Loan amount prediction methods, devices, equipment, storage media and program products

Abstract

This application provides a loan amount prediction method, apparatus, device, storage medium, and program product, relating to the field of artificial intelligence technology. The method includes: acquiring multi-source time-series behavioral data of a target user based on a loan request; the multi-source time-series behavioral data being a collection of multi-dimensional time-series data generated by the target user in a financial business scenario; decomposing the multi-source time-series behavioral data to generate a time-series feature vector corresponding to the target user, the time-series feature vector including multiple component feature vectors; determining a scenario-based feature vector based on the time-series feature vector and loan domain knowledge, the scenario-based feature vector indicating the target user's repayment ability and risk characteristics; and predicting the scenario-based feature vector using a preset prediction model to obtain the predicted loan amount corresponding to the target user. This method improves the accuracy of loan approval.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, storage medium, and program product for predicting loan amounts. Background Technology

[0002] In the approval process for personal consumer loans or micro and small enterprise loans, the dynamic assessment of the customer's repayment ability is the core element in determining the loan amount.

[0003] In related technologies, loan approval processes rely on static materials submitted by customers (such as income statements and credit reports). However, this data suffers from poor timeliness and fails to reflect recent changes in customer behavior. For example, a customer may be at the peak of their income when applying for a loan, but their repayment ability may subsequently plummet due to economic fluctuations or unforeseen events (such as unemployment or medical expenses). Such dynamic changes cannot be captured in real time, resulting in low accuracy in loan approval. Summary of the Invention

[0004] This application provides a method, apparatus, device, storage medium, and program product for predicting loan amounts, in order to solve the technical problem of low accuracy in loan approval.

[0005] Firstly, this application provides a method for predicting loan amounts, including:

[0006] Based on the target user's loan request, obtain the target user's multi-source time-series behavioral data, which is a multi-dimensional time-series data set generated by the target user in financial business scenarios.

[0007] The multi-source time-series behavioral data is decomposed to generate a time-series feature vector corresponding to the target user, and the time-series feature vector includes multiple component feature vectors.

[0008] Based on the time-series feature vector and loan domain knowledge, a scenario-based feature vector is determined, which is used to indicate the repayment ability and risk characteristics of the target user.

[0009] By using a preset prediction model, the scenario-based feature vector is processed to predict the loan amount corresponding to the target user.

[0010] Secondly, embodiments of this application provide a loan amount prediction device, including an acquisition module, a generation module, a determination module, and a prediction processing module:

[0011] The acquisition module is used to acquire multi-source time-series behavioral data of the target user based on the target user's loan request. The multi-source time-series behavioral data is a set of multi-dimensional time-series data generated by the target user in financial business scenarios.

[0012] The generation module is used to decompose the multi-source time-series behavioral data to generate a time-series feature vector corresponding to the target user, wherein the time-series feature vector includes multiple component feature vectors.

[0013] The determining module is used to determine a scenario-based feature vector based on the time-series feature vector and loan domain knowledge. The scenario-based feature vector is used to indicate the repayment ability and risk characteristics of the target user.

[0014] The prediction processing module is used to perform prediction processing on the scenario-based feature vector through a preset prediction model to obtain the predicted loan amount corresponding to the target user.

[0015] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0016] The memory stores computer-executed instructions;

[0017] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0019] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0020] The loan amount prediction method, apparatus, equipment, storage medium, and program products provided in this application can acquire multi-source time-series behavioral data of a target user based on the target user's loan request; decompose the multi-source time-series behavioral data to generate a time-series feature vector corresponding to the target user; fuse loan domain knowledge based on the time-series feature vector to determine a scenario-based feature vector; and predict the scenario-based feature vector through a preset prediction model to obtain the predicted loan amount corresponding to the target user. It can decompose multi-source time-series behavioral data in real time to generate time-series feature vectors, and then fuse scenario-based information to form scenario-based feature vectors, thereby accurately depicting the dynamic changes in the user's repayment ability, improving the model's ability to analyze the dynamic changes in customer repayment ability, avoiding the evaluation bias caused by traditional static models ignoring the dynamic characteristics of time series, and improving the accuracy of loan approval. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 A schematic diagram illustrating an application scenario provided in an embodiment of this application;

[0023] Figure 2 A flowchart illustrating a loan amount prediction method provided in this application embodiment;

[0024] Figure 3 A flowchart illustrating another loan amount prediction method provided in this application embodiment;

