An artificial intelligence-based dynamic credit risk assessment method

Abstract

This invention discloses an artificial intelligence-based dynamic credit risk assessment method. This method continuously integrates multi-source real-time data streams with static historical data, utilizes a deep reinforcement learning-trained agent to generate dynamic risk adjustment factors, and employs an attention mechanism to non-linearly couple these factors with a basic credit score. The final output is a real-time dynamic credit score and an interpretable credit decision. By constructing a closed-loop system from real-time perception and dynamic assessment to interpretable output, this invention achieves minute-level continuous monitoring and proactive early warning of borrower credit risk. This solves the problem of delayed risk capture caused by the static nature of data and models in traditional methods, significantly improving the timeliness, accuracy, and automation of risk management.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and intelligent risk control technology, and in particular to a dynamic credit risk assessment method based on artificial intelligence. Background Technology

[0002] Traditional credit risk assessment primarily relies on borrowers' historical static information, such as historical credit records, personal identification information, and fixed asset status from the central bank's credit report. Credit scoring is then built based on this static data, using scoring card models or simple machine learning models. In recent years, with the development of big data technology, some methods have begun to incorporate broader data dimensions, such as users' consumption records and social networks, to improve the accuracy of assessments. However, these methods are still essentially based on summarizing and analyzing past events, with long data update cycles and relatively rigid model decision-making logic.

[0003] Current technologies lack the capability for real-time, continuous, and dynamic assessment of borrowers' risk profiles. The models and data used in current methods are essentially static or periodically updated in batches. This results in risk assessments failing to reflect timely changes in borrowers' current and recent financial situations, behavioral patterns, and willingness to repay. For example, a borrower may exhibit early warning signs of risk shortly after receiving credit, such as abnormal transactions, cash flow problems, or sudden behavioral changes. However, existing assessment systems based on static snapshots cannot accurately capture and respond to these dynamic signals, causing a significant lag in risk warnings. This fundamental contradiction between the static nature of assessments and the dynamic nature of actual risks exposes financial institutions to a high degree of "time lag risk"—meaning that risk control systems only react passively after risks have already occurred or escalated, potentially leading to credit losses.

[0004] Therefore, in response to the problems mentioned above, this invention proposes a dynamic credit risk assessment method based on artificial intelligence. Summary of the Invention

[0005] To overcome the problem that existing credit risk assessment methods cannot achieve real-time dynamic risk capture and early warning due to the static nature of data and models, this invention proposes an artificial intelligence-based dynamic credit risk assessment method. This method continuously aggregates real-time behavioral data from multiple sources and uses a dynamic intelligent agent that integrates advanced artificial intelligence technologies such as reinforcement learning for real-time analysis, thereby achieving continuous and dynamic assessment of borrower credit risk and effectively improving the timeliness and foresight of risk response.

