Financial delinquency risk intelligent early warning system and algorithm fusing time series data
By fusion of six types of multi-source time-series data in a hierarchical and slicing manner and extracting adaptive time-series features, combined with dynamic weight iterative optimization, the problems of fragmented time-series data processing and model rigidity in financial overdue risk warning are solved. This achieves high-precision and low-cost risk warning, meeting the high adaptability and security requirements of financial institutions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 张万才
- Filing Date
- 2026-05-28
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies for financial overdue risk early warning suffer from problems such as fragmented time-series data processing, rigid early warning algorithm models, single risk quantification, lack of operability of early warning results, simple data preprocessing methods, and insufficient security. These issues result in early warning lag and insufficient accuracy, failing to meet the high precision, high timeliness, and high adaptability requirements of financial institutions.
Employing a distributed server architecture, the system integrates six types of multi-source time-series data through hierarchical slicing and fusion, adaptive time-series feature extraction, dynamic weight iteration optimization, and risk classification. Combined with a unique fusion weight calculation, risk index quantification, and dynamic iteration optimization formula, it forms a closed-loop intelligent management and control system that enables deep coupling of multi-dimensional time-series data and adaptive risk warning.
It achieves accurate early warning of financial overdue risks 7-60 days in advance, with an accuracy rate of over 90%, and the false positive and false negative rates are controlled within 5%. The risk classification is clear and operable, the data security is high, the adaptability is strong, the operation and maintenance costs are reduced, and the actual needs of financial institutions are met.
Smart Images
Figure CN122390858A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of financial risk management and artificial intelligence algorithm integration technology, and specifically relates to a financial overdue risk intelligent early warning system and algorithm that integrates time series data. Background Technology
[0002] Currently, the digital transformation of the financial industry is accelerating, and the scale of various credit businesses continues to expand. Overdue risk management has become a core pain point and key link in the sound operation of financial institutions. Traditional overdue risk early warning methods mostly rely on static credit data, single financial indicators (such as asset-liability ratio, revenue growth rate), or simple threshold judgment methods, which can only achieve "post-event early warning" and cannot capture the dynamic pattern of risk evolution in advance. The problem of early warning lag is prominent, often causing financial institutions to miss the best intervention opportunity and resulting in a large accumulation of non-performing assets. With the popularization of time series data in the financial field, some existing technologies have begun to try to introduce time series data to optimize early warning models, but there are still many technical shortcomings that are difficult to overcome. Moreover, the core logic is highly overlapping with existing technologies, which cannot meet the actual needs of financial institutions for highly accurate, timely, and adaptable early warnings. The specific defects are as follows: First, time-series data processing suffers from severe fragmentation and insufficient integration. Most existing technologies only perform simple statistical analysis on single time-series data (such as repayment record time series), failing to achieve deep coupling of multi-source time-series data, including user repayment behavior time series, account fund flow time series, personal / corporate credit fluctuation time series, business operation behavior time series, macroeconomic environment time series, and industry prosperity time series. They often employ a "piecing together" fusion method, failing to capture the linkages and collaborative risks between different time-series data, resulting in incomplete risk feature extraction and significantly reduced early warning accuracy. For example, some existing patents only focus on users' historical repayment time-series data, ignoring the impact of short-term changes in account fund flow (such as large fund transfers or flow interruptions) and dynamic fluctuations in credit scores on delinquency risk, thus failing to comprehensively depict the user's risk status.
[0003] Secondly, the early warning algorithm models are rigid and lack adaptive optimization capabilities. Existing early warning algorithms mostly employ machine learning models with fixed weight parameters (such as traditional logistic regression and fixed-structure neural networks). Once these weight parameters are set, they remain unchanged for a long period, failing to adapt to changes in risk characteristics across different business scenarios (such as the difference in risk characteristics between personal consumer loans and business loans) and different user groups (such as the difference in repayment habits between young users and middle-aged and elderly users). This results in persistently high false positive and false negative rates. Furthermore, existing algorithms lack dynamic iterative optimization mechanisms and cannot automatically update model parameters based on real-time business data and changes in the market environment. Over long-term use, the model's accuracy will continuously decline, making it difficult to adapt to the dynamic changes in the financial market.