[0025] Figure 4 A schematic diagram of the architecture of a scenario-based feature construction and integration model provided in an embodiment of this application;

[0026] Figure 5 A schematic diagram of the architecture of a strategy adjustment engine provided in an embodiment of this application;

[0027] Figure 6 A schematic diagram illustrating the architecture of a loan amount prediction method provided in an embodiment of this application;

[0028] Figure 7 A schematic diagram of a dynamic timing decomposition and optimization architecture provided for an embodiment of this application;

[0029] Figure 8 This application provides a schematic diagram of the structure of a loan amount prediction device;

[0030] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0031] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0032] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0033] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0034] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0035] It should be noted that the loan amount prediction method, device, equipment, storage medium and program products provided in this application can be used in the field of artificial intelligence technology, or in any field other than artificial intelligence technology. The application field of the loan amount prediction method, device, equipment, storage medium and program products in this application is not limited.

[0036] Figure 1 This is a schematic diagram illustrating an application scenario provided by an embodiment of this application. Please refer to [link / reference]. Figure 1 This application scenario includes user device 101 and processing device 102. A target user can submit a loan request to processing device 102 through the business system set up in user device 101. Processing device 102 can, based on the loan request, obtain multi-source time-series behavioral data within a preset time period. Through decomposition and feature extraction of the multi-source time-series behavioral data, it determines the time-series feature vector corresponding to the target user. Based on the time-series feature vector, it integrates scenario-based domain knowledge to generate a scenario-based feature vector, thereby determining the predicted loan amount for the target user.

[0037] In related technologies, loan approval processes rely on static materials submitted by customers (such as income statements and credit reports). However, this data suffers from poor timeliness and fails to reflect recent changes in customer behavior. For example, a customer may be at the peak of their income when applying for a loan, but their repayment ability may subsequently plummet due to economic fluctuations or unforeseen events (such as unemployment or medical expenses). Such dynamic changes cannot be captured in real time, resulting in low accuracy in loan approval.

[0038] The loan amount prediction method provided in this application can obtain multi-source time-series behavioral data of a target user based on their loan request; decompose the multi-source time-series behavioral data to generate a time-series feature vector corresponding to the target user; fuse loan domain knowledge based on the time-series feature vector to determine a scenario-based feature vector; and use a preset prediction model to predict the scenario-based feature vector to obtain the predicted loan amount for the target user. This method can decompose multi-source time-series behavioral data in real time to generate time-series feature vectors, and then fuse scenario-based information to form scenario-based feature vectors, thereby accurately depicting the dynamic changes in a user's repayment ability. This improves the model's ability to analyze the dynamic changes in a customer's repayment ability, avoids the evaluation bias caused by traditional static models ignoring the dynamic characteristics of time series, and improves the accuracy of loan approval.

[0039] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0040] Figure 2 This is a flowchart illustrating a loan amount prediction method provided in an embodiment of this application. Please refer to... Figure 2 The method may include:

[0041] S201. Based on the target user's loan request, obtain multi-source time-series behavioral data of the target user.

[0042] The execution entity in this application embodiment can be a processing device or a loan amount prediction device installed in the processing device. The loan amount prediction device can be implemented by software or by a combination of software and hardware.

[0043] It can receive loan requests sent by target users through user devices, and the loan requests may include the target user's user identifier.

[0044] After obtaining explicit authorization from the target user, customer multi-source time-series behavioral data can be obtained from the bank's core system, third-party payment platform, etc., through user identification and API or database interface.

[0045] Multi-source time-series behavioral data refers to a collection of multi-dimensional time-series data generated by target users in financial business scenarios. Examples include bank transaction records, third-party payment transaction records, and consumption behavior logs.

[0046] Multi-source time-series behavioral data can be data within a preset time period, such as the target user's bank statements for the past 30 days. These statements include key records such as the time, amount, and transaction counterparty of daily income, expenditure, and transfer operations.

[0047] S202. Decompose the multi-source time-series behavioral data to generate the time-series feature vector corresponding to the target user.

[0048] The dynamic temporal decomposition model can be used to extract feature components of different dimensions from non-stationary multi-source temporal behavioral data, forming a structured temporal feature vector.

[0049] Dynamic time series decomposition models can employ algorithms that support real-time decomposition, such as improved seasonal-trend decomposition using LOESS (STL) models based on local weighted regression or adaptive Prophet models. Both types of models can handle the non-stationarity of financial time series data (such as income fluctuations and sudden expenditures) and support high-frequency data updates, adapting to the needs of real-time loan scenarios.

