Self-adaptive dynamic adjustment method and system of intelligent prediction model

Abstract

The invention provides a self-adaptive dynamic adjustment method and system for an intelligent prediction model, which are applied to the field of supply chain financial risk control, and the method and system comprise the steps: collecting supply chain data to determine an initial deviation, calling a deviation quantification model to generate a deviation score when a threshold value is exceeded, and cooperating with a machine learning model to generate a preliminary strategy; and obtaining a final strategy through risk assessment and correction, and automatically adjusting the deviation quantification model. According to the scheme, the real-time performance and accuracy of credit decision can be improved, and the financial risk of the supply chain can be effectively controlled.

Description

Technical Field

[0001] This invention relates to the field of supply chain finance risk control technology, and in particular to an adaptive dynamic adjustment method and system for an intelligent prediction model. Background Technology

[0002] Intelligent predictive models, especially machine learning-based models, have been widely applied in various technological fields such as industry, finance, and healthcare to predict the future state of complex systems and provide decision support. However, when applying these general models to specific business scenarios, especially dynamic and high-risk scenarios, the following technical bottlenecks are commonly encountered: First, existing intelligent forecasting models often lack sufficient adaptability after deployment. Many forecasting models have fixed internal parameters after training. When external environments, such as market conditions, logistics efficiency, or raw material prices in supply chain finance, change dynamically, the model cannot quickly adapt, leading to a gradual increase in the deviation between forecast results and actual business performance—a technical problem known as "model drift." This makes decision-making based on the model sluggish.

[0003] Secondly, existing technologies for assessing prediction deviations are rather superficial. When a predicted value deviates from the actual value, the system typically only monitors the magnitude of the deviation. However, the system lacks a technical means to combine the numerical value of the deviation with the specific business context in which it occurred, in order to deeply quantify the true level of risk hidden behind the deviation. A deviation within the normal fluctuation range has drastically different technical implications and potential risk levels than a deviation of the same magnitude occurring at a critical business node, and existing technologies struggle to effectively distinguish between these.

[0004] Furthermore, automated decision-making logic based on predictive models is often quite isolated. When faced with multi-source, complex, or even conflicting input signals, such as when historical trends contradict real-time market feedback, the system lacks a technical mechanism that can collaboratively invoke multiple functional modules for intelligent trade-offs, making it difficult to generate a globally optimal decision instruction and resulting in rigid decision-making logic.

[0005] Finally, the iterative optimization process of existing predictive models heavily relies on human intervention. Traditionally, the optimization and iteration of predictive models is an offline process requiring the participation of data scientists and domain experts, which is costly and inefficient. Systems generally lack a closed-loop feedback mechanism that can automatically optimize and adjust their internal parameters based on the final business results after the execution of decision instructions; in other words, they lack the ability to self-evolve.

[0006] The aforementioned technical issues are particularly prominent in the field of supply chain finance risk control. As a highly dynamic and complex system, the risk status of a supply chain is constantly influenced by multiple factors. Applying intelligent predictive models to this system, if the aforementioned technical shortcomings—slow response, superficial risk perception, rigid decision-making, and lack of self-learning ability—cannot be addressed, will directly lead to delayed adjustments in credit strategies, potentially resulting in missed market opportunities or the accumulation of potential financial risks.

[0007] Therefore, there is an urgent need in this field for an adaptive dynamic adjustment technology for intelligent prediction models to solve the technical problems of poor adaptability, insufficient bias assessment, isolated decision-making logic, and lack of self-optimization closed loop in existing prediction models.

[0008] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides an adaptive dynamic adjustment method and system for intelligent prediction models, aiming to solve the following technical problems existing in the current supply chain finance risk control and credit decision-making systems: slow response, inability to adapt to dynamic changes in the supply chain, overly simplistic assessment of business performance deviations, rigid decision-making system logic, difficulty in handling complex signals, and lack of adaptive learning capabilities. By constructing a complete and adaptive intelligent decision-making closed loop, the invention achieves real-time and accurate credit decisions, deeply and dynamically quantifies deviation risks, intelligently and collaboratively processes multi-source, complex, and even conflicting business signals, and enables the credit decision-making system to have adaptive learning capabilities.

[0010] This invention provides an adaptive dynamic adjustment method for an intelligent prediction model, applied in the field of supply chain finance, comprising the following steps executed by a computer: By collecting supplier credit data and downstream enterprise repayment records in the supply chain through data interfaces, real-time business performance indicators are obtained, and the initial deviation value between the real-time business performance indicators and the prediction results of the intelligent prediction model is determined. When the initial deviation value exceeds the preset threshold, the deviation quantification model is invoked to process the business scenario data associated with real-time business performance indicators in order to generate a deviation severity score that characterizes the severity of the deviation. Based on the severity score of the bias, historical pattern characteristics, and market feedback information, at least two machine learning models with different functions are invoked in a coordinated manner to generate an initial credit strategy. Before implementing the initial credit strategy, a risk assessment is conducted using a risk simulation module, and the initial credit strategy is cyclically revised based on the assessment results to obtain the final credit strategy. Based on the actual business performance data during the evaluation period after the final credit strategy is implemented, the internal parameters in the deviation quantification model are automatically adjusted to improve the accuracy of subsequent deviation quantification.

[0011] In some optional embodiments, the real-time business performance indicator is a composite supply chain health index, which is calculated by weighted fusion algorithm based on the historical fulfillment rate and credit rating data in supplier credit data, as well as the payment period and overdue amount ratio data in the repayment records of downstream enterprises.

[0012] In some optional embodiments, the business scenario data includes at least user purchase frequency data, product category preference data, and historical return rate data associated with deviations from real-time business performance metrics.

[0013] In some optional embodiments, prior to the collaborative invocation of at least two functionally distinct machine learning models, the following steps are also included: Historical pattern features are extracted from historical time series data of bias severity scores using a sliding window algorithm. These historical pattern features include moving average, variance, and trend slope. The market feedback information is processed, including sentiment analysis scores extracted from user comment data through natural language processing models, and transaction volume change rate data.

[0014] In some optional embodiments, at least two functionally distinct machine learning models are collaboratively invoked, specifically: Historical pattern features and market feedback information are input into a pre-trained random forest model to determine whether there are policy conflict points. When a policy conflict is identified, the severity score of the bias is used as input to invoke at least two machine learning models with different functionalities.

[0015] In some alternative embodiments, at least two functionally distinct machine learning models are invoked in a collaborative manner, one of which is a neural network model for optimizing the credit line of the core enterprise, and the other is a support vector regression model for correcting the trade finance interest rate of upstream and downstream enterprises.

[0016] In some optional embodiments, the risk simulation module performs a risk assessment on the initial credit strategy, specifically as follows: The expected default rate index under the initial credit strategy was calculated using the Monte Carlo simulation method.

[0017] In some optional embodiments, the initial credit strategy is cyclically revised based on the evaluation results, specifically as follows: If the expected default rate index is higher than the risk tolerance threshold, the expected default rate index will be used as new market feedback information to re-invoke at least two machine learning models with different functions in order to revise the initial credit strategy.