[0004] Third, the risk quantification dimension is too narrow, and the problem of delayed early warning has not been fundamentally solved. Current risk quantification technologies mostly focus on traditional dimensions such as historical delinquency records and static financial indicators, ignoring key temporal characteristics such as short-term behavioral changes (e.g., multiple delinquencies in a short period of time, a sudden increase in the frequency of violations), medium- and long-term trend deviations (e.g., a continuous decline in credit scores for three consecutive months, a continuous shrinking of cash flow), and cross-dimensional related risks (e.g., the linkage between a decline in industry prosperity and a decline in corporate repayment ability). Warnings can only be issued after users show obvious signs of delinquency, and it is impossible to achieve accurate prediction 7-30 days in advance, which makes it difficult to meet the core needs of financial institutions to intervene in advance and reduce the non-performing loan ratio.
[0005] Fourth, risk classification is crude, and early warning results lack operability. Most existing technologies only classify risks into two categories: "overdue" and "not overdue," or simply into three levels: low, medium, and high. The classification standards are fixed and crude, and they do not combine time-series anomalies and risk evolution trends to formulate differentiated intervention strategies. As a result, the early warning results lack operability, and risk control personnel cannot quickly locate the core causes of risks and take targeted intervention measures based on early warning information, which reduces the practical application value of the early warning system.
[0006] Fifth, the data preprocessing methods are simplistic, making it difficult to guarantee data quality. Existing technologies for preprocessing time-series data often employ single methods such as missing value imputation (e.g., fixed value imputation, mean imputation) and outlier removal (e.g., simple threshold removal). They fail to develop targeted preprocessing rules that consider the temporal continuity and contextual relevance of the time-series data. This can easily lead to the accidental deletion of valid data and the retention of invalid data, affecting the accuracy and reliability of subsequent algorithmic calculations. Furthermore, some existing technologies do not perform rigorous anonymization and encryption of the collected data, posing a risk of user privacy leaks and failing to comply with financial industry data security standards.
[0007] Furthermore, existing patents and technical solutions either focus solely on the processing and analysis of single time-series data, employ simple fitting using general machine learning models, or rely on human experience for weight setting. None of these solutions form a complete closed-loop technical system encompassing "full-dimensional fusion of time-series data - adaptive feature extraction - dynamic weight iteration - risk grading and quantification - real-time early warning push - data traceability and review." Moreover, the core algorithms, data processing logic, and system architecture exhibit significant overlap with existing technologies, failing to meet the actual needs of financial institutions for highly accurate, timely, adaptable, and secure overdue risk early warning systems. Therefore, developing a financial overdue risk intelligent early warning system and algorithm that completely circumvents existing technologies, possesses an original time-series fusion mechanism, a proprietary algorithm, and a complete closed-loop system has become an urgent technical problem to be solved in this field. This is of great significance for promoting the digital and intelligent development of financial risk management. In addition, existing solutions generally do not consider model drift issues under "black swan" events such as pandemics and policy changes, nor do they cover the computing power limitations and data availability differences of small and medium-sized financial institutions. They also lack forward-looking designs regarding algorithm fairness, interpretability, and cross-border data compliance, making it difficult to adapt to increasingly stringent financial regulatory requirements. Summary of the Invention
[0008] In view of this, the present invention provides a financial overdue risk intelligent early warning system and algorithm that integrates time series data to solve or alleviate one of the technical problems existing in the prior art, and at least provides a beneficial option.
[0009] The technical solution of this invention is implemented as follows: A financial overdue risk intelligent early warning system and algorithm integrating time-series data includes seven collaborative modules: a hardware support layer, a data acquisition layer, a time-series data fusion and processing layer, an algorithm core operation layer, a risk early warning output layer, an iterative optimization layer, and a data traceability layer. This forms a closed-loop intelligent management system covering the entire process of "acquisition-preprocessing-fusion-operation-early warning-optimization-traceability." The functions of each module are as follows: Hardware support layer: Adopting a distributed server architecture, it provides server computing power, massive time-series data storage, high-speed network transmission and data security protection support, and has built-in data encryption module and security audit module; Data Acquisition Layer: This layer collects six types of multi-source time-series data from financial business scenarios, including time-series data on user / corporate repayment behavior, account fund flow, personal / corporate credit fluctuations, business operation behavior, macroeconomic environment, and industry prosperity. The acquisition process uses AES-256 de-identification and encryption, and the time-series data is divided into slices at daily, weekly, monthly, and quarterly granularities. Different time-series data use independent acquisition channels and verification mechanisms. Time series data fusion processing layer: The collected multi-source time series data is processed by layer cleaning, deduplication, missing value imputation, outlier removal and normalization. The layered slicing fusion logic completes the deep coupling of multi-dimensional time series data. The initial fusion weights are allocated according to the time series distance, data stability and risk correlation to generate a fused comprehensive time series feature dataset. The core computational layer of the algorithm has a built-in self-developed intelligent early warning algorithm for overdue risk. It loads the time-series fused data to complete adaptive time-series feature extraction, risk index quantification and refined risk classification. It has unsupervised adaptive learning capabilities and does not require manual intervention in feature selection and parameter adjustment. Risk warning output layer: Generates multi-level warning instructions based on risk classification results, and pushes them synchronously to financial risk control terminals, mobile risk control APPs and relevant business managers. It marks the risk level, warning reason, time-series anomaly points, risk evolution trend and targeted intervention suggestions, and supports encrypted push and hierarchical viewing of warning information. Iterative optimization layer: Real-time acquisition of actual overdue result data and intervention effect data after risk warning, which are then fed back into the core computing layer of the algorithm. The model weights are adaptively updated through dynamic weight iteration formulas. At the same time, the risk classification threshold and feature extraction rules are automatically optimized according to changes in business scenarios and market environment. Data traceability layer: Records the collection time, processing process, fusion logic, algorithm operation process, early warning results and intervention effect data of multi-source time series data, forming a complete data traceability chain, supporting reverse tracing of early warning results, and meeting the regulatory compliance requirements of the financial industry.