[0050] Taking the improved STL model as an example, decomposition can be achieved through three-layer iterative smoothing: ① Smooth the original time series data (such as daily income flow) using the LOESS (Locally Weighted Regression) method to separate the initial trend term; ② Based on the preset window length (such as monthly, bi-weekly, to adapt to the salary cycle of loan scenarios), extract the periodic fluctuation part to form the periodic component; ③ Introduce holiday calibration factors (such as statutory holidays, industry special nodes) to remove the abnormal fluctuations brought by holidays from the remaining data to obtain the holiday component; ④ The final remaining random fluctuation data that cannot be explained by trends, cycles, or holidays is the residual component (reflecting sudden abnormal situations such as income and expenditure).

[0051] Taking the adaptive Prophet model as an example, the time series modeling formula is used... Decomposition: ① Automatically fit the nonlinear trend term (e.g., long-term revenue growth / decline trend); ② Identify the periodic components corresponding to the preset window length. (e.g., weekly or monthly consumption cycles, adapting to key cycles in loan scenarios such as bi-weekly paydays and monthly salaries); ③ Embed holiday datasets, using binary variables to label the impact of holidays and generate holiday components. ④ Residual components The error term after model fitting is used to filter out abnormal fluctuations based on the normal distribution assumption.

[0052] For residual components, abnormal fluctuations can be identified using mutation point detection algorithms (such as CUSUM, Bayesian Change Point, etc.).

[0053] The time series feature vector is a set of multi-dimensional components output after decomposition. The time series feature vector may include trend components (reflecting long-term income and expenditure trends), periodic components (reflecting regular income and expenditure fluctuations), holiday components (reflecting the special income and expenditure impact of holidays), and residual components (reflecting sudden abnormal income and expenditure). Each component is stored in the form of a time series sequence (such as the corresponding component value for each day).

[0054] S203. Based on time-series feature vectors and loan domain knowledge, determine scenario-based feature vectors.

[0055] Contextualized feature vectors are used to indicate a target user's repayment ability, willingness to repay, and potential risks. Contextualized feature vectors can include the target user's income feature vector and contextual feature vector.

[0056] In some embodiments, the income feature vector can be determined based on the time-series feature vector; the scenario feature vector can be determined based on the time-series feature vector and loan domain knowledge; and the income feature vector and the scenario feature vector can be concatenated to obtain the scenario-based feature vector.

[0057] Loan-related knowledge refers to business rules, experiential indicators, and data dimensions that have been proven in practice and are directly related to repayment ability / risk in loan approval scenarios. Examples include safety cushion days, historical repayment records, and debt-to-income ratio.

[0058] Income feature vectors can be used to indicate the trend, cycle, risk resistance, and integration characteristics of target users; scenario feature vectors can be used to indicate safety cushion, performance history, debt-to-income ratio, and scenario-adaptability characteristics.

[0059] In this application, by introducing knowledge from the lending field to form a scenario feature vector, it is possible to integrate time-series dynamic features with the lending scenario to characterize the repayment ability, repayment willingness and potential risks of the target user, thereby improving the accuracy of subsequent predictions.

[0060] S204. By using a preset prediction model, the scenario-based feature vector is processed to obtain the predicted loan amount for the target user.

[0061] The preset prediction model can adopt a hierarchical integrated architecture, which can include decision tree models and temporal neural network models. The two models can work together to leverage the advantages of traditional machine learning in processing high-dimensional heterogeneous features and the advantages of neural networks in capturing temporal dynamic patterns, thereby improving the accuracy and generalization ability of prediction.

[0062] In some embodiments, a decision tree model can be used to predict the scenario-based feature vector to obtain a first intermediate result; a temporal neural network model can be used to predict the scenario-based feature vector to obtain a second intermediate result; and the first and second intermediate results can be weighted to obtain the predicted loan amount.

[0063] The decision tree model can be the LightGBM model, an efficient machine learning model based on gradient boosting decision trees, suitable for handling high-dimensional heterogeneous features. For example, the LightGBM model can handle features such as trend stability index and periodic amplitude stability in contextualized feature vectors.

[0064] Temporal neural network models can refer to neural network models specifically designed for processing time-series data, such as Long Short-Term Memory (LSTM) networks or Temporal Convolutional Networks (TCNs). For example, LSTM models can capture long-term growth or decline patterns in trend components.

