all inflows and outflows on a daily basis to calculate the daily net inflow value (sum of inflows minus sum of outflows). It also sums all inflows and outflows on a weekly basis to calculate the weekly net inflow value (sum of weekly inflows minus sum of weekly outflows), and summarizes all inflows and outflows on a monthly basis to calculate the monthly net inflow value (sum of monthly inflows minus sum of monthly outflows). Simultaneously, it calculates the net outflow value for each time granularity, forming a time-series sub-database of fund flows. This sub-database includes fields for date stamp, daily net inflow value, daily net outflow value, weekly net inflow value, weekly net outflow value, monthly net inflow value, and monthly net outflow value.

[0022] In some embodiments, the transaction location and transaction type in the consumer behavior data are desensitized and encoded. The transaction location is mapped to an anonymous identifier using a hash function, and the transaction type is encoded as a numerical category according to a preset classification dictionary. The proportion of non-essential consumption per unit time is calculated by combining the transaction amount. The unit time is set to a calendar month, and the formula for calculating the proportion of non-essential consumption is as follows:

[0023] in: This indicates the proportion of non-essential consumption. Indicates the first The amount of the pen was categorized as a non-essential consumer transaction. This indicates the number of non-essential consumer transactions. Indicates the first Total transaction amount This represents the total number of consumer transactions. The monthly non-essential consumption percentage sequence is calculated, forming a consumption structure time-series sub-data set, which includes a month stamp and a non-essential consumption percentage field. Optionally, the employment date, resignation date, and job salary information in the occupational change data are parsed to construct a time-series state sequence reflecting occupational stability. The state sequence is in days, and each day's state value is assigned based on employment status: "on-the-job" status is assigned the current job salary value, and the resignation interval is assigned zero, forming an occupational status time-series sub-data set, which includes a date stamp and an occupational status value field.

[0024] It is understandable that the repayment and query records in the credit history data are processed and statistically analyzed using a time window sliding window. The time window is set to the most recent 90 days. The recent delinquency frequency is calculated by dividing the number of delinquencies within the time window by the total number of payments due. The number of hard queries is calculated by the number of times the credit report is queried by financial institutions within the time window, forming a time-series sub-data of credit performance. The time-series sub-data of credit performance includes fields such as date stamp, delinquency frequency, and number of hard queries. In practice, the generated time-series sub-data on cash flow, consumption structure, occupational status, and credit performance is integrated. Missing time points are filled in using time interpolation, which employs linear interpolation. For missing daily-level data in the cash flow time-series sub-data, the average of data from adjacent days is used to fill the gaps. For missing monthly data in the consumption structure time-series sub-data, the average of data from adjacent months is used to fill the gaps, achieving strict alignment of the entire timeline and forming a clean time-series data stream. The clean time-series data stream is a multi-dimensional data table with a unified timestamp, where each row corresponds to a time point and each column corresponds to an indicator field of a sub-data.

[0025] In one embodiment of the present invention, a multidimensional risk feature space is constructed based on a clean time-series data stream. The clean time-series data stream includes time-series sub-data on cash flow, time-series sub-data on consumption structure, and time-series sub-data on credit performance. The cash flow time-series sub-data in the clean time-series data stream is processed, and a moving average of the daily net inflow value sequence is calculated. The window length of the moving average is set to 30 days. The slope of this moving average within the most recent calculation window is extracted as an indicator reflecting the income growth trend, and the variance of this moving average within the same calculation window is calculated as an indicator reflecting the income volatility. The slope and variance are combined to form an income stability feature set. The consumption structure time-series sub-data in the clean time-series data stream is processed, and the Shannon entropy value of the non-essential consumption proportion sequence in the consumption structure time-series sub-data is calculated. The Shannon entropy value calculation formula is:

[0026] in: Represents Shannon's entropy value. Indicates the first The proportion of this type of spending to the total spending. The total number of consumption categories is represented by Shannon entropy, which reflects the dispersion of consumption behavior. At the same time, the kurtosis of the non-essential consumption proportion sequence is calculated to reflect the concentration of abnormal consumption. The Shannon entropy and kurtosis are combined to form the expenditure volatility characteristic set.