[0010] As a preferred option, it includes three core, original mathematical formulas, which are used for time-series data fusion weight calculation, risk index quantification, and dynamic weight iterative optimization, respectively. The specific formulas are as follows: Formula 1: Formula for calculating the weights of hierarchical fusion of multi-source time series data:
[0011] In the formula: Let be the fusion weight of the i-th class of time series data under the t-th time series slice, with a value range of [0,1]. The sum of the fusion weights of all classes of time series data is 1. The stability coefficient of the i-th type of time series data is calculated from the variance of the fluctuations of this type of time series data in the t-th slice. For the timeliness weight of the i-th type of time series data, ,in This represents the number of days between the time-series slice and the current time. This is the time-dependent attenuation coefficient, with a value ranging from 0.01 to 0.05. Let be the risk correlation coefficient of the i-th type of time series data, with a value range of [0,1]; n is the total number of time series data types, n=6; This is the dynamic correction coefficient for time-series slices, with a value range of [0.8, 1.2]. Formula 2: Quantitative Formula for the Comprehensive Index of Financial Overdue Risk:
[0012] In the formula: This is a comprehensive index for overdue risk, with a value range of [0, 100]. represents the risk feature value of the i-th type of time series data, with a value range of [0, 100]. This is the trend offset correction coefficient, with a value range of [0.1, 0.5]. This represents the offset of short-term temporal behavior abrupt changes. ; C is the cross-dimensional association risk coefficient, with a value range of [0.05, 0.2]; C is the cross-dimensional association risk value, with a value range of [0, 20]. Formula 3: Formula for dynamic weight iteration optimization of the algorithm model:
[0013] In the formula: The fusion weights are the fusion weights for the i-th type of time series data in the (t+1)-th time series slice after iteration; The fusion weights for the i-th type of time series data in the t-th time slice before iteration; For the iterative learning rate:
[0014] Where Acc represents the model's early warning accuracy (%). This is the actual overdue risk label value, with overdue payments recorded as 100, non-overdue payments recorded as 0, and partial overdue payments recorded as the corresponding risk percentage value; Predict risk indices for the model; These are the partial derivatives of the weights; The value of is the iterative constraint coefficient, which ranges from [0.9, 1.1].
[0015] Preferably, the missing value imputation rule of the time-series data fusion processing layer is as follows: Missing data with a missing rate of <10% are imputed using the weighted mean or mean of adjacent time series data. For missing data with a missing rate of 10% ≤ 30%, imputation was performed using the average value of similar users / enterprises during the same period. Slices with a missing rate >30% are removed from the corresponding time series. The outlier removal rule is as follows: based on different time series data types, remove data that exceed a reasonable threshold or deviate from the mean ± 2 standard deviations.
[0016] Preferably, the risk level of the risk warning output layer is divided into 5 levels, namely low risk [0,20), low-medium risk [20,40), medium risk [40,60), medium-high risk [60,80], and high risk [80,100]. Different risk levels correspond to different warning intensities and intervention strategies.
[0017] Preferably, the iteration cycle of the iterative optimization layer can be adjusted according to the business scenario. For personal credit scenarios, daily granular iteration is adopted, and for corporate business loan scenarios, monthly granular iteration is adopted. The iteration reaches the convergence state when the model's warning accuracy fluctuation is <3% and the false positive rate and false negative rate are both <5% for 10 consecutive time slices.