[0065] The first layer, the LightGBM model, handles high-dimensional heterogeneous features (such as trend-cycle coupling factors and income coverage ratios), while the second layer, a temporal neural network model (such as LSTM), handles the long-term dependencies of time-series components. The outputs of the two models are then combined through a weighted average or a neural network fusion processor to generate the final loan amount recommendation.

[0066] In this application, the synergistic effect of the hierarchical ensemble model enables in-depth mining of high-dimensional heterogeneous features and time-series patterns, improves the model's ability to fuse multi-dimensional features, and thus further enhances the accuracy of loan amount prediction.

[0067] The data processing method provided in this application embodiment can decompose multi-source time-series behavioral data in real time, generate time-series feature vectors, and then integrate scenario-based information to form scenario-based feature vectors, thereby accurately depicting the dynamic changes in users' repayment ability, improving the model's ability to analyze the dynamic changes in customers' repayment ability, avoiding the evaluation bias caused by traditional static models ignoring the dynamic characteristics of time series, and improving the accuracy of loan approval.

[0068] Figure 3 This is a flowchart illustrating another loan amount prediction method provided in an embodiment of this application. Please refer to... Figure 3 The method may include:

[0069] S301. Based on the target user's loan request, obtain multi-source time-series behavioral data of the target user.

[0070] The execution process of S301 can be found in the execution process of S201, and will not be repeated here.

[0071] S302. Update the preset parameters of the dynamic time series decomposition model to the target parameters using the sliding window optimizer.

[0072] The sliding window optimizer is a dynamic parameter adjustment module based on recent data performance, which allows the dynamic time series decomposition model to adapt to changes in user behavior patterns and avoid decomposition bias caused by fixed parameters.

[0073] In some embodiments, a sliding window optimizer can be used to adjust the preset window length to the target window length and the preset decomposition parameters to the target decomposition parameters based on the prediction performance of the feature vectors of each component within the recent window.

[0074] The nearest window can be a window preceding the current time. The predictive efficacy of each component feature vector within the nearest window refers to the contribution of each component feature vector within the nearest window to the prediction of subsequent loan amounts, evaluated through quantitative indicators.

[0075] For example, prediction performance can be measured by prediction error; the larger the prediction error, the lower its contribution to the prediction within the adjacent window.

[0076] When the prediction performance within a neighboring window meets the adjustment threshold, the sliding window optimizer initiates adjustments to the preset parameters. For example, if the prediction error increases for three consecutive window periods, the preset parameters need to be adjusted.

[0077] Preset parameters may include preset window length and preset decomposition parameters, where preset decomposition parameters may be preset smoothing coefficients for STL models or preset trend flexibility parameters for Prophet models.

[0078] When the preset decomposition parameter is the preset smoothing coefficient, the target decomposition parameter is the preset smoothing coefficient; when the preset decomposition parameter is the preset trend flexibility parameter, the target decomposition parameter is the preset trend flexibility parameter.

[0079] Regarding window size adjustments, if the user data pattern changes (e.g., income shifts from short-cycle fluctuations with bi-weekly paychecks to a monthly fixed income plus long-term growth pattern), the time window of the STL model can be expanded to enhance trend capture, or the window can be reduced to adapt to the new short-cycle patterns.

[0080] For parameter fine-tuning, if the residual components after decomposition If volatility increases abnormally (indicating insufficient model fit), then fine-tune the LOESS smoothing coefficient of the STL model and the holiday impact weight of the Prophet model to improve the decomposition accuracy.

[0081] The original preset window length of the dynamic time series decomposition model can be updated to the target window length that adapts to the current data pattern, and the fine-tuned core parameters of the model can be output simultaneously.

[0082] In this application, the real-time evaluation and dynamic adjustment capabilities of the sliding window optimizer enable adaptive optimization of the decomposition model parameters, thereby improving the sensitivity and adaptability of the time-series decomposition to changes in user behavior patterns, reducing feature distortion caused by decomposition lag, and further reducing the risk of bad debts.

[0083] S303. Using a dynamic temporal decomposition model, multi-source temporal behavior data is decomposed based on target parameters to generate temporal feature vectors.

[0084] The dynamic temporal decomposition model will perform accurate decomposition on real-time acquired multi-source temporal behavior data based on the updated target window length and target decomposition parameters.

[0085] The decomposition process strictly follows the core logic of the model: ① Extracting cyclical components that are strongly correlated with the loan scenario based on the target window length (such as the monthly salary payment cycle and the concentrated consumption cycle during holidays to adapt to the new window); ② Combining the fine-tuned target decomposition parameters, more accurately separating trend components (such as long-term income growth / decline trends) and holiday components (such as the adjusted impact of holidays); ③ Finally, obtaining the residual components (containing only sudden abnormal income and expenditure, and eliminating the fitting error after model adaptation).