[0027] In some embodiments, the cash flow time-series sub-data and credit performance time-series sub-data in the clean time-series data stream are comprehensively processed and calculated using two dimensions: debt-to-income ratio and remaining credit line utilization rate. The debt-to-income ratio is calculated by dividing the target entity's total debt repayment amount in the past month by the monthly net inflow value in the previous month's cash flow time-series sub-data. The remaining credit line utilization rate is calculated by dividing the total used amount of all current credit products of the target entity by the total credit line. The debt-to-income ratio and remaining credit line utilization rate are extracted as indicators reflecting short-term debt repayment pressure and together form a debt pressure feature set. The constructed income stability feature set, expenditure volatility feature set, and debt pressure feature set are concatenated along the feature dimensions to form a multi-dimensional risk feature space. The feature vector corresponding to each time point in the multi-dimensional risk feature space is formed by sequentially combining the corresponding elements of the income stability feature set, the expenditure volatility feature set, and the debt pressure feature set.

[0028] Optional, see below Figure 3 The multidimensional risk feature space is input into a state transition-based probabilistic graphical model. Three mutually exclusive repayment capacity state nodes are defined in the model: robust, stressed, and deteriorating. A 3x3 state transition matrix is ​​initialized, where each element represents the conditional probability of transitioning from one repayment capacity state to another. The initial values ​​of the conditional probabilities are set to a uniform distribution, and their dynamic values ​​are determined by the Euclidean distance difference between feature vectors at adjacent time points in the multidimensional risk feature space; a larger difference results in a higher weight for the conditional probability of a state transition. Based on the current time-series data points, feature vectors corresponding to the current time points are extracted from the multidimensional risk feature space, and the values ​​in the state transition matrix are updated using maximum likelihood estimation to better match the transition probabilities of the state transition matrix with the evolution patterns of current and historical feature vectors.

[0029] It can be understood that the forward algorithm traverses all time-series data points in the multidimensional risk feature space, starting from the repayment ability state probability distribution and the updated state transition matrix at the previous time point, and combining the feature vector observations at the current time point, recursively calculating the state probability at the current time point, and cumulatively calculating the posterior probabilities of the target object being in a robust state, a stressed state, and a deteriorating state at the current time point, respectively. These three posterior probability values ​​are then summarized to form the repayment ability state probability distribution, which is a three-dimensional probability vector.

[0030] In one embodiment of the present invention, the construction steps of the inversion inference engine include the following process: A mirror model is created in memory that is identical to the probabilistic graphical model structure based on state transitions. The mirror model contains the same state nodes and directed edges for state transitions. In the mirror model, the transition direction of each directed edge in the state transition matrix is ​​physically reversed to construct a reverse link. In the reverse link, all predecessor state nodes can be traced back from the current state node. A weighted backtracking algorithm is defined for each reversed directed edge in the reverse link. The weighted backtracking algorithm takes the posterior probability calculated by forward inference as input and calculates the probability contribution value of the reverse transition according to Bayes' theorem. At each predecessor feature node in the reverse link, a contribution calculator is deployed. The contribution calculator is used to execute the weighted backtracking algorithm and outputs the contribution of each predecessor feature node to the target abnormal state. The completed computational framework, including the reverse link, the weighted backtracking algorithm, and the contribution calculator, is encapsulated as an inversion inference engine.