[0018] Preferably, the six types of time-series data in the data acquisition layer are as follows: Repayment behavior timeline data: including repayment amount, repayment time, repayment delay duration, number of overdue payments, number of overdue days, and amount of outstanding interest; Account fund flow time sequence data: including income / operating income amount, expenditure amount, balance, large fund transfer out / in records, number of flow interruptions, and upstream and downstream payment records; Credit fluctuation time series data: including personal / corporate credit scores / ratings, number of new overdue records, debt ratio, number of inquiries, guarantee status, and litigation records; Business operation behavior time sequence data: including the number of loan / credit card applications, the number of credit limit adjustments, the number of repayment method changes, the number of violations, and guarantee change records; Macroeconomic environment time series data: including the Consumer Price Index (CPI), GDP growth rate, industrial added value, short-term loan interest rate, unemployment rate, and inflation rate; Industry prosperity time series data: including the corresponding industry prosperity index, industry non-performing loan ratio, and industry revenue growth rate.
[0019] Preferably, the adaptive time-series feature extraction algorithm can automatically identify the core risk features of different financial business scenarios without the need for manual feature pre-setting, and can extract three types of core time-series features: short-term behavioral mutations, medium- and long-term trend shifts, and cross-dimensional correlation risks.
[0020] Preferably, the system can be adapted to all scenarios of financial business, such as personal loans, business loans, consumer finance, credit card business, and micro-loans, and automatically adapts to the differences in risk characteristics of different scenarios through iterative optimization layers.
[0021] Preferably, the hardware support layer can adapt to the real-time collection, processing and storage needs of tens of millions of user / enterprise time-series data, and the data encryption module adopts the AES-256 encryption standard to ensure data security and user privacy, which complies with the data security specifications of the financial industry.
[0022] Preferably, the algorithm can accurately predict financial delinquency risks 7-60 days in advance, with 7-30 days in advance for personal consumer loans and credit card scenarios, and 15-60 days in advance for business loan scenarios.
[0023] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: 1. This invention innovatively adopts a six-category multi-source time-series data hierarchical slicing and fusion mechanism, which differs from the existing single or simple splicing time-series data processing methods. Combined with an original fusion weight calculation formula (Formula 1), it comprehensively considers data stability, timeliness, and risk correlation, and achieves deep coupling of multi-dimensional time-series data such as repayment behavior, cash flow, credit fluctuations, macroeconomics, and industry prosperity. It can comprehensively capture the dynamic laws of risk evolution and completely solve the defects of fragmented time-series data processing and incomplete risk feature extraction in existing technologies.
[0024] 2. The algorithm of this invention incorporates an original dynamic weight iteration optimization formula (Formula 3), which can automatically update the model weights and feature extraction rules based on the real-time overdue results and intervention effects, without manual intervention. This solves the problems of existing algorithms such as model solidification, fixed weights, and continuous decline in accuracy. At the same time, the adaptive time-series feature extraction algorithm can automatically adapt to different business scenarios without the need for manual feature pre-setting, further improving the accuracy of early warning. Experimental verification shows that the early warning accuracy rate can reach over 90%, and the false positive rate and false negative rate are both controlled within 5%, which is significantly better than existing traditional algorithms.
[0025] 3. Through the original comprehensive risk index quantification formula (Formula 2), combined with the quantification of short-term behavioral change offset and cross-dimensional correlation risk, it can accurately predict overdue risk and achieve early warning 7-60 days in advance. Among them, personal consumer loans and credit card scenarios are 7-30 days in advance, and business loan scenarios are 15-60 days in advance. This completely solves the pain points of existing technology's lagging early warning and inability to intervene in advance, giving financial institutions sufficient time to intervene in risks and effectively reducing the non-performing asset ratio.
[0026] 4. The system and algorithm of this invention can be flexibly adapted to financial businesses across all scenarios, including personal loans, business loans, consumer finance, credit card business, and micro-loans. Through iterative optimization, the iteration cycle, weight allocation, and risk classification threshold are automatically adjusted to adapt to the differences in risk characteristics of different scenarios. There is no need to develop separate early warning models for different scenarios, which greatly reduces the operation and maintenance costs and development costs of financial institutions. Its adaptability is significantly better than existing single-scenario early warning technologies.
[0027] 5. This invention divides risks into 5 levels, each corresponding to different warning intensities and intervention strategies. It clearly marks risk anomalies, evolution trends, and targeted recommendations. Unlike the crude risk classification methods of existing technologies, this invention allows risk control personnel to quickly locate the core causes of risks and take precise intervention measures, significantly improving the efficiency of risk control work. At the same time, the data traceability layer forms a complete traceability chain, meeting the regulatory compliance requirements of the financial industry and improving the credibility and traceability of warning results.