[0086] After decomposition, the output is a structured time-series feature vector consisting of trend components, periodic components, holiday components, and residual components. This ensures that the feature vector accurately reflects the user's current behavior pattern and provides high-quality input for subsequent scenario-based feature construction.

[0087] S304. Construct an income feature vector based on the time-series feature vector.

[0088] Income feature vectors are used to quantify the stability, growth potential, and risk resistance of user income, and can include multiple income sub-vectors.

[0089] Specifically, for any income sub-vector, at least one component feature vector corresponding to the income sub-vector can be determined, and the income sub-vector can be determined based on the at least one component feature vector.

[0090] Among them, multiple income sub-vectors can be short-term growth slope, cycle amplitude stability, residual volatility, and trend cycle coupling factor.

[0091] If the income subvector is the short-term growth slope, then at least one component feature vector is the trend component. Based on a preset sliding window, a linear regression can be performed on the trend component to obtain the regression coefficient within the preset sliding window, and this regression coefficient can be determined as the short-term growth slope.

[0092] The short-term growth slope can be used to indicate the strength of the recent growth / decline trend of user income. The larger the slope is and the greater the absolute value, the stronger the income growth momentum and the higher the potential for improving repayment ability.

[0093] If the income subvector is periodically stable, then at least one component feature vector is a periodic component. The relevant periods that are strongly correlated with the loan scenario can be extracted from the periodic components. The fluctuation amplitude of multiple consecutive relevant periods (such as the last 3 relevant periods) (the difference between the maximum and minimum income in each period) can be determined. Then, by calculating the coefficient of variation (standard deviation / mean) of the fluctuation amplitude, the periodic amplitude stability can be obtained.

[0094] Relevant cycles can be selected based on knowledge of the loan industry; for example, relevant cycles could be bi-weekly payday or monthly payday.

[0095] The smaller the coefficient of variation, the more stable the fluctuation of cyclical income and the stronger the predictability of repayment sources.

[0096] If the income subvector is the residual volatility, then at least one component eigenvector is a residual component. The residual component can be calculated using the rolling standard deviation (Rolling StdDev). The residual component is calculated period by period using the target window length or a preset window length (which is consistent with the sliding window of the short-term growth slope) to obtain a time series of residual volatility with the same sequence length as the residual component.

[0097] Residual volatility is used to quantify the impact of sudden income and expenditure (such as unexpected expenses or temporary income) on user income. The lower the volatility, the smaller the dispersion of residual data within the window, indicating that users have fewer irregular sudden income and expenditure, stronger income stability, and better resistance to interference in repayment ability. The higher the volatility, the more drastic the fluctuation of residual data within the window, indicating that user income is greatly affected by unexpected factors, repayment ability is uncertain, and loan risk is relatively high.

[0098] If the income subvector is a trend-cycle coupling factor, then at least one component eigenvector is a trend component and a cycle component, and the ratio of the trend component to the cycle component can be determined as the trend-cycle coupling factor.

[0099] The trend-cycle coupling factor is used to indicate the synergy between long-term income trends and short-term cyclical fluctuations. A ratio close to 1 with small fluctuations indicates a stable match between long-term trends and short-term cycles, resulting in a healthier income structure.

[0100] In this application, by extracting targeted features from time-series feature vectors, abstract time-series data is transformed into quantifiable and interpretable income capacity indicators, thereby achieving an accurate depiction of the dynamic changes in users' repayment ability. This effectively avoids the limitations of traditional static income assessment and improves the quality of underlying data for loan amount prediction.

[0101] S305. Construct scenario feature vectors based on time-series feature vectors and loan domain knowledge.

[0102] Scenario feature vectors can deeply bind dynamic income features with core risk points in loan approval (i.e., loan domain knowledge), allowing features to directly serve the risk assessment needs of loan business scenarios.

[0103] Specifically, based on knowledge of the loan domain, multiple scenario sub-vectors can be determined in the time-series feature vector; these multiple scenario sub-vectors are then concatenated to obtain the scenario feature vector.

[0104] Multiple scenario sub-vectors can be used to measure revenue coverage ratio, safety cushion days, and scenario adaptability.