[0018] In some alternative embodiments, actual business performance data includes actual supply chain cash flow efficiency data and bad debt rate data.

[0019] In some optional embodiments, the intrinsic parameters in the deviation quantification model are automatically adjusted, specifically as follows: Based on supply chain cash flow efficiency data and bad debt rate data, a comprehensive score for optimization effect is calculated. When the overall optimization score is lower than the performance benchmark, the overall optimization score is used as a feedback signal, and the weight parameters in the deviation quantization model are iteratively updated through the gradient boosting optimization algorithm. The weight parameters are the internal parameters.

[0020] In some optional embodiments, the deviation quantification model is a weighted scoring model, which generates a deviation severity score by weighting and summing the initial deviation value with multi-dimensional data in the business scenario data.

[0021] In some optional embodiments, the intrinsic parameters in the deviation quantification model are automatically adjusted, specifically as follows: Automatically adjust the weight coefficients corresponding to multi-dimensional data in the weighted scoring model.

[0022] In some optional embodiments, the termination condition for the cyclic feedback correction is: the expected default rate index is lower than the risk tolerance threshold, or the number of cyclic corrections reaches a preset maximum number.

[0023] In some optional embodiments, prior to determining the initial deviation value, the following steps are also included: Standardized preprocessing is performed on the collected supplier credit data and downstream enterprise repayment records to eliminate dimensional differences between different data sources.

[0024] In some alternative embodiments, preliminary credit strategies generated by at least two functionally distinct machine learning models are invoked in a coordinated manner, including credit line adjustment instructions and financing rate correction instructions.

[0025] This invention provides an adaptive dynamic adjustment system for an intelligent prediction model, applied in the field of supply chain finance, comprising: The data acquisition and quantification module is configured to: acquire supplier credit data and downstream enterprise repayment records in the supply chain through the data interface to obtain real-time business performance indicators, and determine the initial deviation value between the real-time business performance indicators and the prediction results; and when the initial deviation value exceeds a preset threshold, call the deviation quantification model to process the business scenario data associated with the real-time business performance indicators to generate a deviation severity score that characterizes the severity of the deviation. The collaborative decision-making module is configured to: collaboratively invoke at least two machine learning models with different functions based on the severity score of the deviation, historical pattern characteristics, and market feedback information to generate an initial credit strategy; The closed-loop feedback control module is configured to: conduct a risk assessment of the initial credit strategy through the risk simulation module before execution, and perform cyclical feedback correction of the initial credit strategy based on the assessment results to obtain the final credit strategy; and automatically adjust the internal parameters in the deviation quantification model based on the actual business performance data within the assessment period after the final credit strategy is executed, so as to improve the accuracy of subsequent deviation quantification by the deviation quantification model.

[0026] In some optional embodiments, the collaborative decision-making module is also configured to: Extract historical pattern features from historical time-series data of bias severity scores; and The market feedback information is processed, including sentiment analysis scores extracted from user comment data through natural language processing models, and transaction volume change rate data.

[0027] In some optional embodiments, the closed-loop feedback control module is configured to perform cyclical feedback corrections to the initial credit strategy based on the evaluation results, specifically through the following methods: If the expected default rate index generated by the risk simulation module is higher than the risk tolerance threshold, the expected default rate index will be used as new market feedback information and fed back to the collaborative decision-making module to revise the initial credit strategy.

[0028] In some optional embodiments, the closed-loop feedback control module is configured to automatically adjust the internal parameters in the deviation quantization model, specifically by means of the following: The overall score for optimization effectiveness is calculated based on actual business performance data. When the overall score of the optimization effect is lower than the performance benchmark, the weight parameters in the deviation quantization model are iteratively updated using the gradient boosting optimization algorithm. The weight parameters are the internal parameters.

[0029] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention.

[0030] The adaptive dynamic adjustment method and system for an intelligent prediction model of the present invention has the following beneficial effects: This invention employs a weighted scoring model to quantify deviations and incorporates multi-dimensional business scenario data, enabling more accurate assessment of supply chain finance credit risk and improving the accuracy of risk identification. This method combines initial deviation values ​​with factors such as user purchase frequency, product category preferences, and historical return rates, achieving a deep quantification of deviation risk. This allows the system to more effectively identify potential risks, thereby enhancing the reliability of credit decisions. Compared to traditional risk control methods that rely solely on a single deviation value, this invention provides a more comprehensive and detailed risk assessment, helping to reduce bad debt rates, optimize capital allocation, and improve the efficiency and security of supply chain finance services. Attached Figure Description

[0031] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0032] Figure 1 This is a flowchart of an adaptive dynamic adjustment method for an intelligent prediction model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an adaptive dynamic adjustment system for an intelligent prediction model according to an embodiment of the present invention. Detailed Implementation

[0033] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0034] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0035] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined. Therefore, the actual execution order may change depending on the specific circumstances.

[0036] In supply chain finance, risk control and credit decisions rely on the integrated analysis of data from all stages of the supply chain. Deviation quantification models identify potential risks by quantifying the discrepancy between actual business performance and expectations. Machine learning models learn patterns from historical data to predict future risks and formulate corresponding credit strategies. Closed-loop feedback control technology monitors the actual effects of strategy execution and feeds this information back to the model, enabling continuous optimization and adaptive adjustment. Combining deviation quantification models with closed-loop feedback control technology allows for dynamic adjustments to credit strategies based on real-time changes in the supply chain, improving the effectiveness of risk control and the accuracy of credit decisions.

[0037] like Figure 1 As shown, this embodiment of the invention provides an adaptive dynamic adjustment method for an intelligent prediction model, applied in the field of supply chain finance. The method includes the following steps: Step S100: Data Acquisition and Initial Deviation Determination. The system collects data reflecting the supply chain's operational status from multiple data sources through a data interface and determines the initial deviation value based on this data. In one embodiment, the data interface can be an application programming interface (API), such as collecting data from an enterprise resource planning (ERP) system or a supply chain management (SCM) system. Data sources may include, but are not limited to, supplier operational data, downstream customer financial data, and logistics and transportation data. The data reflects the supply chain's operational status, such as its overall efficiency, stability, or risk level. The system can employ a predictive model, such as a regression model, to predict current expected values ​​based on historical data. The initial deviation value represents the difference between the actual collected data and the predicted value. In some optional implementations, the data interface can also be a message queue, a database connection, etc.

[0038] Step S200: Deviation Quantification. When the initial deviation value determined in step S100 exceeds a preset threshold, the system invokes the deviation quantification model to process business scenario data associated with real-time business performance indicators to generate a deviation severity score characterizing the severity of the deviation. The deviation quantification model is a model used to assess the risk level of a deviation. Its inputs are the initial deviation value and various contextual information related to the deviation. Business scenario data refers to various factors that can affect the deviation risk assessment, such as macroeconomic data, industry dynamics, and the company's own operational status. The deviation severity score is the output of the deviation quantification model, reflecting the actual risk level of the deviation in the current business scenario. In some other optional implementations, the deviation quantification model can be a rule engine, a decision tree model, or a neural network model.