[0006] The technical solution of this invention is: a dynamic credit risk assessment method based on artificial intelligence, executed by a credit risk assessment system deployed on a cloud server. This system includes a data aggregation module, a dynamic intelligent agent module, a risk assessment engine, a trusted execution environment module, a dynamic baseline generation module, and a counterfactual reasoning and interpretation module. The method includes the following steps: S1, with explicit authorization from the assessment subject, the data aggregation module continuously acquires raw dynamic data and traditional static data related to the assessment subject from multiple authorized data sources; wherein, the raw dynamic data includes at least: Real-time transaction data is obtained from payment networks and banking systems; Online behavioral footprint data is collected through a secure application programming interface (API) and its obfuscated digital behavior logs are collected. Specifically, it includes the distribution of APP usage time and frequency, search keyword clustering analysis results and public service query records. All personally identifiable information is hashed on the local device. When the assessment subject is a micro or small enterprise and is authorized, IoT sensor data can be obtained through compliant sensors installed at its business premises to acquire anonymized data on customer traffic, logistics frequency and equipment energy consumption. Partner ecosystem data, through the Trusted Execution Environment module, based on blockchain and multi-party secure computation or federated learning technology, obtains anonymized performance, consumption or operation data from partner institutions while ensuring data privacy; S2 cleans, standardizes, and performs real-time feature engineering on the original dynamic data stream, extracting multi-dimensional time-series features with temporal correlation. These multi-dimensional time-series features include at least the consumption stability index calculated by a statistical model, the network density and node influence index obtained by analyzing the capital flow network through a graph neural network, the time-series pattern index reflecting the regularity of digital life, and the performance behavior sequence under specific business scenarios. S3, the dynamic baseline generation module uses unsupervised learning technology to perform cluster analysis on the anonymized dynamic characteristics and historical default performance of all assessment subjects, and generates and updates dynamic risk baselines and risk pattern change warnings by industry and region in real time. S4, the dynamic agent module receives multi-dimensional time-series features, traditional static data and corresponding dynamic risk baseline information, and integrates them into a hybrid state vector. The dynamic agent module contains an agent trained by deep reinforcement learning. In a simulated credit environment, the agent optimizes its strategy with the goal of maximizing long-term risk-adjusted returns. Its return function integrates expected losses, gains and customer retention costs. The agent outputs a dynamic risk adjustment factor in real time based on the current hybrid state vector. S5, the risk assessment engine adopts an attention mechanism model, using dynamic risk adjustment factors as query vectors and initial credit scores and their feature subspaces generated based on traditional static data as key and value vectors. By calculating attention weights, it achieves non-linear coupling of dynamic risk information to static credit scores, and finally generates real-time dynamic credit scores. S6. When the real-time dynamic credit score fluctuates significantly or triggers the warning threshold, the counterfactual reasoning and explanation module is activated. This module performs counterfactual reasoning simulation by perturbing the key dynamic feature values ​​of the agent, quantitatively analyzes the marginal contribution of different features to the score change, and generates an interpretable risk driver report. S7 generates dynamic credit decision recommendations based on real-time dynamic credit scores, dynamic risk baseline comparison results, and interpretable reports. These recommendations include differentiated credit limits, flexible interest rate floating schemes, and tiered risk warning signals. The recommendations are then pushed to the credit decision terminal.

[0007] Preferably, in step S2, the analysis of the capital flow network through a graph neural network specifically includes: constructing a time-weighted directed graph with the evaluation subject and its counterparty as nodes and the capital flow relationship as edges; then using a graph attention network model to learn the embedding representation of nodes in the time-series graph structure. This embedding representation simultaneously captures information on transaction size, frequency, path depth, and network structure evolution, thereby comprehensively calculating the network density and node influence indicators.

[0008] The beneficial effects of this invention are: 1. This invention constructs a high-frequency, multi-dimensional, and real-time borrower status perception network by continuously aggregating payment transaction data, IoT sensor signals, online behavioral footprints, and partner ecosystem data fused through privacy computing. This replaces the lagging static data snapshots in traditional risk control with continuous data, thereby transforming risk identification from a summary of past facts to an insight into the current state, solving the problem of lagging information updates.

[0009] 2. The dynamic agent based on deep reinforcement learning introduced in this invention continuously learns how to make decisions under the goal of maximizing long-term risk and return in a simulated credit environment. The agent can continuously output dynamic risk adjustment factors based on real-time inflow feature data and macro risk baselines. This makes risk assessment no longer a one-time or periodic static judgment, but a never-ending dynamic process that interacts with the environment. This changes the problem of traditional model parameters being fixed and unable to proactively adapt to new risk patterns, and realizes intelligent adaptation and proactive defense against sudden and evolving risks.