[0105] Based on the knowledge in the lending field that "income must cover rigid expenses and repayment obligations", the income coverage ratio is determined. The income coverage ratio can be (average of trend component + average of period component) / (monthly necessary expenses + expected monthly repayment amount).

[0106] Among them, monthly essential expenditures can be filtered by historical consumption flow (time series data) (such as rigid expenditures such as food, housing, and transportation), and the expected monthly repayment amount can be preset based on the loan product type (such as credit loan and business loan).

[0107] The income coverage ratio can directly quantify the ability of income to cover rigid expenses and loan repayments. When the income coverage ratio is ≥1.5 (the industry experience threshold), it indicates that the repayment pressure is manageable.

[0108] Based on the knowledge of "matching repayment nodes with income nodes" in the field of lending, the payday and loan bill repayment date can be determined by combining the periodic component in the time series feature vector. The safety cushion days = next payday - bill repayment date.

[0109] The safety cushion days can be directly adapted to the risk of loan repayment time nodes. When the safety cushion days are ≥7 (industry experience threshold), the risk of delinquency caused by salary delays can be reduced. It is the core indicator of the time dimension safety cushion in loan approval.

[0110] Based on the knowledge in the lending field that "different loan purposes correspond to different risk characteristics," we can customize scenario adaptation flexibility for different loan types:

[0111] For business loans, the elasticity of operating income (growth slope / variance coefficient of periodic amplitude) can be determined based on the stability of operating cash flow in the periodic component and the growth slope of the trend component. The elasticity of operating income can be determined as the elasticity of scenario adaptation, and the elasticity of operating income in supporting repayment can be assessed.

[0112] For consumer loans, the consumption expenditure elasticity (residual volatility / income coverage ratio) can be determined based on the volatility of sudden expenditures in the residual component. This consumption expenditure elasticity can be defined as scenario-adaptive elasticity, which can assess the impact of sudden consumption on repayment ability.

[0113] The scenario-adaptive flexibility enables precise matching between income dynamics and loan purposes, making it more scenario-specific.

[0114] In this application, by using knowledge of the loan domain as a logical framework for feature construction, time-series features are transformed from general data into scenario-based risk indicators, achieving a deep binding between features and the core requirements of loan business, avoiding the problem of features being disconnected from risks, and further improving the accuracy of loan amount prediction and risk controllability.

[0115] S306. Concatenate the income feature vector and the scene feature vector to obtain the scene-based feature vector.

[0116] Revenue feature vectors and scenario feature vectors need to be generated based on the same time dimension (e.g., both using "the last 30 days" as the target window length) to ensure that the time range of user behavior corresponding to the features is consistent and to avoid feature misalignment across periods.

[0117] The income feature vector and the scenario feature vector are horizontally concatenated in the order of "income capacity first, scenario risk second" to form a complete scenario-based feature vector.

[0118] That is, the total number of dimensions of the concatenated contextual feature vectors = the number of dimensions of the income feature vectors + the number of dimensions of the contextual feature vectors.

[0119] For example, the income feature vector (4 dimensions): [Short-term growth slope = 0.08 (average monthly income growth of 8%), cycle amplitude stability = 0.15 (coefficient of variation), residual volatility = 1200 (yuan), trend cycle coupling factor = 1.05]; the scenario feature vector (3 dimensions): [income coverage ratio = 1.8, dynamic safety cushion days = 12 (days), scenario adaptation elasticity = 0.7 (consumer loan scenario)]; the scenario-based feature vector is [0.08, 0.15, 1200, 1.8, 12, 0.7].

[0120] S307. Using a decision tree model, the scenario-based feature vector is predicted to obtain the first intermediate result.

[0121] In the offline phase, a training set is constructed based on the contextualized feature vectors of historical users and their actual repayment performance (such as whether they are overdue and their repayment ability is sufficient). The model is trained to learn the mapping relationship between features and loan amounts, using a reasonable loan amount range as the label. At the same time, hyperparameters such as learning rate and tree depth are optimized through cross-validation.

[0122] In the online prediction phase, the current user's contextualized feature vector is input into the trained model. The model outputs a first intermediate result through ensemble voting of multiple decision trees and weighted summation. Based on the basic loan amount (e.g., 150,000 yuan) based on the nonlinear correlation of features, this result is the "static matching degree between current repayment ability and risk".

[0123] The first intermediate result is a single numerical credit limit (unit: yuan), with weights for the importance of features (e.g., income coverage ratio weight 35%, dynamic safety cushion days weight 25%).