[0039] Step S300: Preliminary Credit Strategy Generation. Based on the deviation severity score, historical pattern features, and market feedback information generated in Step S200, the system collaboratively invokes at least two machine learning models with different functionalities to generate a preliminary credit strategy. Historical pattern features are characteristics reflecting the changing patterns of deviations, obtained by the system through analysis of historical data. Market feedback information is information from the market that reflects market sentiment and trends, such as text data collected from publicly available online sources and social media data. This information, along with the deviation severity score, serves as input to the machine learning model for generating the preliminary credit strategy. The preliminary credit strategy is a preliminary credit decision plan generated by the system based on the current situation, such as adjusting credit limits or interest rates. In some optional implementations, historical pattern features may include statistical features, time series features, or spectral features. Market feedback information can be obtained through web scraping or third-party APIs.

[0040] Step S400: Looping Feedback Correction. Before executing the preliminary credit strategy generated in step S300, the system performs a risk assessment on the preliminary credit strategy using a risk simulation module, and performs looping feedback correction based on the assessment results to obtain the final credit strategy. The risk simulation module is a module used to simulate the execution effect of the credit strategy. Through risk simulation, the system can assess the potential risks of the strategy before actual execution. If the risk assessment results do not meet expectations, the system will adjust the preliminary credit strategy based on the assessment results, such as adjusting strategy parameters or changing the strategy direction. Looping feedback correction refers to the system repeatedly performing risk simulation and strategy adjustment until the risk assessment results of the strategy meet expectations. In some optional implementations, the risk simulation module can use Monte Carlo methods, stress testing methods, or scenario analysis methods for risk assessment. The termination condition for looping feedback correction can be reaching a preset upper limit for the number of loops, or the risk assessment results meeting a preset threshold.

[0041] Step S500: Parameter Self-Optimization. Based on the actual business performance data within the evaluation period after the final credit strategy implementation, the system automatically adjusts the internal parameters in the deviation quantification model to improve the accuracy of subsequent deviation quantification. The evaluation period refers to the timeframe during which the system evaluates the effectiveness of the credit strategy implementation. Actual business performance data refers to data collected by the system within the evaluation period that reflects the actual effectiveness of the credit strategy, such as capital turnover efficiency and bad debt rate. By analyzing this data, the system can evaluate the effectiveness of the credit strategy and adjust the internal parameters of the deviation quantification model based on the evaluation results. Parameter self-optimization means that the system automatically adjusts the parameters without manual intervention. Through parameter self-optimization, the system can continuously improve the accuracy of the deviation quantification model, thereby improving the quality of credit decisions. In some optional implementations, the self-optimization algorithm can be a gradient descent algorithm, a genetic algorithm, or a particle swarm optimization algorithm. The internal parameters of the deviation quantification model can be weight coefficients, thresholds, or model structures.

[0042] Steps S100 to S500 describe an adaptive dynamic adjustment method for an intelligent prediction model. First, through data acquisition and initial deviation determination, the system can perceive the operational status of the supply chain in real time. Then, through deviation quantification, the system can accurately assess the risk level of the deviation. Next, through preliminary credit strategy generation, the system can generate preliminary credit decision plans based on the risk level. Subsequently, through cyclical feedback correction, the system can perform risk simulation and adjustment of the credit strategy to ensure its effectiveness. Finally, through parameter self-optimization, the system can continuously improve the accuracy of the deviation quantification model, thereby improving the quality of credit decisions. These technical features work together to construct a complete, adaptive intelligent decision-making closed loop, solving the technical problems of slow response, simplistic deviation assessment, rigid decision logic, and lack of self-learning ability in existing supply chain finance risk control and credit decision-making systems, achieving real-time, accurate, intelligent, and adaptive credit decisions.

[0043] Through the above solution, this embodiment can achieve accurate and rapid response to dynamic supply chain risks, transforming the traditional post-event analysis into pre-event, context-based risk prediction and proactive correction. It can eliminate manual intervention, achieve self-learning and evolution based on real business results, and optimize the overall capital allocation efficiency of the supply chain.

[0044] In one specific implementation, based on the above embodiments, firstly, the real-time business performance indicator is defined as a composite supply chain health index. Specifically, this index is calculated by integrating historical fulfillment rates and credit rating data from supplier credit data, as well as payment terms and overdue amount percentages from downstream enterprise repayment records, using a weighted fusion algorithm. Then, appropriate quantification methods are selected for different data types. For example, historical fulfillment rates are directly expressed as percentage values; credit rating data is first mapped to numerical scores (e.g., AAA corresponds to 100 points, AA corresponds to 90 points) before being included in the calculation. Next, for downstream enterprise repayment records, payment terms are directly expressed as average days, and overdue amount percentages are converted to percentages. The weighted fusion algorithm uses the following formula:

[0045] in: Composite supply chain health index. Supplier credit rating numerical score. Supplier historical fulfillment rate. The average payment period (in days) for downstream enterprises needs to be normalized. : Percentage of overdue amounts for downstream enterprises (%). , , , Weighting coefficients can be adjusted according to the actual business scenario, for example, set to 0.3, 0.3, 0.2, and 0.2 respectively, and satisfy the following conditions: .

[0046] In other alternative implementations, the weighted fusion algorithm can use linear regression models, neural network models, or other machine learning models to automatically learn the weight coefficients of each indicator through training data. Furthermore, more data dimensions can be introduced, such as logistics timeliness data and market demand forecast data, to more comprehensively reflect the health of the supply chain.

[0047] Through the above approach, this embodiment can more accurately quantify the overall health of the supply chain, providing a more reliable basis for subsequent risk assessment and credit decisions. Compared with using only a single indicator or simple summation, this embodiment can more comprehensively and objectively reflect the true operational status of the supply chain.

[0048] In one specific implementation, based on the above embodiments, firstly, in the deviation context quantization step of the above steps, the system obtains the initial deviation value. Next, the system further acquires business scenario data related to this deviation. Specifically, the system acquires this data in the following ways: User purchase frequency data related to the deviation entity (e.g., a specific supplier or downstream enterprise) is obtained through API calls on the e-commerce platform. This API returns data such as the number of purchases, the number of buyers, and the repurchase rate of products associated with the deviation entity within a specified time period (e.g., the past 30 days).

[0049] Product category preference data is obtained from the inventory management system. The system queries historical sales data for the product categories involved in the deviation, calculates the sales percentage of each category, and uses this to reflect the user's product category preference. For example, if the deviation mainly involves category A products, and category A products account for 80% of high-profit products, it indicates that the user has a high preference for high-profit products.

[0050] Historical return rate data is obtained through the API interface of the after-sales service system. The system queries the historical return records of the products involved in the deviation and calculates the return rate (returned quantity / total sales quantity).

[0051] Then, calculate the severity score of the deviation in the above steps. In this process, this data will be incorporated into the calculation. In other alternative implementations, user purchase frequency data can be obtained by crawling user review data from e-commerce platforms and analyzing the frequency of mentions of purchase behavior in the reviews; product category preference data can be inferred by analyzing users' historical browsing records; and historical return rate data can be estimated by analyzing the number of posts users make on social media about product returns.