[0010] 3. The counterfactual reasoning and explanation module integrated in this invention can be automatically activated when the dynamic score fluctuates significantly. By simulating the marginal impact of changes in key features on the scoring results, it quantitatively traces and explains the fundamental driving factors of score changes and outputs them in an understandable report format. This solves the problem that traditional artificial intelligence models are difficult to apply to financial decision-making with high compliance requirements, making every risk warning or limit adjustment have a clear and credible business basis, and greatly enhancing the transparency, credibility and auditability of the risk control system. Attached Figure Description

[0011] Figure 1 The diagram shown is a schematic representation of the system framework of this invention. Figure 2 The diagram shown is a schematic representation of the method flow of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Please see Figure 1 and Figure 2 This invention provides an embodiment: a dynamic credit risk assessment method based on artificial intelligence. This method is implemented through a dynamic credit risk assessment system deployed in the cloud, specifically: S1 first uses the data aggregation module to acquire and merge multi-source data, specifically including: (1) For real-time transaction data, encrypted real-time transaction records are obtained directly from the payment network and banking system through a financial data gateway. The focus is not only on the transaction amount and frequency, but also on changes in counterparties, abnormal transaction times (such as large transfers outside of business hours), and sudden reversals in net fund flows. For example, for micro and small enterprises, the matching degree between daily acquiring transactions and corporate account expenditures is an excellent dynamic health indicator.

[0014] (2) For online behavioral footprint data, with explicit user authorization and in compliance with the Personal Information Protection Law, highly abstract and fuzzy behavioral logs are collected through interaction with a secure SDK deployed on the user's device. All directly identifying personal information is irreversibly anonymized on the user's device using differential privacy or local hashing techniques, and only behavioral pattern tags (such as "frequent use of financial apps at night", "recent clustering of multiple searches for the keyword 'debt restructuring'", "sudden increase in the frequency of provident fund account inquiries") are uploaded. This data can be used in an invisible way to explore potential financial pressure or changes in consumer confidence while ensuring privacy and security.

[0015] (3) For IoT sensor data, compliant IoT sensors are installed at the borrower's business premises, warehouses, or transport vehicles for micro and small enterprises or supply chain finance scenarios. These sensors continuously collect anonymized data, such as daily customer flow heat maps statistically obtained through infrared or Bluetooth beacons, the frequency of goods entering and leaving the warehouse read through smart weighbridges or RFID, and the energy consumption curves of production equipment monitored through smart meters. This data provides real-time and difficult-to-falsify physical operating methods that cannot be reflected in traditional financial statements, and can cross-verify the operating conditions declared by the enterprise. (For example, if a retail store that claims to be busy has low customer flow for a long period of time, it will immediately trigger risk attention).

[0016] (4) For partner ecosystem data, a privacy computing platform built on blockchain and a trusted execution environment (TEE) facilitates data collaboration with partners such as e-commerce platforms, logistics companies, and core supply chain enterprises. Specifically, federated learning algorithms or multi-party secure computation protocols are run within a secure TEE enclave. For example, without exposing their respective users' original order data, they can jointly calculate the borrower's cross-platform monthly consumption stability index with e-commerce platforms, or jointly verify the authenticity of shipment volume with logistics companies. Throughout the process, the original data remains within the domain; only encrypted intermediate computation results or gradients are exchanged, ultimately merging to generate enhanced ecosystem credit features.

[0017] S2, through a risk assessment engine, transforms raw data into high-value signals that can be understood by AI models. Specifically: (1) Use a time series decomposition model (such as STL) to decompose the consumption amount series into a trend term, a seasonal term, and a residual term. The stability index is determined by the volatility (standard deviation) of the residual term and the recent slope of the trend term. A sudden increase in volatility or a sharp decline in the trend slope constitutes a risk signal.

[0018] (2) Transaction records from the past N days are rolled out periodically (hourly) to construct a dynamic and directed weighted graph network of fund flows. Nodes represent transaction entities, edges represent fund transfers, and weights are a combination of transaction amount and frequency. Then, a graph neural network (GAT) is used to learn the representation of this dynamic graph. The GAT model can learn the importance (attention weight) of each node (borrower) in its local network, thus outputting an embedding vector. This vector not only encodes the transaction scale of the node, but also its position in the entire fund network (whether it is a core hub or a peripheral node), the quality of its neighboring nodes (the credit status of the counterparty), and the temporal changes in the network structure. When an entity suddenly establishes contact with multiple unfamiliar high-risk nodes, or is rapidly marginalized from its original stable transaction network, its network embedding vector will drift significantly. This change can reveal its potential business shift or credit deterioration earlier than any single transaction data.