[0031] In some embodiments, an inversion engine is used for root cause analysis. Repayment ability states with a probability value below a preset warning threshold are selected from the probability distribution of repayment ability states as anomalous states to be analyzed. The preset warning threshold is set to 0.1. The inversion engine is activated, and the anomalous state is set as the termination node. Backtracking is performed along the reverse path of the probabilistic graphical model based on state transitions. During the backtracking process, the contribution of each predecessor feature node to the anomalous state is calculated. The contribution is obtained by calculating the partial derivative of the change in feature variables with the change in state transition probability. The contribution calculation formula is:

[0032] in: Indicates the first The contribution of each predecessor feature node. This represents the change in the state transition probability. Indicates the first The change in each characteristic variable.

[0033] Optionally, feature variables with an absolute contribution value greater than a preset threshold (set to 0.5) are identified and marked as key feature variables. The value sequence of these key feature variables is extracted along the time axis in a forward direction over a historical period (90 days prior to the occurrence of the anomaly). The trajectory of these key feature variables is plotted as a line graph with time on the horizontal axis and the normalized value of the feature variable on the vertical axis. In essence, the inverse analysis engine makes it possible to trace back from probability output to specific influencing factors, and the weighted backtracking algorithm achieves the reverse decomposition of the probability flow. The contribution calculator quantifies the impact of each feature, and the trajectory of key feature variables visually demonstrates the evolution of risk factors in a time-series manner.

[0034] In one embodiment of the present invention, a pre-configured single-dimensional risk threshold table for each key feature variable is retrieved. This table defines the normal range, attention range, and danger range for each feature variable in a structured form. Using an example scenario, assume three key feature variables are identified from the inversion engine: debt-to-income ratio, proportion of non-essential consumption, and occupational status value. The specific contents of the single-dimensional risk threshold table are shown in Table 1. Each data point in the trajectory of the key feature variable's change is mapped to the corresponding range in the risk threshold table. For example, if the debt-to-income ratio is 0.5 at time point T1, it falls within the normal range according to Table 1. The time point in the trajectory of the key feature variable's change that first falls into the danger range is then found. Assuming the debt-to-income ratio is 0.85 at time point Tn, first meeting the danger range condition of "greater than 0.8," time point Tn is marked as the risk trigger threshold. The number of key characteristic variables falling into the danger zone within the time window approaching the risk trigger threshold was counted. The time window was set to 7 days before and after the risk trigger threshold. Assuming that within the window [Tn-7, Tn+7], the debt-to-income ratio remained in the danger zone, the proportion of non-essential consumption also fell into the danger zone at time point Tn+2, while the occupational status value remained in the normal range, then the number of key characteristic variables falling into the danger zone would be 2. See Table 1.

[0035] Table 1: Single-Dimensional Risk Threshold Table ; In some embodiments, risk levels are determined based on the number of key characteristic variables falling into the danger zone. The classification rule is as follows: one key characteristic variable is considered low risk, two are considered medium risk, and three or more are considered high risk. Based on the aforementioned example, the number of key characteristic variables falling into the danger zone is two, therefore the risk level is determined to be medium risk. The corresponding early warning signal is activated based on the determined risk level, and a risk diagnosis report containing the risk triggers and key characteristic fluctuations is generated. The specific implementation includes the following process: Based on the determined risk level, a corresponding early warning signal template is matched from the early warning strategy library. The early warning signal template specifies the color coding and notification frequency for the early warning; for example, medium risk corresponds to an "orange" early warning code and a notification frequency of "once a day." Placeholders in the early warning signal template are filled to generate a specific early warning instruction, which includes the early warning level, the risk trigger threshold time, and the names of the key characteristic variables involved.