[0052] Through the above approach, this embodiment can more comprehensively consider business scenario information related to the subject of the deviation, thereby more accurately quantifying the severity of the deviation and providing a more reliable basis for subsequent credit decisions.

[0053] In one specific implementation, based on the above embodiments, the system first extracts historical pattern features from the historical time-series data of deviation severity scores. Specifically, the system employs a sliding window algorithm, sliding across the historical data with a preset window size (e.g., 14 days, 30 days, or 60 days), and calculating statistical features for the data within each window. These statistical features include, but are not limited to: the average deviation severity score within the window period, reflecting the overall level of historical deviation; the variance of the deviation severity score within the window period, reflecting the degree of fluctuation in historical deviation; and the trend slope obtained by fitting the time-series data of deviation severity scores within the window period through linear regression, reflecting the changing trend of historical deviation. In other optional implementations, other time-series analysis methods, such as exponential smoothing, can also be used to extract historical pattern features.

[0054] The system then processes market feedback information. Specifically, it collects supply chain-related market feedback data from multiple sources, including but not limited to social media platforms, e-commerce platforms, and news websites. For user comment data collected from social media platforms, the system uses a pre-trained natural language processing model to perform sentiment analysis, obtaining a sentiment score that represents the user's overall emotional inclination towards supply chain-related products. In some implementations, Transformer models such as BERT, RoBERTa, or XLNet can be used for sentiment analysis. Additionally, the system obtains transaction volume change rate data from the transaction system, reflecting changes in market demand for supply chain-related products.

[0055] Through the above approach, this embodiment can extract more comprehensive historical pattern features and market feedback information, providing more accurate input for subsequent credit decisions and thus improving the quality of credit decisions.

[0056] In one specific implementation, based on the above embodiments, firstly, the extracted historical pattern feature vector and the acquired market feedback information vector are concatenated. Specifically, the historical pattern feature vector includes three components: moving average, variance, and trend slope; the market feedback information vector includes two components: sentiment analysis score and transaction volume change rate. The final concatenated feature vector has a dimension of 5. Then, this 5-dimensional feature vector is input into a pre-trained random forest model. This random forest model consists of an ensemble model containing 100 decision trees, with each tree having a maximum depth limit of 10 layers. The model outputs a probability value between 0 and 1, which represents the likelihood of a strategy conflict in the current credit decision scenario. Next, this probability value is compared with a preset conflict threshold (e.g., 0.75). If the probability value is greater than the conflict threshold, a strategy conflict is determined to exist; otherwise, it is determined not to exist. Only when a strategy conflict is determined to exist will the system use the deviation severity score as input to collaboratively call the neural network model for credit limit optimization and the support vector regression model for interest rate pricing. In other alternative implementations, besides the random forest model, ensemble learning models such as gradient boosting decision trees (GBDT) or extreme gradient boosting (XGBoost) can be used to identify policy conflict points, or other machine learning models such as logistic regression or support vector machines can be used to achieve similar results.

[0057] Through the above solution, this embodiment can predict potential strategic conflict risks in the credit decision-making process based on historical pattern characteristics and market feedback information, thereby avoiding direct decision-making in complex or contradictory business scenarios and improving the robustness and adaptability of the credit strategy.

[0058] In one specific implementation, based on the above embodiments, firstly, the neural network model is configured to receive a score indicating the severity of the bias. The system takes into account historical credit limits, financial statement data, and supply chain transaction data as inputs. Specifically, the neural network model employs a deep learning architecture with multiple hidden layers, such as three layers with 128, 64, and 32 neurons respectively. Each neuron uses the ReLU activation function. The model then calculates a suggested percentage adjustment to the enterprise's credit limit through forward propagation; for example, +2% suggests a 2% increase in the credit limit. During training, the neural network model uses historical credit data and actual bad debt rate data, with the optimization objective being to minimize the bad debt rate while maximizing capital utilization efficiency.

[0059] Next, the Support Vector Regression (SVR) model was configured to receive bias severity scores. The inputs include the current market benchmark interest rate, credit rating data of upstream and downstream enterprises, and historical interest rate data. Specifically, this SVR model uses a radial basis function (RBF) as the kernel function and employs a grid search algorithm to optimize model parameters, such as the penalty factor C and kernel function coefficients. The model then outputs a revised trade finance rate for upstream and downstream companies; for example, +0.05% suggests a 0.05 percentage point increase above the benchmark rate. This SVR model is trained using historical interest rate data and actual repayment data, aiming to predict the optimal interest rate level to balance risk and return.

[0060] In some other alternative implementations, the neural network model can be replaced with other types of neural networks, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), to handle time-series data. The support vector regression model can also be replaced with other types of regression models, such as gradient boosting regression trees or random forest regression models.

[0061] Through the above approach, this embodiment can optimize credit limits and interest rates by using different machine learning models for enterprises and upstream and downstream enterprises respectively, thereby achieving more refined risk control and more efficient capital allocation.

[0062] In one specific implementation, based on the above embodiments, the risk simulation module first obtains a preliminary credit strategy, which may include, but is not limited to, suggested credit line adjustment percentages and financing interest rate correction instructions. Specifically, the module uses the Monte Carlo method to perform stochastic simulations of supply chain finance activities, where each simulation represents various scenarios that may occur in the supply chain under a given credit strategy. Each simulation considers multiple uncertainties, such as raw material price fluctuations, changes in market demand, logistics delay probabilities, and the credit risk of downstream enterprises. These uncertainties are modeled using probability distributions; for example, raw material price fluctuations follow a normal distribution, and logistics delay times follow an exponential distribution.

[0063] Then, the module generates random numbers for each simulation, based on a set number of simulations (e.g., 10,000). These random numbers are used to sample the parameter values ​​required for the simulation from a predefined probability distribution. Based on these parameter values, the module simulates the cash flow process in the supply chain, including upstream suppliers' payment collection, the company's production and operation, and downstream companies' sales revenue collection. During the simulation, preliminary credit strategies are taken into account as factors influencing corporate behavior and cash flow.

[0064] Next, after completing all simulations, the module compiles statistics on default events that occurred in all simulation results. The definition of a default event can be adjusted according to the specific business scenario; for example, it can be defined as a downstream company failing to repay the loan in full within the specified period, or a supplier failing to deliver qualified materials on time.

[0065] Finally, the module divides the number of default events by the total number of simulations to obtain the expected default rate index. This index reflects the likelihood of supply chain default under the current credit strategy. A higher expected default rate index indicates a higher risk associated with the credit strategy.

[0066] In other alternative implementations, the risk simulation module can employ a backtesting approach based on historical data to assess the risk of initial credit strategies. The accuracy of risk assessment can also be improved by constructing more refined simulation models that consider a wider range of factors. Furthermore, the probability distribution parameters in the Monte Carlo simulation can be dynamically adjusted using real-time data to better reflect changes in the current market environment.