[0019] (3) Analyze online behavior logs using natural language processing and sequence models. For example, model users' behaviors such as logging into the bank's APP to repay loans on time and regularly checking credit reports as a sequence of "good performance habits," and use recurrent neural networks to predict the stability of their subsequent behavior sequences. Disorder in behavior patterns is a precursor to potential changes in performance willingness.

[0020] S3, the dynamic baseline generation module, continuously receives anonymized dynamic characteristics and their final historical default labels (delayed feedback) from all assessed entities. It employs an online clustering algorithm combined with time-series analysis to categorize entities with similar dynamic behavioral patterns. For example, it may automatically identify clusters such as "young blue-collar workers with active nighttime consumption," "small and micro-sized manufacturing enterprises with stable transaction networks," and "groups experiencing sudden changes in online behavior." For each cluster, it calculates its central characteristics and historical default rates in real time, forming a dynamic risk baseline for that subgroup. This makes the assessment of individual borrowers no longer isolated. When assessing a borrower, the system compares their dynamic characteristics with the baseline of their cluster. If their consumption stability begins to deviate significantly from the center of their "healthy blue-collar" cluster, even if their absolute value has not yet reached the global threshold, it will attract the attention of the intelligent agent. This achieves a "relative" risk assessment, providing better adaptability to industry-wide and regional systemic risks (such as the widespread downturn in the catering industry during the pandemic), and avoiding the misjudgment of high-quality entities temporarily facing difficulties due to the overall environment.

[0021] S4 creates a highly realistic credit simulation environment containing millions of historical and virtual borrowers, whose states are represented by the hybrid feature vectors generated in steps S2 and S3. The dynamic agent module's action is to output a continuous risk adjustment factor (e.g., -0.5 to +0.5). The environment adjusts its credit policy for borrowers based on the agent's actions, and after a virtual time span (e.g., the next 12 months), it calculates rewards based on simulated default outcomes, interest income, and simulated customer churn due to overly strict policies.

[0022] Its reward function Rt is the core of the training-oriented approach, balancing risk, reward, and customer relationship: ; in, It is the simulated value at risk (negative value, the smaller the better) of a portfolio over a future period of time. It is net interest income; It's a customer retention score, used to penalize the churn of low-risk customers in the simulation due to excessive risk control. Weighting coefficients. , , It is set by business experts and can be adjusted according to strategy.

[0023] This invention employs a reinforcement learning algorithm with proximal policy optimization to train an agent. After training, the agent is deployed as an online service. In practical applications, it can process the latest state vectors of tens of thousands of borrowers per second and, based on its learned policies, output a dynamic adjustment factor that matches the borrower's current unique risk situation. Thus, the agent learns to make optimal trade-offs in uncertainty. For example, for a small business with a slightly declining network influence index but whose IoT data shows continued busy production, the agent may only slightly increase the risk; while for an individual whose online behavior suddenly shows high anxiety and whose transaction network begins to disintegrate, even with a good historical credit history, the agent may output a strong risk-increasing signal.

[0024] S5 uses the adjustment factor output by the dynamic agent as the "query" and the borrower's static initial credit score and its corresponding original static features (such as age, number of past delinquencies, and total assets) as the "key" and "value." The attention mechanism calculates the correlation between the "query" and each "key" (attention weight). This is equivalent to having the dynamic risk signal "examine" each component of the static score. For example, if the dynamic signal strongly suggests "fluctuations in willingness to repay," the attention weight will be significantly amplified on the static feature of "number of past delinquencies," making the current dynamic score more influenced by this negative history; conversely, if the dynamic signal shows "increased income stability," it may assign a higher weight to the "occupational stability" feature. Finally, by weighted aggregation of the "values," a new, context-aware, real-time dynamic credit score is generated.