[0036] Optionally, original data snapshots of key characteristic variables within two time windows before and after the risk trigger threshold are extracted from the multidimensional risk feature space. Each time window is set to 3 days. The original numerical sequences of the debt-to-income ratio and the proportion of non-essential consumption within the time window [Tn-3, Tn+3] are extracted. The original data snapshots of the key characteristic variables are integrated with their corresponding trend charts, which are historical trend charts covering a longer time range. The textual description of the risk trigger is automatically generated based on the range and direction of change of the key characteristic variables, such as "The debt-to-income ratio continues to rise and exceeds the danger threshold, while the proportion of non-essential consumption has recently shown an abnormal peak." A risk diagnosis report is compiled, presented in a structured document format, including a risk summary, threshold information, detailed analysis of key characteristic variables, and visualization charts. The risk level determination formula can be understood as follows:

[0037] in: This indicates the number of items obtained through statistics. It is an indicator function that takes the value 1 when the condition inside the parentheses is true and 0 otherwise. Indicates the first Key feature variables, Indicates the danger zone. This represents the total number of key characteristic variables. Matching of the warning signal template is based on risk level. The generation of risk diagnostic briefings, accomplished through numerical indexing, encapsulates the process from data to risk interpretation.

[0038] See Figure 4 This chart compares the core characteristics before and after the risk threshold, showcasing the changing trends of two key risk features. The debt-to-income ratio steadily climbed from 0.656 over 17 days, breaking the danger threshold of 0.8 for the first time on day 21, and then remaining within the danger range; this is the core driving factor behind this risk. The proportion of non-essential consumption also rose, from 0.325 over 17 days to 0.497 over 23 days. Although it did not enter the danger range, its continued rise exacerbated overall debt repayment pressure. The continued expansion of debt, coupled with the lack of effective control over non-essential consumption, led to a rapid accumulation of cash flow pressure. An orange alert should be triggered at this time, with notifications issued daily, focusing on the subsequent trend of the debt-to-income ratio. Prioritize reducing debt while restricting non-essential consumption expenditures to prevent the risk from escalating further into a high-level risk.

[0039] In one embodiment of the present invention, sensitivity analysis is performed on key characteristic variables in the risk diagnosis briefing to quantify the impact coefficient of each key characteristic variable on the transition of repayment ability status. Using an example scenario, it is assumed that the risk diagnosis briefing contains three key characteristic variables: debt-to-income ratio, proportion of non-essential consumption, and occupational status value. In the probabilistic graphical model based on state transition, all other characteristic variables except for one key characteristic variable are frozen; for example, the proportion of non-essential consumption and occupational status value are frozen, leaving only the debt-to-income ratio as a variable input. A small perturbation is artificially added to the value of the debt-to-income ratio, set at 0.01 times its original value. The magnitude of the change in the probability distribution of repayment ability status is observed; this magnitude is obtained by calculating the Euclidean distance between the probability distribution vectors before and after the perturbation. The ratio of the magnitude of the change in the probability distribution of repayment ability status to the small perturbation is calculated, and this ratio is defined as the univariate influence coefficient of the debt-to-income ratio, the key characteristic variable. The formula for calculating the influence coefficient is:

[0040] in: Indicates the first Univariate influence coefficients of key characteristic variables, This represents the probability distribution vector of repayment ability after the perturbation is applied. This represents the probability distribution vector of the original repayment ability state. Represents the Euclidean distance between vectors. Indicates application to the first The absolute value of small perturbations on key feature variables.

[0041] In some embodiments, the perturbation and calculation process is repeated for all key feature variables. Small perturbations with the same rules are applied sequentially to the proportion of non-essential consumption and the occupational status value, and the magnitude of change in the probability distribution of repayment ability is observed. The influence coefficients of the proportion of non-essential consumption and the occupational status value are calculated respectively. The influence coefficients are used to fine-tune the parameters of the probabilistic graphical model based on state transition to correct the model's bias in predicting future states. The absolute values ​​of the influence coefficients are ranked; for example, the debt-to-income ratio has the largest absolute value, followed by the proportion of non-essential consumption, and then the occupational status value. Several key feature variables with the highest ranking are selected as dominant variables. Setting the selection quantity to two, the debt-to-income ratio and the proportion of non-essential consumption are determined as the dominant variables.