[0067] Through the above approach, this embodiment can more precisely assess the risks of initial credit strategies, thereby providing a more accurate basis for subsequent strategy adjustments and optimizations. Compared to risk assessment methods that rely solely on expert experience or simple rules, this embodiment can more comprehensively consider various uncertainties, thus more effectively identifying potential risks.

[0068] In one specific implementation, based on the above embodiments, if the expected default rate index is higher than the risk tolerance threshold, the system first converts the expected default rate index, for example, 3.5%, into a standardized risk level score, for example, mapped to between 0 and 1 using the Sigmoid function, resulting in a risk score of 0.8. Specifically, the system maintains a market feedback information pool containing risk signals from different channels, such as counts of negative sentiment events extracted from public online information sources and downward trends in macroeconomic data. Then, the aforementioned risk score of 0.8 is used as a new risk signal and fused with other existing risk signals in the information pool. The fusion method can be a simple weighted average or a more complex Bayesian network or deep learning model. Next, the fused market feedback information vector, along with the historical pattern feature vector generated in the steps of the above embodiments, is re-inputted into the collaborative decision-making module. The random forest model within the collaborative decision-making module re-evaluates whether there are policy conflict points based on the new input information. If a conflict exists, the neural network model and support vector regression model regenerate credit limit adjustment instructions and financing interest rate correction instructions based on the updated input information. Compared to the initial credit strategy, the generated credit strategy is more conservative to reduce the expected default rate. In other optional implementations, the expected default rate index can also be converted into a risk report in natural language, such as "High default risk, mainly due to…", and this report can be passed to the collaborative decision-making module to help the model better understand the source and nature of the risk.

[0069] Through the above-described solution, this embodiment enables the credit decision-making system to adjust its credit strategy in a timely manner when facing high risks, thereby reducing potential losses. Compared with existing technologies, the advantages of this embodiment are that it can more effectively utilize the results of risk simulation to achieve dynamic adjustment and optimization of credit strategies.

[0070] In one specific implementation, building upon the above embodiments, the system first extracts actual data from the data storage system to evaluate the final business performance. Specifically, the system can be configured to access data from the enterprise's financial system, order management system, and customer relationship management (CRM) system. Data extracted from the financial system includes the actual collection cycles of all upstream and downstream enterprises participating in supply chain finance business within the evaluation period, as well as the total amount of bad debts generated. The collection cycle is defined as the time span from order confirmation to actual receipt of funds. Then, the system obtains the total amount of all relevant orders within the period and the average order processing time from the order management system. From the CRM system, the system can access customer satisfaction scores to assess the potential impact of supply chain finance services on customer experience. Next, the system integrates and cleans the data from these different sources to ensure data consistency and accuracy. For example, data alignment can be achieved through a unified timestamp format and enterprise identifier. In other optional implementations, the system can use different data sources, such as logistics information platforms, to obtain more accurate logistics timeliness data, or use reports from industry risk assessment agencies to supplement information on bad debt risk assessment.

[0071] Through the above approach, this embodiment can provide more comprehensive and accurate data on actual business performance, providing a solid data foundation for the subsequent self-optimization process.

[0072] In one specific implementation, based on the above embodiments, firstly, the data acquisition module extracts supply chain cash flow efficiency data and bad debt rate data corresponding to all executed credit strategies within the evaluation period (e.g., one month) from the data warehouse. Specifically, supply chain cash flow efficiency is quantified by calculating the average order-to-payment cycle in days, which comes from the financial module of the Enterprise Resource Planning (ERP) system; the bad debt rate is quantified by statistically analyzing the proportion of accounts receivable overdue for more than 90 days within the period to the total accounts receivable, which also comes from the financial module of the ERP system. Then, the system uses the formula:

[0073] Calculate the overall score of the optimization effect, where: : Represents the overall score of optimization effectiveness. This indicates the number of days the funding cycle has shortened, and is a positive indicator. This indicates the new bad debt rate, which is a negative indicator. and : is a preset weighting coefficient used to balance the importance of improving capital efficiency and risk control. For example, it can be set to 0.6 and 0.4 respectively.

[0074] Next, the calculated Compare with a preset performance benchmark. The performance benchmark is the system's monthly performance over the past year. The average value. If the current If the value falls below this benchmark, a self-optimization process is triggered. This process is implemented using a gradient boosting algorithm, specifically the XGBoost framework. The algorithm's goal is to maximize... The algorithm is based on the aforementioned and As input features, the internal weight parameters of the bias quantization model in the above steps are used. , , and These weights are adjustable parameters. During algorithm initialization, they use preset initial values ​​(e.g., all set to 0.25). During algorithm execution, they will... The negative gradient is used as the residual to iteratively update the weight parameters. During the iteration process, L1 and L2 regularization can be used to prevent overfitting. After the iteration is complete, the updated weight parameters will overwrite the original parameter values ​​and be used for subsequent credit decisions. In some alternative implementations, the gradient boosting optimization algorithm can be replaced with other optimization algorithms, such as genetic algorithms or particle swarm optimization.

[0075] Through the above scheme, this embodiment can quantify the actual effect of credit strategy based on the efficiency of supply chain capital turnover and bad debt rate, and adaptively adjust the internal parameters of deviation quantification model using gradient boosting algorithm, thereby improving the system's accuracy in quantifying deviation risk and the level of intelligence in credit decision-making.

[0076] In one specific implementation, based on the above embodiments, the deviation quantification model is implemented as a weighted scoring model. First, the weighted scoring model receives initial deviation values ​​provided by the data acquisition module. Specifically, the model includes multiple weight coefficients, each associated with a specific business scenario data dimension, such as user purchase frequency, product category preference, and historical return rate. Then, the initial deviation values ​​are multiplied by the standardized values ​​of each business scenario data to obtain a series of weighted values. Next, all these weighted values ​​are summed, and the summation result is normalized and mapped to the [0,1] interval to obtain the final deviation severity score. This score reflects the risk level after comprehensively considering the initial deviation and various business context factors. In other optional implementations, different normalization methods, such as Min-Max standardization or the Sigmoid function, can be used to adapt to different data distributions and business needs. Furthermore, the weighted scoring model can be flexibly adjusted, not limited to weighted summation, and can also include nonlinear transformations, such as logarithmic or exponential functions, to enhance the model's sensitivity to specific risk factors.

[0077] Through the above solution, this embodiment can effectively integrate isolated deviation values ​​with multi-dimensional business scenario data, thereby generating a deviation severity score that is more business interpretable and risk-discriminating, providing a more accurate risk assessment basis for subsequent credit decisions.

[0078] In one specific implementation, based on the above embodiments, firstly, during the initialization phase, the system quantizes each weight coefficient in the deviation quantization model ( , , , Assign initial values. These initial values ​​can be preset based on empirical data or expert knowledge; for example, they can be set to 0.5, 0.2, 0.2, and 0.1 respectively. Then, in the outer loop self-optimization step of the above embodiment, the gradient boosting optimization algorithm will calculate a comprehensive score based on the optimization effect. Calculate the gradient for each weight coefficient. Specifically, assume the algorithm calculates... The gradient is ,but The new value will be updated using the following formula:

[0079] in: : Represents the updated weight coefficients. : Represents the current weighting coefficient. : represents the learning rate, which is a hyperparameter that controls the update step size, for example, 0.01. : express right The gradient.