[0025] S6: When the dynamic score changes significantly (e.g., a single-day drop of more than 50 points) or triggers an alert, the counterfactual reasoning and explanation module automatically activates, generating an easy-to-understand risk report. This module fixes other inputs to the agent model, performing "if...then..." simulations only on the few dynamic features most critical to the score change (e.g., "Trading network dispersion over the past 7 days" and "specific keyword search index"). (For example: "If the client's trading network dispersion over the past 7 days drops from 0.8 to its historical normal of 0.2, then its dynamic score will rebound by 45 points. Therefore, the increase in network dispersion is the main reason for this score drop (contributing 70%).") The report also includes a comparison chart of the client's current feature values ​​and the dynamic baseline of their group. This allows credit auditors or account managers to see not just a declining score, but a diagnostic report clearly stating, "Please pay close attention to whether there have been any abnormal changes in the client's recent trading circle and verify them," significantly improving the efficiency and accuracy of manual intervention.

[0026] S7, finally, integrates real-time dynamic credit scores, positions relative to the baseline, and interpretable reports to generate specific and differentiated credit decision recommendations through counterfactual reasoning and interpretation modules. The recommendations are no longer binary "approve / reject" decisions, but rather a scheme that includes flexible credit limits (e.g., dynamically adjusting the credit limit from 100,000 to 50,000), floating interest rates (e.g., increasing the base rate by 50 basis points), and warning levels and handling suggestions (e.g., "Level 2 warning, suggesting a gentle reminder-style follow-up call from an AI robot"). This scheme is pushed to the front-end business system in real time via API, driving automated credit strategy execution or assisting human decision-making.

[0027] This invention provides a comparative example: This comparative example includes a simulation experiment to compare the present invention with two existing technologies, wherein... Comparative Example 1 is a traditional static model, a scorecard model based on logistic regression, which only uses static data such as central bank credit information and social security records at the time of application, and the model is updated once a year. Comparative Example 2 is a big data batch update model, using the XGBoost model. In addition to static data, the features also include big data such as transactions and consumption over the past 6 months, but the features are updated in batches monthly, and the model is updated quarterly.

[0028] This experiment uses anonymized historical credit datasets and simulates the generation of a 24-month dynamic behavioral data stream (transactions, app behavior, etc.) and final default labels. The dataset is divided into a training set (first 12 months) and a test set (last 12 months) by time. For this invention, the system performs real-time evaluation during the test period; for the two comparative sets, batch evaluation is performed using historical data up to that point in each fixed evaluation cycle, and the results are shown in the table below.

[0029]

[0030] As shown in the table above, this invention significantly outperforms traditional static models and big data batch update models in all core risk control and business indicators. In terms of the forward-looking nature of risk identification, this invention achieves a qualitative leap, with an average warning lead time of 89.5 days, nearly double that of the best existing methods, providing a crucial time window for risk mitigation. Regarding assessment accuracy, its KS value of 0.583 demonstrates that the model's ability to distinguish between good and bad customers is industry-leading. Simultaneously, this invention exhibits excellent model stability (monthly PSI of only 0.05), indicating that its scoring criteria remain consistent despite dynamic changes, avoiding model drift. In terms of actual business results, this invention significantly reduced the non-performing loan rate in the simulated environment to 2.15% and successfully controlled the churn rate of high-quality customers to 1.8%. This confirms the success of its reinforcement learning agent in balancing risk and return and avoiding false positives. Furthermore, its 100% explainable report coverage completely changes the decision-making mode of traditional models, ensuring that every risk warning has clear action guidelines.

[0031] 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 dynamic credit risk assessment method based on artificial intelligence, characterized in that, This method is executed by a credit risk assessment system deployed on a cloud server, which includes a data aggregation module, a dynamic intelligent agent module, and a risk assessment engine; the method includes the following steps: S1, the data aggregation module continuously acquires dynamic data related to the assessment subject from multiple authorized data sources; S2, clean and standardize the acquired data, and extract multi-dimensional time series features, including but not limited to consumption stability indicators, capital flow network tightness indicators, digital life regularity indicators and performance behavior sequences in specific business scenarios; S3, the dynamic agent module receives multi-dimensional time-series features and merges them with historical static features from traditional credit databases to form a hybrid feature vector. The dynamic agent module includes an agent trained by reinforcement learning, which outputs a dynamic risk adjustment factor based on the current hybrid feature vector and a preset risk adjustment reward function. S4, the risk assessment engine inputs the dynamic risk adjustment factor into a basic credit scoring model, which generates an initial credit score based on historical static features. The risk assessment engine calculates the real-time dynamic credit score by non-linearly coupling the dynamic risk adjustment factor with the initial credit score. S5 generates dynamic credit decision suggestions based on real-time dynamic credit scoring. These suggestions include credit limit, interest rate fluctuation plan, and risk warning level. The decision suggestions and key characteristics are then pushed to the credit decision terminal.