[0028] 6. The data acquisition process employs AES-256 de-identification encryption, and the hardware support layer incorporates a data encryption and security audit module to strictly protect user / enterprise privacy and data security. This addresses the issues of insufficient data security protection and high privacy leakage risks in existing technologies, fully complying with financial industry data security standards. Simultaneously, targeted missing value imputation and outlier removal rules ensure data quality, provide reliable support for algorithm computation, and further improve the accuracy of early warning results.
[0029] 7. This invention constructs a closed-loop system covering the entire process of "data collection-preprocessing-fusion-computation-early warning-optimization-tracing". Each module works in concert, from data collection to early warning output and iterative optimization, without the need for manual intervention, which greatly reduces the manpower operation and maintenance costs of financial institutions. At the same time, the iteration frequency can be automatically adjusted after the model converges, taking into account both optimization efficiency and system energy consumption, so as to achieve efficient and low-cost risk management.
[0030] 8. Strong robustness in extreme scenarios: The model drift problem during special periods is solved through the circuit breaker mechanism for abnormal events and the public opinion adjustment factor; the sliding window incremental computing and edge deployment scheme take into account both high concurrency and real-time performance and computing cost.
[0031] 9. Highly inclusive and adaptable: Supports modular purchasing and SaaS-based pay-as-you-go pricing, lowering the access threshold for small and medium-sized city commercial banks and rural commercial banks; provides industry data alternatives to address the pain point of data gaps for small and medium-sized institutions.
[0032] 10. Strong regulatory compliance: Built-in fairness verification and algorithm decision logs avoid implicit discrimination based on region, age, etc., and meet the requirements of domestic algorithm recommendation regulation and GDPR cross-border transmission.
[0033] 11. Low barrier to entry: The functions of visualized risk transmission, plain-language explanation of weights, and automatic parameter optimization allow risk control personnel to quickly locate the core risks without needing professional modeling capabilities.
[0034] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 This is a flowchart illustrating the overall architecture of the intelligent early warning system for financial overdue risks of the present invention. Figure 2 This is a flowchart of the core algorithm processing of the present invention; Figure 3 This is a flowchart of the dynamic weight iterative optimization process of the present invention. Detailed Implementation
[0037] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.
[0038] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0039] Example 1: Example of Early Warning of Overdue Risk in Personal Consumer Loans
[0040] This embodiment is applied to the personal consumer loan business of a commercial bank. This business targets natural persons aged 18-60, with loan amounts ranging from 1,000 to 50,000 yuan, loan terms of 3-24 months, and repayment methods of equal principal and interest. The core risks are short-term funding gaps for users, repayment delays, and fluctuations in credit scores. The system and algorithm described in this invention are used to achieve early warning of overdue risks. The specific implementation process is as follows: 1. Data Acquisition and Preprocessing: Six types of time-series data were collected from 5,000 personal consumer loan users of the commercial bank through the data acquisition layer. The collection period was 12 months, and the time granularity was divided into daily granularity (core data), weekly granularity (auxiliary data), and monthly granularity (trend data). The specific data types, collection frequency, and preprocessing rules are shown in Table 1 below:
[0041] 2. Time Series Data Fusion Processing: The time series data fusion processing layer divides the data into daily, weekly, and monthly granular slices. Daily slices are used for short-term risk prediction (7-15 days in advance), weekly slices for medium-term risk prediction (15-30 days in advance), and monthly slices for medium- to long-term trend analysis. Formula 1 is used to calculate the fusion weights for each type of time series data. The specific fusion weight allocation is shown in Table 2 below (taking daily slices as an example).