[0124] S308. Using a temporal neural network model, the scene-specific feature vector is predicted to obtain a second intermediate result.

[0125] In the offline phase, the training data is the time-series features of historical users and their long-term repayment performance (such as repayment stability over 6-12 months). The model is trained to learn the mapping relationship between time-series patterns and loan amounts using long-term sustainable loan amounts. The model's generalization ability is improved by adjusting hyperparameters such as the number of network layers, the number of hidden units, and the dropout rate.

[0126] In the online prediction phase, the time-series features of the current user are input into the model. LSTM uses a gating mechanism to memorize key information about income growth and cyclical fluctuations, while TCN uses dilated convolution to capture multi-scale cyclical patterns, and finally outputs the second intermediate result.

[0127] The second intermediate result is a single numerical credit limit (unit: yuan), accompanied by a time-series evaluation result (such as "significant long-term income growth trend, credit limit can be increased by 20%" or "stable cyclical fluctuations, credit limit stability score of 85 points").

[0128] S309. Weight the first and second intermediate results to obtain the predicted loan amount.

[0129] In some embodiments, based on a multimodal attention mechanism, the first weight corresponding to the decision tree model and the second weight corresponding to the temporal neural network model are determined according to the contextual feature vector; the sum of the product of the first intermediate result and the first weight, and the product of the second intermediate result and the second weight, is determined as the predicted loan amount.

[0130] The "time-related features" (including trend stability index, periodic amplitude stability, residual volatility, trend-period coupling factor, and sequence statistical features of the original trend component and periodic component) in the contextualized feature vector are used as input to the attention mechanism, and the features are encoded into a unified-dimensional attention vector through a fully connected layer.

[0131] Based on the encoded feature vectors, the temporal stability signal and the current matching signal are determined, and the temporal stability signal and the current matching signal are normalized to obtain the first weight corresponding to the decision tree model. ) and the second weight corresponding to the temporal neural network model ( ),in, .

[0132] Among them, the time-series stability signal can be characterized by low periodic amplitude stability and high residual volatility, indicating large dynamic fluctuations in user income, requiring an increase in the weight of the time-series neural network model; the current matching signal can be characterized by high income coverage ratio and sufficient safety cushion days, indicating clear current repayment ability, requiring an increase in the weight of the decision tree model.

[0133] A weighted summation formula is used to merge the intermediate results of the two-layer model:

[0134]

[0135] in: It is the first weight (the weight of the decision tree model). This is the second weight (weight of the temporal neural network model).

[0136] For example, User A (consumer loan applicant): Time-series characteristics are "cycle amplitude stability = 0.12 (low volatility), residual volatility = 800 yuan (low impact), trend growth slope = 0.05 (stable growth)", and scenario-based characteristics are "income coverage ratio = 2.1 (high coverage), safety cushion days = 15 days (sufficient)"; attention mechanism calculation results: (Decision tree models have higher weights because current repayment ability is clear.) (The time series model has appropriate weights and a stable long-term trend); if the first intermediate result is 160,000 yuan and the second intermediate result is 170,000 yuan, then .

[0137] Figure 4 This is a schematic diagram of the architecture of a scenario-based feature construction and integration model provided in an embodiment of this application. Please refer to... Figure 4 By integrating temporal features and domain knowledge features, the temporal features extracted after dynamic temporal decomposition are combined with loan domain knowledge features to generate scenario-based feature vectors that cover both user income and expenditure dynamics and loan risk requirements.

[0138] Once feature construction is complete and the integration and preprocessing of scenario-based feature vectors (such as standardization and missing value imputation) are finished, the feature construction phase is finished, and the feature vectors are input into the subsequent integrated model module.

[0139] In the hierarchical model processing, the LightGBM / XGBoost model layer first inputs the scenario-based feature vectors into the gradient boosting decision tree model, leveraging its advantage of efficiently processing high-dimensional heterogeneous features to capture the non-linear correlation between features and output the first intermediate prediction result based on the matching of current risk and capability. Then, through the temporal neural network / temporal attention mechanism layer, the temporal sequence-like features in the scenario-based features are input into the temporal model such as LSTM / TCN, leveraging its ability to capture long-term dependencies to output the second intermediate prediction result based on the temporal dynamic trend.

[0140] In the model fusion strategy determination, a fusion strategy is selected for the intermediate results of the two-layer model: if the determination is "continue", the current fusion logic is maintained; if the determination is "yes", the adaptive fusion process is started.