2. The dynamic credit risk assessment method based on artificial intelligence according to claim 1, characterized in that: The data obtained in step S1 includes real-time transaction behavior data, online behavior footprint data, IoT sensor data, and de-identified ecosystem data provided by partners.

3. The dynamic credit risk assessment method based on artificial intelligence according to claim 2, characterized in that, The process of acquiring online behavior footprint data in step S1 specifically includes: with the explicit authorization of the assessment subject, collecting its obfuscated digital behavior logs through a secure application programming interface. These digital behavior logs include the distribution of APP usage time and frequency, search keyword clustering analysis results, and public service query records. All personally identifiable information is hashed on the local device before being uploaded.

4. The dynamic credit risk assessment method based on artificial intelligence according to claim 3, characterized in that, The calculation method of the capital flow network density index in step S2 includes: constructing a dynamic capital flow graph network based on the counterparty data of the assessment subject over a period of time, then calculating the node centrality, clustering coefficient and cross-community connection strength of the subject in the graph network, using a graph neural network to perform embedding representation learning on the capital flow graph network, and using the learned embedding vector as the capital flow network density index.

5. The dynamic credit risk assessment method based on artificial intelligence according to claim 4, characterized in that, The training method for the agent in step S3 specifically includes: constructing a simulated credit environment, using the mixed feature vector of the assessment subject as the state, the risk adjustment action as the action space, defining a return function that comprehensively considers the expected loss rate of the adjusted credit asset, the return on investment, and the cost of customer churn due to misjudgment, and then using a proximal policy optimization algorithm to continuously train the agent so that it learns to output the optimal dynamic risk adjustment factor under the goal of maximizing long-term risk and return.

6. The dynamic credit risk assessment method based on artificial intelligence according to claim 5, characterized in that, The nonlinear coupling in step S4 is specifically as follows: using an attention mechanism model, the dynamic risk adjustment factor is used as the query vector, the initial credit score and its corresponding historical static feature subspace are used as the key and value vectors, and then the influence strength of the dynamic risk adjustment factor on different components of the initial credit score is determined by the attention weight calculation, thereby generating a weighted aggregated real-time dynamic credit score.

7. The dynamic credit risk assessment method based on artificial intelligence according to claim 6, characterized in that: The method also includes performing privacy-preserving computations when using a trusted execution environment module to retrieve data from a partner data source.

8. The dynamic credit risk assessment method based on artificial intelligence according to claim 7, characterized in that: The method also includes using a dynamic baseline generation module to establish dynamic risk baselines by industry and region. This dynamic baseline generation module continuously analyzes the anonymized feature distribution and default performance of all assessment subjects, uses unsupervised learning clustering methods to identify risk pattern changes, and compares the real-time dynamic credit score with the dynamic risk baseline of the corresponding group to output a relative risk position report.

9. The dynamic credit risk assessment method based on artificial intelligence according to claim 8, characterized in that: The method also includes a counterfactual reasoning and explanation module, which generates an interpretable report based on counterfactual reasoning technology when the real-time dynamic credit score changes significantly.

10. The dynamic credit risk assessment method based on artificial intelligence according to claim 9, characterized in that, The IoT sensor data in step S1 specifically refers to: when the assessment subject is a micro or small enterprise and is authorized, anonymized data obtained through compliant sensors installed at its business premises, including but not limited to the daily average customer flow trend, the frequency of warehouse logistics inbound and outbound, and the energy consumption pattern of production equipment. This data is used to cross-verify its statement on operational stability.

Images