[0042] 3. Algorithm Operation and Risk Classification: The core operation layer of the algorithm loads the fused comprehensive time-series feature dataset, extracts core risk features through an adaptive time-series feature extraction algorithm, and calculates the overdue risk comprehensive index R for each user using Formula 2. Based on the characteristics of personal consumer loan business, a preset risk classification threshold is set, dividing the risk into 5 levels. The specific classification standards and intervention strategies are shown in Table 3 below:
[0043] 4. Iterative Optimization and Implementation Results: The iterative optimization layer collects real-time data on the actual overdue results of 5000 users. Dynamic iterative optimization of the model weights is achieved using Formula 3, with an iteration cycle of daily granularity. After three months of continuous optimization, the model reaches convergence (accuracy fluctuation of the early warning rate for 10 consecutive daily granularity slices <3%, and both false positive and false negative rates <5%). A comparative experiment was conducted between the algorithm of this invention and a traditional early warning algorithm (logistic regression model). The experimental results are shown in Table 4 below:
[0044] Experimental results show that the algorithm of this invention significantly outperforms traditional algorithms in terms of early warning accuracy, false positive rate, and false negative rate in personal consumer loan scenarios. It can achieve an average early warning of 22 days, effectively reducing the non-performing asset ratio and fully meeting the risk management needs of commercial banks' personal consumer loan business. At the same time, with the fairness verification function enabled, the early warning accuracy of different age groups has been monitored for three consecutive months, and the difference has remained stable within 2.1%, with no implicit discrimination. For edge inference using 7 days of time series data for a single application, the average computation time is 37ms, meeting the requirements for second-level loan disbursement. The high-risk early warning is accompanied by a transmission path display: "recent surge in query frequency → increase in debt ratio → weakening repayment ability", which improves the verification efficiency of risk control personnel by 62%.
[0045] Example 2: Example of Early Warning of Overdue Business Loans
[0046] This embodiment is applied to the business loan business of a joint-stock bank. This business targets small and micro enterprises, with loan amounts ranging from 50,000 to 1,000,000 yuan, loan terms of 1 to 5 years, and repayment methods including monthly interest payments and principal repayment at maturity or equal principal and interest payments. The core risk points are fluctuations in the enterprise's operating cash flow, decline in revenue, increase in debt, and decline in industry prosperity. The system and algorithm described in this invention are used to achieve early warning of overdue risks. The specific implementation process is as follows: 1. Data Acquisition and Preprocessing: Six types of time-series data were collected from 300 small and micro enterprises of the bank through the data acquisition layer. The collection period was 24 months, and the time granularity was divided into weekly granularity (core data), monthly granularity (auxiliary data), and quarterly granularity (trend data). The specific data types, collection frequency, and preprocessing rules are shown in Table 5 below:
[0047] 2. Time Series Data Fusion Processing: Through the time series data fusion processing layer, time series slices are divided according to weekly, monthly, and quarterly granularity. Weekly slices are used for short-term risk prediction (15-30 days in advance), monthly slices for medium-term risk prediction (30-60 days in advance), and quarterly slices for medium- to long-term trend analysis. The fusion weights for each type of time series data are calculated using Formula 1. The specific fusion weight allocation is shown in Table 6 below (taking the monthly slice as an example):
[0048] 3. Algorithm Calculation and Risk Classification: The core computation layer of the algorithm loads the fused comprehensive time-series feature dataset. An adaptive time-series feature extraction algorithm extracts core risk features (such as fluctuations in enterprise revenue growth rate, cash flow gaps, changes in debt ratio, and shifts in industry prosperity). Formula 2 is used to calculate the comprehensive overdue risk index R for each enterprise. Based on the characteristics of enterprise operating loan business, a preset risk classification threshold is established, dividing the risk into 5 levels. The specific classification standards and intervention strategies are shown in Table 7 below.
[0049] 4. Iterative Optimization and Implementation Results: The iterative optimization layer collects real-time data on actual overdue results and intervention effects from 300 enterprises. Dynamic iterative optimization of model weights is achieved using Formula 3, with an iteration cycle of monthly granularity. After 6 months of continuous optimization, the model reaches convergence (accuracy fluctuation of granular slice warnings is <3% for 10 consecutive months, with both false positive and false negative rates <5%). A comparative experiment is conducted between the algorithm of this invention and existing enterprise warning algorithms (traditional gradient boosting tree model). The experimental results are shown in Table 8 below:
[0050] In 2023, when a region experienced a sudden adjustment in industry policies, the system automatically triggered the circuit breaker mechanism, switching to the emergency weight template for operational difficulties, and the accuracy of the early warning did not fluctuate significantly. For three micro and small enterprise clients who could not provide data on the prosperity of their specific industries, publicly available industry report data was automatically used as a substitute, with the weight reduced to 2%, and the final risk assessment result was consistent with the actual overdue situation. The intervention effect tracking showed that telephone follow-up measures reduced the risk index by an average of 23%, and the system has automatically increased the intervention priority of such measures.