[0080] This adjustment process is repeated iteratively until... Convergence or reaching the preset maximum number of iterations.

[0081] In some alternative implementations, instead of using a uniform gradient adjustment step size, an independent learning rate is set for each weight coefficient. For example, it could be... Set a small learning rate (e.g., 0.001) to ensure the impact of the initial bias value remains relatively stable; while for A larger learning rate (e.g., 0.01) can be set so that the system can adapt to changes in product category preferences more quickly. In addition to gradient boosting optimization, other optimization algorithms, such as the Adam algorithm or L-BFGS algorithm, can be used to adjust the weight coefficients.

[0082] Through the above approach, this embodiment can more precisely control the influence of each factor in the deviation quantification model, enabling the system to identify and quantify risks in different business scenarios more quickly and accurately, thereby improving the level of intelligence in credit decision-making.

[0083] In one specific implementation, based on the above embodiments, the system first continuously monitors the expected default rate index output by the risk simulation module. Specifically, after each feedback correction cycle, the system will update the current... With the preset risk tolerance threshold ( Compare them. Then, if Less than or equal to If the loop termination condition is met, the process ends, and the current credit strategy is determined as the final credit strategy. Next, if the number of loop feedback corrections reaches the preset maximum number (…), the process continues. For example, 3 times, even Still greater than The system will also forcibly terminate the loop and select the lowest-performing strategy from all tried strategies. The strategy is used as the final credit strategy.

[0084] In some other alternative implementations, The value can be adjusted based on actual business needs and risk tolerance, for example, set to 5 times or more. The maximum number of cycles can also be configured according to different credit product types.

[0085] Through the above approach, this embodiment ensures that the credit decision-making process will not loop indefinitely, provided that the expected risks are controllable. This avoids resource waste and decision delays caused by excessively pursuing the absolute optimal solution. Furthermore, even if the ideal risk level cannot be achieved within a limited number of attempts, a relatively superior strategy can be selected, ensuring the system's availability in complex scenarios.

[0086] In one specific implementation, based on the above embodiments, after the data collection is completed in the above steps and before calculating the composite supply chain health index, the collected raw data is first preliminarily processed using data cleaning rules to remove fields with more than 20% missing values ​​and obviously erroneous outliers. Specifically, the criteria for judging outliers can be as follows: The principle is that data points deviating from the mean by more than three standard deviations are considered outliers. Then, the cleaned data undergoes standardization preprocessing to eliminate the influence of different dimensionality of the indicators.

[0087] Specifically, for the historical contract fulfillment rate in supplier credit data, the Min-Max standardization method can be used to linearly map it to the [0, 1] interval. The calculation formula is as follows:

[0088] in: : Represents the value obtained after standardization, with a range of [0, 1]. : Represents the original fulfillment rate value to be processed. : Represents the minimum historical fulfillment rate in a predefined sample set. : Represents the maximum historical fulfillment rate within the same pre-defined sample set.

[0089] For the average accounts receivable period in the repayment records of downstream enterprises, since a larger value indicates higher risk, negative Min-Max standardization is adopted, and the calculation formula is as follows:

[0090] In other alternative implementations, Z-score standardization can be used to transform the data into a distribution with a mean of 0 and a standard deviation of 1. Alternatively, different standardization methods can be selected for different types of data; for example, Z-score standardization can be used for data that conforms to a normal distribution, while Min-Max standardization can be used for data with a well-defined distribution range. Furthermore, other nonlinear transformation methods, such as logarithmic transformation and exponential transformation, can be used to better adapt to the distribution characteristics of the data.

[0091] Through the above scheme, this embodiment can eliminate the differences in the units of measurement between different data sources and indicators, making the subsequent weighted fusion calculation more reasonable and effective, improving the accuracy of the composite supply chain health index, and thus enhancing the reliability of the overall credit decision.

[0092] In one specific implementation, based on the above embodiments, a neural network model and a support vector regression model are collaboratively invoked to generate a preliminary credit strategy. Specifically, firstly, the neural network model predicts and generates a credit line adjustment instruction based on the severity score of the bias and financial indicators such as the company's historical credit line and debt-to-equity ratio. This instruction can be a suggested percentage increase or decrease in the credit line. Then, the support vector regression model, based on the same severity score and combined with financial market data such as the current central bank benchmark interest rate and interbank lending rates, predicts and generates a financing interest rate correction instruction. This instruction can be a suggested percentage increase or decrease based on the benchmark interest rate. For example, the neural network model outputs a credit line adjustment instruction of "suggesting a 10% reduction in the credit line," while the support vector regression model outputs a financing interest rate correction instruction of "suggesting a 0.2 percentage point increase in the financing interest rate." Finally, the system integrates these two instructions to form a complete preliminary credit strategy that includes both the credit line adjustment percentage and the interest rate adjustment magnitude. In other alternative implementations, the credit limit adjustment instruction can be a direct suggestion of a specific adjusted credit limit value, and the financing interest rate correction instruction can be a suggestion of a specific adjusted interest rate value, rather than an increment based on the original value. Furthermore, the machine learning model can be, but is not limited to, decision tree models, gradient boosting tree models, or deep learning models.

[0093] Through the above solution, this embodiment can provide a credit strategy that includes two dimensions: credit limit adjustment and financing interest rate adjustment. This makes the final credit decision more refined, better balances risk control and return maximization, and thus improves the quality and efficiency of supply chain financial services.

[0094] like Figure 2 As shown, this embodiment of the invention provides an adaptive dynamic adjustment system for an intelligent prediction model, applied in the field of supply chain finance, for implementing the adaptive dynamic adjustment method of the intelligent prediction model in any of the above embodiments. The system includes: a data acquisition and quantification module M100, a collaborative decision-making module M200, and a closed-loop feedback control module M300.

[0095] The M100 data acquisition and quantification module is configured to collect supplier credit data and downstream enterprise repayment records from the supply chain via a data interface to obtain real-time business performance indicators. For example, the data interface can be built on a RESTful API and interface with the enterprise resource planning (ERP) system to periodically (e.g., hourly) retrieve supplier credit ratings, historical on-time delivery rates, and downstream enterprise accounts receivable terms and historical delinquency rates. This data, after being cleaned and transformed, is used to calculate real-time business performance indicators, such as the supply chain health index.

[0096] The M100 data acquisition and quantification module is also configured to determine the initial deviation between real-time business performance indicators and forecast results. The forecast results are generated by an independent forecasting model, which can be a time-series forecasting model trained on historical data, such as a seasonal ARIMA model. The initial deviation can be the absolute difference between the actual and forecast values ​​of the real-time business performance indicators.