[0042] Optionally, the bias parameter of the state transition matrix in the probabilistic graphical model based on state transition can be adjusted according to the sign of the influence coefficient of the dominant variable. A positive influence coefficient indicates that the growth of the variable will increase the state transition probability, and the bias parameter is adjusted in the direction of increasing the transition probability; a negative influence coefficient is adjusted in the direction of decreasing the transition probability. Gradient descent is used to minimize the model's prediction error on historical data. The prediction error is calculated by comparing the cross-entropy loss between the state probability distribution predicted by the model and the subsequently observed state labels. The adjusted bias parameter and other relevant parameters are iteratively optimized. This iterative optimization process continues until the model's prediction error converges to an acceptable range. The convergence condition is set to the loss function value decreasing by less than a preset threshold for five consecutive iterations. The iteration stops, and the model parameters at this point are saved as the fine-tuned parameter configuration. The probabilistic graphical model based on state transition, after parameter fine-tuning, will be applied to the data of the next time window and begin a new round of dynamic prediction of repayment ability and overdue warning cycle.

[0043] See Figure 5 This is a time-series analysis chart of multidimensional risk characteristics. Income stability rose continuously from January to March, peaking at 0.92 in March, far exceeding the stability threshold of 0.7, demonstrating excellent performance. It gradually declined from April to August, falling to 0.69 in August, breaking below the stability threshold for the first time and entering a volatile range. It fluctuated and rebounded from September to December, returning to 0.84 in December, remaining at a healthy level overall, but with significant volatility. Expenditure volatility declined continuously from January to April, bottoming out at 0.11 in April, far below the 0.5 threshold, indicating highly controllable spending behavior. It rebounded rapidly from May to June, reaching 0.52 in June, breaking the high volatility threshold for the first time. It fell back to a controllable range from July to September, bottoming out at 0.10 in September; it rebounded again from October to December, exhibiting a cyclical fluctuation of "low-high-low-high". Debt pressure remained at 0.62 from February to March, slightly above the 0.6 threshold, entering a short-term high-pressure state. From April to June, it fell back to the 0.2–0.3 range, significantly easing the pressure. From July to August, it climbed again to 0.63, returning to the high-pressure range; after September, it gradually declined, bottoming out at 0.17 in November, and then rebounded slightly in December.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data, characterized in that, include: Acquire multi-source time-series data of the target object within a continuous time period, perform time axis normalization and data quality repair operations on the multi-source time-series data, and generate a set of clean time-series data streams with unified timestamps; Based on the clean time-series data stream, an income stability feature set, an expenditure volatility feature set, and a debt pressure feature set are constructed respectively. The income stability feature set, the expenditure volatility feature set, and the debt pressure feature set together constitute a multidimensional risk feature space. The multidimensional risk feature space is input into the state transition-based probabilistic graphical model, which, through inference calculation, outputs the probability distribution of the target object's repayment ability status at the current time point. The inversion and inference engine is invoked to perform root cause analysis on the abnormal states that are in the low probability range of the repayment ability state probability distribution. Through state inversion technology, the key feature variables that caused the abnormal state and their change trajectory are traced in reverse. The key feature variables and their change trajectories obtained from the inversion are compared with the pre-set risk thresholds to determine the critical point for risk triggering and the risk level.

2. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 1, characterized in that, Perform time axis normalization and data quality repair operations on the multi-source time-series data to generate a set of clean time-series data streams with unified timestamps, including: The multi-source time-series data includes at least bank transaction data, consumer behavior data, credit history data, and occupational change data; The transaction records in the bank statement data are summarized by three time granularities: daily, weekly, and monthly. The net inflow and net outflow of funds at each time granularity are calculated to form the time series sub-data of fund flow. The transaction location and transaction type in the aforementioned consumer behavior data are desensitized and encoded. Combined with the transaction amount, the proportion of non-essential consumption per unit time is calculated to form a time-series sub-data of the consumption structure. The entry date, exit date, and job salary information in the occupational change data are analyzed to construct a time-series state sequence reflecting occupational stability, forming occupational state time-series sub-data. The repayment and query records in the credit history data are statistically analyzed using a time window sliding window to calculate the recent overdue frequency and hard query count, forming a time series sub-data of credit performance. The generated time-series sub-data of fund flow, consumption structure, occupation status, and credit performance are interpolated using time interpolation to fill in missing time points, achieving strict alignment of the entire timeline and forming the clean time-series data stream.

3. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 2, characterized in that, Based on the aforementioned clean time-series data stream, income stability feature set, expenditure volatility feature set, and debt pressure feature set are constructed respectively, including: For the time series sub-data of the cash flow in the clean time series data stream, calculate the slope and variance of its moving average line, extract indicators that reflect the income growth trend and fluctuation range, and form the income stability feature set; For the consumption structure time-series sub-data in the clean time-series data stream, calculate its Shannon entropy and kurtosis, extract indicators reflecting the dispersion of consumption behavior and the concentration of abnormal consumption, and form the expenditure volatility feature set; The cash flow time series sub-data and the credit performance time series sub-data in the clean time series data stream are calculated using two dimensions: debt-to-income ratio and remaining credit line utilization rate. Indicators reflecting short-term debt repayment pressure are extracted and formed into the debt pressure feature set. The constructed income stability feature set, expenditure volatility feature set, and debt pressure feature set are concatenated to form the multidimensional risk feature space, wherein the feature vector at each time point is composed of the corresponding elements of these three feature sets.

4. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 3, characterized in that, The multidimensional risk feature space is input into a state transition-based probabilistic graphical model. This model, through inference calculations, outputs the probability distribution of the target object's repayment ability at the current time point, including: In the probabilistic graphical model based on state transition, several mutually exclusive repayment ability state nodes are defined, and the repayment ability state nodes include at least a robust state, a stressed state, and a deteriorating state. Initialize a state transition matrix, where each element of the state transition matrix represents the conditional probability of transitioning from one repayment ability state to another, and the conditional probability is determined by the feature differences between adjacent time points in the multidimensional risk feature space. Based on the current time series data points, extract their feature vectors and update the values ​​in the state transition matrix using the maximum likelihood estimation method; The forward algorithm is run to traverse all time-series data points in the multidimensional risk feature space, and the posterior probability of the target object being in each repayment ability state at the current time point is calculated cumulatively, and the summaries form the repayment ability state probability distribution.

5. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 4, characterized in that, The inversion and deduction engine is invoked to perform root cause analysis on abnormal states in the low-probability range of the repayment ability status probability distribution. Through state inversion technology, the key characteristic variables leading to the abnormal state and their change trajectories are traced back, including: In the probability distribution of repayment ability status, repayment ability statuses with a probability value lower than a preset warning line are selected as abnormal statuses to be analyzed. Activate the inversion and inference engine, set the abnormal state as the termination node, and backtrack along the reverse link of the probabilistic graphical model based on state transition; During the backtracking process, the contribution of each predecessor feature node to the abnormal state is calculated. The contribution is obtained by calculating the partial derivative of the change in the feature variable with the change in the state transition probability. Identify feature variables whose absolute contribution value is greater than a preset threshold and mark them as key feature variables; Along the positive time axis, extract the value sequence of the key feature variable within the historical time period and plot the change trajectory of the key feature variable.

6. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 5, characterized in that, The step of comparing the key feature variables and their change trajectories obtained from the inversion with a pre-set risk threshold to determine the critical point for risk triggering and the risk level includes: Retrieve a pre-configured single-dimensional risk threshold table for each key feature variable, wherein the risk threshold table defines normal range, attention range and danger range; Map each data point in the change trajectory of the key feature variables to the corresponding interval in the risk threshold table; Find the point in time when the change trajectory of the key characteristic variable first falls into the danger zone, and mark that point as the risk triggering critical point; The number of key characteristic variables falling into the danger zone within the time window approaching the risk trigger threshold is counted, and the risk level is divided according to the number of variables: one is considered low risk, two are considered medium risk, and three or more are considered high risk.

7. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 6, characterized in that, Also includes: Based on the determined risk level, activate the corresponding early warning signal and generate a risk diagnosis report containing the risk triggers and key characteristic fluctuations. Sensitivity analysis was performed on the key feature variables in the risk diagnosis briefing to quantify the impact coefficient of each key feature variable on the transition of repayment ability status. Using the influence coefficient, the parameters of the probabilistic graphical model based on state transition are fine-tuned to correct the model's bias in predicting the state at future time points. The state transition-based probabilistic graphical model, after parameter fine-tuning, is applied to the data of the next time window to begin a new cycle of dynamic prediction of repayment ability and overdue warning. Based on the determined risk level, the corresponding early warning signal is activated, and a risk diagnosis report containing the risk triggers and key characteristic fluctuations is generated, specifically including: Based on the determined risk level, a corresponding early warning signal template is matched from the early warning strategy library. The early warning signal template specifies the color code and notification frequency for the early warning. Fill in the placeholders in the warning signal template to generate specific warning instructions; Extract the original data snapshots of the key feature variables within two time windows before and after the risk triggering critical point from the multidimensional risk feature space; The original data snapshots of the key feature variables and their corresponding change trajectory charts are integrated together with the textual descriptions of the risk triggers to form the risk diagnosis brief.

8. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 7, characterized in that, Sensitivity analysis was performed on the key characteristic variables in the aforementioned risk diagnosis report to quantify the impact coefficient of each key characteristic variable on the transition of repayment ability status, including: In the state transition-based probabilistic graphical model, all other feature variables except for one key feature variable are frozen. By artificially adding a small perturbation to the value of one of the key characteristic variables, the magnitude of the change in the probability distribution of the repayment ability status is observed. Calculate the ratio of the variation amplitude of the probability distribution of the repayment ability status to the amount of small disturbance, and define the ratio as the univariate influence coefficient of a certain key characteristic variable; By iterating through all the key feature variables and repeating the perturbation and calculation process, the influence coefficient of each key feature variable is obtained.

9. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 8, characterized in that, Using the aforementioned influence coefficients, the parameters of the state transition-based probabilistic graphical model are fine-tuned to correct the model's bias in predicting states at future time points, including: The absolute values ​​of the influence coefficients are sorted, and the top-ranked key feature variables are selected as the dominant variables. Based on the sign of the influence coefficient of the dominant variable, adjust the bias parameter of the state transition matrix in the probabilistic graphical model based on state transition; The gradient descent method is used to iteratively optimize the adjusted bias parameter and other relevant parameters with the goal of minimizing the prediction error of the model on historical data. When the model's prediction error converges to an acceptable range, the iteration stops, and the model parameters at this point are saved as the fine-tuned parameter configuration.

10. The method for dynamic prediction of repayment ability and overdue early warning based on multi-source time-series data as described in claim 9, characterized in that, The construction steps of the inversion inference engine include: Create a mirror model in memory that is exactly the same as the state transition-based probabilistic graphical model structure, the mirror model containing the same state nodes and directed edges for state transitions; In the mirror model, the transition direction of each directed edge in the state transition matrix is ​​physically reversed to construct a reverse link, so that in the reverse link, it is possible to trace back to all the predecessor state nodes from the current state node. A weighted backtracking algorithm is defined for each reversed directed edge in the reverse link. The weighted backtracking algorithm takes the posterior probability calculated by forward inference as input and calculates the probability contribution value of the reverse transition according to Bayes' theorem. At each preceding feature node of the reverse link, a contribution calculator is deployed. The contribution calculator is used to execute the weight backtracking algorithm and output the contribution of each preceding feature node to the target abnormal state. The completed computational framework, which includes the reverse link, the weight backtracking algorithm, and the contribution calculator, is encapsulated as the inversion inference engine. When an abnormal state to be analyzed is input, the reverse tracing process can be automatically started.

Images