[0051] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A financial overdue risk intelligent early warning system and algorithm integrating time-series data, characterized by: It comprises seven collaborative modules: a hardware support layer, a data acquisition layer, a time-series data fusion and processing layer, a core algorithm operation layer, a risk warning output layer, an iterative optimization layer, and a data traceability layer. These modules form a closed-loop intelligent management system covering the entire process of "acquisition-preprocessing-fusion-operation-early warning-optimization-traceability." The functions of each module are as follows: Hardware support layer: Adopting a distributed server architecture, it provides server computing power, massive time-series data storage, high-speed network transmission and data security protection support. It has built-in data encryption module and security audit module, supports edge computing node deployment, and can be deployed to the local server of financial institutions to run lightweight inference models to meet the needs of second-level loan disbursement scenarios; it supports cross-border data local storage configuration, and overseas user data is only processed locally. Data Acquisition Layer: This layer collects six types of multi-source time-series data from financial business scenarios, including user / corporate repayment behavior time-series data, account fund flow time-series data, personal / corporate credit fluctuation time-series data, business operation behavior time-series data, macroeconomic environment time-series data, and industry prosperity time-series data. The acquisition process uses AES-256 de-identification and encryption, and the time-series data is divided into slices at daily, weekly, monthly, and quarterly granularities. It also connects to external news and public opinion APIs to automatically generate public opinion risk tags. If an institution lacks industry prosperity data, it supports the use of publicly available industry report data as a substitute. Time series data fusion processing layer: Performs layered cleaning, deduplication, missing value imputation, outlier removal and normalization on the collected multi-source time series data. It completes the deep coupling of multi-dimensional time series data through layered slice fusion logic. The initial fusion weight is allocated according to the time series distance, data stability and risk correlation. Sliding window incremental calculation is used for daily granular data. Only the newly added slice features are updated to avoid full duplication processing. The core computational layer of the algorithm has a built-in self-developed intelligent early warning algorithm for overdue risk. It loads time-series fusion data to complete adaptive time-series feature extraction, risk index quantification and refined risk classification. It has unsupervised adaptive learning capabilities and a built-in fairness verification module. It regularly detects the differences in early warning accuracy among different genders, ages and regions. If the difference is greater than 5%, it automatically triggers weight calibration to avoid implicit discrimination. Risk warning output layer: Generates multi-level warning instructions based on risk classification results, and pushes them simultaneously to financial risk control terminals, mobile risk control APPs, and relevant business managers. It marks the risk level, warning reason, time-series anomalies, risk evolution trend, and targeted intervention suggestions. It supports the visualization of risk transmission path and provides a plain-language interpretation of the weighted results. A new field for tracking intervention effect has been added to record risk control response actions and subsequent risk changes. Iterative optimization layer: Real-time acquisition of actual overdue result data and intervention effect data after risk warning, and reverse input to the core operation layer of the algorithm. The model weight is adaptively updated through dynamic weight iteration formula. At the same time, it is adjusted according to changes in business scenarios and market environment. It has a built-in abnormal event circuit breaker mechanism. When macroeconomic / industry data exceeds the preset threshold, dynamic iteration is automatically paused and switched to emergency static weight template. It supports automatic parameter optimization, matching default values of coefficients according to scene type, eliminating the need for manual parameter tuning; Data traceability layer: Records the collection time, processing process, fusion logic, algorithm operation process, early warning results and intervention effect data of multi-source time series data, forming a complete data traceability chain, supplementing algorithm decision logs, recording the feature contribution ratio of individual early warnings, and meeting the requirements for interpretable supervision.
2. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: It includes three core, original mathematical formulas, which are used for time-series data fusion weight calculation, risk index quantification, and dynamic weight iterative optimization, respectively. The specific formulas are as follows: Formula 1: Formula for calculating the weights of hierarchical fusion of multi-source time series data: In the formula: Let be the fusion weight of the i-th class of time series data under the t-th time series slice, with a value range of [0,1]. The sum of the fusion weights of all classes of time series data is 1. The stability coefficient of the i-th type of time series data is calculated from the variance of the fluctuations of this type of time series data in the t-th slice. For the timeliness weight of the i-th type of time series data, ,in This represents the number of days between the time-series slice and the current time. This is the time-dependent attenuation coefficient, with a value ranging from 0.01 to 0.