[0097] When the initial deviation value exceeds a preset threshold, the data acquisition and quantification module M100 is configured to invoke the deviation quantification model to process business scenario data associated with real-time business performance indicators, thereby generating a deviation severity score characterizing the severity of the deviation. The deviation quantification model can be a weighted linear model, whose inputs include the initial deviation value and business scenario data collected from multiple channels, such as macroeconomic data (e.g., GDP growth rate, industry prosperity), market competition data (e.g., changes in competitors' market share), and event data (e.g., natural disasters, textual changes to specific industry regulations). The deviation quantification model weights and fuses this data, outputting a deviation severity score between 0 and 1. In other optional implementations, the deviation quantification model can be a non-linear model, such as a deep neural network model, to capture more complex relationships between business scenario data and risk levels.

[0098] The collaborative decision-making module M200 is configured to collaboratively invoke at least two functionally distinct machine learning models based on deviation severity scores, historical pattern features, and market feedback information to generate an initial credit strategy. For example, the collaborative decision-making module M200 can maintain a model registry registering multiple pre-trained machine learning models, including: a regression model for predicting a company's future revenue, a classification model for assessing the probability of default, and an optimization model for calculating the optimal credit limit. Based on the deviation severity score, historical pattern features obtained from other data sources (such as the company's revenue growth rate over the past three months), and market feedback information (such as sentiment analysis scores extracted from social media data), the collaborative decision-making module M200 selects an appropriate combination of models and collaboratively invokes these models to generate an initial credit strategy, such as a suggested credit limit, interest rate, and repayment period.

[0099] The closed-loop feedback control module M300 is configured to perform a risk assessment of the initial credit strategy through a risk pre-simulation module before execution, and to cyclically adjust the initial credit strategy based on the assessment results to obtain the final credit strategy. The risk pre-simulation module can be an event-driven discrete event simulator that simulates various stages of supply chain finance operations, such as: a company submitting a financing application, a bank approving a loan, the company using the loan, and the company repaying the loan. The risk pre-simulation module takes the initial credit strategy as input parameters, simulates business operations over multiple credit cycles, and evaluates different risk indicators, such as the expected bad debt rate and capital utilization rate. If the risk assessment results do not meet the preset risk preferences, the closed-loop feedback control module M300 will adjust the initial credit strategy and perform another risk assessment until a credit strategy that meets the risk requirements is found.

[0100] The closed-loop feedback control module M300 is also configured to automatically adjust the internal parameters of the deviation quantification model based on actual business performance data within the evaluation period after the final credit strategy is implemented, thereby improving the accuracy of subsequent deviation quantification by the deviation quantification model. For example, the closed-loop feedback control module M300 can collect business indicators such as the actual bad debt rate and capital turnover rate after the credit strategy is implemented, compare them with historical data, and calculate the return on this credit decision. Then, the closed-loop feedback control module M300 uses an optimization algorithm, such as gradient descent, to adjust the weighting coefficients in the deviation quantification model in reverse based on the return on this credit decision, thereby enabling the deviation quantification model to better adapt to the ever-changing market environment.

[0101] The aforementioned data acquisition and quantification module M100, collaborative decision-making module M200, and closed-loop feedback control module M300 work together to achieve intelligent and automated supply chain finance credit decision-making. The data acquisition and quantification module M100 is responsible for sensing risk signals from the supply chain in real time and converting these signals into quantifiable deviation severity scores. Based on these scores, the collaborative decision-making module M200 collaboratively calls multiple machine learning models to generate preliminary credit strategies. The closed-loop feedback control module M300 then performs risk simulations and iterative corrections on these strategies to ensure the robustness of credit decisions and continuously improves the system's decision-making capabilities through an external self-optimization mechanism. This series of collaborative efforts solves the technical problems of existing supply chain finance risk control systems, such as slow response, simplistic deviation assessment, rigid decision-making logic, and lack of self-learning capabilities.

[0102] Through the above solution, this embodiment can achieve real-time perception, accurate quantification, intelligent decision-making, and continuous optimization of supply chain finance risks, thereby improving the efficiency and quality of credit decisions, reducing risks, and enhancing the overall capital allocation efficiency of the supply chain.

[0103] In one specific implementation, based on the above embodiments, the collaborative decision-making module M200 further includes a historical pattern feature extraction unit and a market feedback information processing unit. The historical pattern feature extraction unit is configured to extract historical pattern features from historical time-series data of deviation severity scores. This extraction unit employs a sliding window algorithm with a window size set to 30 days, calculating the moving average, variance, and trend slope of the deviation severity scores within each window. These statistics constitute a three-dimensional historical pattern feature vector, used to characterize the pattern of deviation risk changes over time. In other optional implementations, the size of the sliding window can be adjusted according to actual conditions, for example, selecting 15 days, 60 days, or 90 days. In addition to the moving average, variance, and trend slope, other statistical features can also be calculated, such as the median, quartile range, kurtosis, or skewness.

[0104] The market feedback information processing unit is configured to process market feedback information. This unit includes a sentiment analysis subunit and a transaction volume analysis subunit. The sentiment analysis subunit accesses user review data from the e-commerce platform via API, uses a pre-trained BERT model to perform sentiment analysis on the review text, and obtains a sentiment score between -1 and 1, where -1 represents strong negative sentiment, 1 represents strong positive sentiment, and 0 represents neutral sentiment. The transaction volume analysis subunit obtains transaction volume data from the transaction system and calculates the rate of change of the current transaction volume compared to the average transaction volume of the past week. The sentiment analysis score and the transaction volume change rate constitute a two-dimensional market feedback information vector, used to represent the market's real-time feedback on the company's current operating status. In other optional implementations, the sentiment analysis model can be other natural language processing models, such as RoBERTa or XLNet. The data source for sentiment analysis can also be extended to social media platforms, such as Weibo or WeChat. In addition to calculating the rate of change compared to the average transaction volume of the past week, the transaction volume analysis can also calculate the rate of change compared to the average transaction volume of the past month.

[0105] Through the above-described scheme, this embodiment can more comprehensively extract historical pattern characteristics and market feedback information related to deviation risk, thereby providing more accurate and richer input information for subsequent credit decisions. Compared with the prior art, the beneficial effects of this embodiment are: by extracting historical pattern characteristics, it can more accurately identify the changing trends and cyclical patterns of deviation risk; by processing market feedback information, it can timely capture the market's real-time evaluation of the company's operating conditions, thereby more comprehensively assessing the company's credit risk.

[0106] In one specific implementation, based on the above embodiments, the closed-loop feedback control module M300 includes a comparison unit and a signal reconstruction unit. The comparison unit is configured to compare the expected default rate index output by the risk pre-simulation module with a preset risk tolerance threshold. When the expected default rate index exceeds the risk tolerance threshold, the comparison unit triggers the signal reconstruction unit. The signal reconstruction unit is configured to convert the excess expected default rate index into a new market feedback information, which includes two sub-steps: quantification and encoding. The quantification sub-step maps the difference between the expected default rate index and the risk tolerance threshold to a predefined set of risk levels (e.g., {"mild", "moderate", "severe"}). The encoding sub-step selects a corresponding feature vector based on the risk level, which is added to the input feature set of the collaborative decision-making module M200 to adjust the initial credit strategy. For example, if the difference is greater than 0.05, it is judged as a "serious" risk. The signal reconstruction unit generates a feature vector containing three elements: [-1,-1,0][-1,-1,0]. The first two elements are -1, representing a negative market signal, and the third element is 0, indicating that the signal comes from a risk simulation rather than the actual market. After receiving this feature vector, the collaborative decision-making module M200 adjusts its internal model parameters and outputs a more conservative credit strategy. In some other optional implementations, the signal reconstruction unit can use fuzzy logic or neural network models to dynamically adjust the values ​​of each dimension of the feature vector based on the difference between the expected default rate index and the risk tolerance threshold, in order to more finely control the feedback intensity.