05. Let be the risk correlation coefficient of the i-th type of time series data, with a value range of [0,1]; n is the total number of time series data types, n=6; This is the dynamic correction coefficient for time-series slices, with a value range of [0.8, 1.2]. If an institution lacks industry prosperity data, the weight of this category is automatically reduced to 2%. Formula 2: Quantitative Formula for the Comprehensive Index of Financial Overdue Risk: In the formula: This is a comprehensive index for overdue risk, with a value range of [0, 100]. represents the risk feature value of the i-th type of time series data, with a value range of [0, 100]. This is the trend offset correction coefficient, with a value range of [0.1, 0.5]. This represents the offset of short-term temporal behavior abrupt changes. ; C is the cross-dimensional association risk coefficient, with a value range of [0.05, 0.2]; C is the cross-dimensional association risk value, with a value range of [0, 20]. δ is the public opinion risk adjustment factor, with a value range of [0.9, 1.5]. When there is no negative public opinion, δ=1. The final risk index is calculated as R=R×δ. Formula 3: Formula for dynamic weight iteration optimization of the algorithm model: In the formula: The fusion weights are the fusion weights for the i-th type of time series data in the (t+1)-th time series slice after iteration; The fusion weights for the i-th type of time series data in the t-th time slice before iteration; For the iterative learning rate: Where Acc represents the model's early warning accuracy (%). This is the actual overdue risk label value, with overdue payments recorded as 100, non-overdue payments recorded as 0, and partial overdue payments recorded as the corresponding risk percentage value; Predict risk indices for the model; These are the partial derivatives of the weights; The value of is the iterative constraint coefficient, which ranges from [0.9, 1.1].
3. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: The missing value imputation rule for the time-series data fusion processing layer is as follows: Missing data with a missing rate of <10% are imputed using the weighted mean or mean of adjacent time series data. For missing data with a missing rate of 10% ≤ 30%, imputation was performed using the average value of similar users / enterprises during the same period. Slices with a missing rate >30% are removed from the corresponding time series. The outlier removal rule is as follows: based on different time series data types, remove data that exceed a reasonable threshold or deviate from the mean ± 2 standard deviations.
4. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: The risk level of the risk warning output layer is divided into 5 levels: low risk [0,20], low-medium risk [20,40], medium risk [40,60], medium-high risk [60,80], and high risk [80,100]. Different risk levels correspond to different warning intensities and intervention strategies.
5. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: The iteration cycle of the iterative optimization layer can be adjusted according to the business scenario. Daily granularity iteration is used for personal credit scenarios, and monthly granularity iteration is used for business loan scenarios. The iteration reaches the convergence state when the model's warning accuracy fluctuation is less than 3% and the false positive rate and false negative rate are both less than 5% for 10 consecutive time slices.
6. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: The specific contents of the six types of time-series data in the data acquisition layer are as follows: Repayment behavior timeline data: including repayment amount, repayment time, repayment delay duration, number of overdue payments, number of overdue days, and amount of outstanding interest; Account fund flow time sequence data: including income / operating income amount, expenditure amount, balance, large fund transfer out / in records, number of flow interruptions, and upstream and downstream payment records; Credit fluctuation time series data: including personal / corporate credit scores / ratings, number of new overdue records, debt ratio, number of inquiries, guarantee status, and litigation records; Business operation behavior time sequence data: including the number of loan / credit card applications, the number of credit limit adjustments, the number of repayment method changes, the number of violations, and guarantee change records; Macroeconomic environment time series data: including the Consumer Price Index (CPI), GDP growth rate, industrial added value, short-term loan interest rate, unemployment rate, and inflation rate; Industry prosperity time series data: including the corresponding industry prosperity index, industry non-performing loan ratio, and industry revenue growth rate.
7. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 2, characterized in that: The adaptive time-series feature extraction algorithm can automatically identify the core risk features of different financial business scenarios without the need for manual feature pre-setting. It can extract three types of core time-series features: short-term behavioral mutations, medium- and long-term trend shifts, and cross-dimensional correlation risks.
8. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: The system is adaptable to all scenarios of financial business, including personal loans, business loans, consumer finance, credit card business, and micro-loans. Through iterative optimization, it automatically adapts to the different risk characteristics of different scenarios and supports modular deployment and SaaS service model. Small and medium-sized institutions can choose to enable only the core early warning link and temporarily not enable the iterative optimization layer to reduce initial investment costs.
9. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 1, characterized in that: The hardware support layer can adapt to the real-time collection, processing and storage needs of tens of millions of user / enterprise time-series data. The data encryption module adopts the AES-256 encryption standard to ensure data security and user privacy, which complies with the data security specifications of the financial industry. For lightweight inference tasks for a single small loan application, the computation time is <50ms, which meets the requirements of high concurrency and real-time loan disbursement.
10. The intelligent early warning system and algorithm for financial overdue risk that integrates time-series data according to claim 2, characterized in that: The algorithm can accurately predict financial delinquency risks 7-60 days in advance, with 7-30 days in advance for personal consumer loans and credit card scenarios, and 15-60 days in advance for business loan scenarios. The default configuration of the automatic parameter optimization function is: timeliness decay coefficient β=0.03 for consumer loan scenarios and β=0.02 for business loan scenarios, without the need for manual parameter adjustment.