[0107] Through the above approach, this embodiment can more effectively utilize the results of risk simulation by transforming risk simulation information into market feedback signals that can be identified and utilized by the collaborative decision-making module M200, thereby achieving more precise adjustments to credit strategies.

[0108] In one specific implementation, based on the above embodiments, the closed-loop feedback control module M300 includes an optimization effect comprehensive score calculation unit, a performance benchmark comparison unit, and a gradient boosting optimization unit. The optimization effect comprehensive score calculation unit is configured to receive actual business performance data from the business system, including supply chain cash flow efficiency (expressed as average order-to-payment cycle) and bad debt rate, and calculate the optimization effect comprehensive score using a preset formula. The performance benchmark comparison unit is configured to compare the calculated optimization effect comprehensive score with a preset performance benchmark, which can be a historical average score or a target score. The gradient boosting optimization unit is configured to activate the gradient boosting optimization algorithm when the optimization effect comprehensive score is lower than the performance benchmark, and use the optimization effect comprehensive score as a feedback signal to iteratively adjust the weight parameters in the deviation quantization model. The gradient boosting optimization algorithm can be implemented using frameworks such as XGBoost, LightGBM, or CatBoost. In one iteration, the gradient boosting optimization unit calculates the gradient direction and step size based on the current weight parameters and the optimization effect comprehensive score, and updates the weight parameters. For example, if supply chain cash flow efficiency declines and the bad debt rate increases, the gradient boosting optimization unit will adjust the weight parameters in the deviation quantification model to reduce the weight of high-risk factors and increase the weight of low-risk factors. In other alternative implementations, the comprehensive optimization score calculation unit can use different formulas, such as considering only supply chain cash flow efficiency or bad debt rate. The gradient boosting optimization unit can also use other optimization algorithms, such as genetic algorithms or particle swarm optimization.

[0109] Through the above approach, this embodiment can achieve adaptive optimization of the deviation quantification model, improve the accuracy and efficiency of credit decisions, and reduce risk.

[0110] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An adaptive dynamic adjustment method for an intelligent prediction model, applied in the field of supply chain finance, characterized in that, Includes the following steps performed by a computer: The system collects supplier credit data and downstream enterprise repayment records in the supply chain through a data interface to obtain real-time business performance indicators. Based on the real-time business performance indicators and a prediction result of the intelligent prediction model, the initial deviation value between the two is determined. When the initial deviation value exceeds a preset threshold, the deviation quantification model is invoked to process the business scenario data associated with the real-time business performance indicators in order to generate a deviation severity score that characterizes the severity of the deviation. Based on the aforementioned deviation severity score, historical pattern characteristics, and market feedback information, at least two machine learning models with different functions are collaboratively invoked to generate an initial credit strategy. Before implementing the preliminary credit strategy, a risk assessment is conducted on the preliminary credit strategy through a risk simulation module, and the preliminary credit strategy is cyclically revised based on the assessment results to obtain the final credit strategy. Based on the actual business performance data within the evaluation period after the final credit strategy is implemented, the internal parameters in the deviation quantification model are automatically adjusted to improve the accuracy of subsequent deviation quantification by the deviation quantification model.

2. The method according to claim 1, characterized in that, The real-time business performance indicator is the composite supply chain health index, which is calculated using a weighted fusion algorithm based on the historical fulfillment rate and credit rating data in the supplier credit data, as well as the payment period and overdue amount ratio data in the downstream enterprise repayment records.

3. The method according to claim 1, characterized in that, The business scenario data includes at least user purchase frequency data, product category preference data, and historical return rate data that are associated with the deviation subjects of the real-time business performance indicators.

4. The method according to claim 1, characterized in that, The collaborative invocation of at least two functionally distinct machine learning models also includes: The historical pattern features, including moving average, variance, and trend slope, are extracted from the historical time series data of the deviation severity scores using a sliding window algorithm. The market feedback information is processed, including sentiment analysis scores extracted from user comment data through a natural language processing model, and transaction volume change rate data.

5. The method according to claim 4, characterized in that, The collaborative invocation of at least two machine learning models with different functionalities is specifically as follows: The historical pattern features and the market feedback information are input into a pre-trained random forest model to determine whether there are policy conflict points. When a conflict point in the strategy is determined to exist, the severity score of the deviation is used as input to invoke at least two machine learning models with different functions.

6. The method according to claim 1, characterized in that, The collaborative invocation uses at least two machine learning models with different functions, one of which is a neural network model for optimizing the credit line of the core enterprise, and the other is a support vector regression model for correcting the trade finance interest rate of upstream and downstream enterprises.

7. The method according to claim 1, characterized in that, The risk simulation module performs a risk assessment on the preliminary credit strategy, specifically as follows: The expected default rate index under the initial credit strategy was calculated using the Monte Carlo simulation method.

8. The method according to claim 1, characterized in that, Before determining the initial deviation value, the following steps are also included: The collected supplier credit data and downstream enterprise repayment records are standardized and preprocessed to eliminate dimensional differences between different data sources.

9. The method according to claim 1, characterized in that, The collaborative invocation of preliminary credit strategies generated by at least two functionally different machine learning models includes credit limit adjustment instructions and financing interest rate correction instructions.

10. An adaptive dynamic adjustment system for an intelligent prediction model, applied in the field of supply chain finance, characterized in that, include: The data acquisition and quantification module is configured to: acquire supplier credit data and downstream enterprise repayment records in the supply chain through a data interface to obtain real-time business performance indicators, and determine the initial deviation value between the real-time business performance indicators and a prediction result of the intelligent prediction model. In addition, when the initial deviation value exceeds a preset threshold, the deviation quantification model is invoked to process the business scenario data associated with the real-time business performance indicators in order to generate a deviation severity score that characterizes the severity of the deviation. The collaborative decision-making module is configured to collaboratively invoke at least two machine learning models with different functions based on the deviation severity score, historical pattern characteristics, and market feedback information to generate an initial credit strategy. The closed-loop feedback control module is configured to perform a risk assessment on the preliminary credit strategy through a risk pre-simulation module before executing the preliminary credit strategy, and to perform cyclic feedback correction on the preliminary credit strategy based on the assessment results to obtain the final credit strategy. Furthermore, based on the actual business performance data within the evaluation period after the final credit strategy is implemented, the internal parameters in the deviation quantification model are automatically adjusted to improve the accuracy of subsequent deviation quantification by the deviation quantification model.

Images