[0141] In adaptive fusion processing, the dynamic weights of the two-layer model are first determined based on a multimodal attention mechanism (or error feedback), and the intermediate results are weighted and summed. Then, an end-to-end trained neural network fusion engine further learns the optimal combination of the two-layer results to improve fusion accuracy. After fusion is completed, the final loan amount prediction result is output, marking the end of the ensemble model prediction.

[0142] In this application, a multimodal attention mechanism is used to achieve dynamic weight allocation, replacing the traditional fixed weight fusion. This allows the credit limit prediction to fully utilize the current risk control capability of the decision tree model and the long-term trend capture capability of the temporal neural network model. It achieves adaptive fusion under different user behavior patterns, effectively balancing loan risk and user repayment potential, and further improving the accuracy and personalization of loan limit prediction.

[0143] The loan amount prediction method provided in this application can decompose multi-source time-series behavioral data in real time, generate time-series feature vectors, and then integrate scenario-based information to form scenario-based feature vectors, thereby accurately depicting the dynamic changes in the user's repayment ability, improving the model's ability to analyze the dynamic changes in the customer's repayment ability, avoiding the evaluation bias caused by the traditional static model ignoring the dynamic characteristics of time series, and improving the accuracy of loan approval.

[0144] In some embodiments, after predicting the scenario-based feature vectors using a preset prediction model to obtain the predicted loan amount for the target user, the method further includes: adjusting the loan amount correction coefficient, model hyperparameters, or triggering a manual review process based on the quantitative indicator of the difference between the predicted loan amount and the actual repayment performance using a real-time strategy adjustment engine.

[0145] The real-time strategy adjustment engine refers to an intelligent decision-making module based on a combination of reinforcement learning algorithms and a rule engine. It is used to continuously calibrate prediction results through real-time monitoring and dynamic adjustment, balancing prediction accuracy and risk controllability. Among them, reinforcement learning is responsible for adaptively learning and adjusting strategies (such as optimizing the direction of the credit limit correction coefficient), while the rule engine is responsible for executing explicit risk control thresholds (such as the conditions for triggering manual review).

[0146] For example, when the predicted results deviate significantly from the actual repayment performance, the real-time strategy adjustment engine can trigger a manual review process.

[0147] The real-time strategy adjustment engine continuously tracks the actual repayment performance of target users and calculates quantitative indicators of difference. These quantitative indicators include the credit limit deviation rate and the performance deviation rate. Specifically, the credit limit deviation rate = (predicted loan amount - reasonable amount corresponding to actual affordable repayment) / predicted loan amount; and the performance deviation rate = (predicted performance probability - actual performance probability) × 100% (the actual performance probability is calculated from repayment records).

[0148] When the deviation quantification index is within the preset normal range (e.g., absolute value of the quota deviation rate ≤ 10%, absolute value of the performance deviation rate ≤ 15%), the current quota correction coefficient (initial value 1.0) and model parameters are maintained.

[0149] When the deviation quantification index exceeds the normal range but does not reach the risk threshold (e.g., 10% < absolute value of credit limit deviation rate ≤ 20%), the credit limit correction coefficient is dynamically adjusted. For example, if the system predicts that a customer has a high repayment ability but the actual repayment performance is poor (credit limit deviation rate = 25%, performance deviation rate = -20%), the credit limit correction coefficient will be lowered from 1.0 to 0.8 (i.e., the predicted credit limit for subsequent users of the same type will be output at 80%); if the predicted repayment ability is low but the actual repayment performance is good, the correction coefficient can be raised to 1.1.

[0150] When the deviation quantification indicators reach the risk threshold (such as the absolute value of the credit limit deviation rate > 20%, the absolute value of the performance deviation rate > 25%, or three consecutive periods of overdue payments), the manual review process is immediately triggered. Risk control personnel will check the authenticity of user characteristics, deviations in model feature extraction, abnormal behavior, and other issues, and manually adjust the credit limit or optimize model rules based on the review results.

[0151] The real-time policy adjustment engine feeds back adjustment records (such as correction coefficient changes and manual review results) to the model training module as samples for online learning. This is used to fine-tune the hyperparameters of the preset prediction model (such as the fusion weights of the ensemble model and the window size of the temporal decomposition), thereby achieving collaborative iteration between the policy and the model.

[0152] In this application, the closed-loop feedback mechanism of the real-time strategy adjustment engine is used to realize the dynamic optimization of the model prediction results, improve the model's adaptability and risk control capabilities, and thus further improve the automation and intelligence of the entire loan amount prediction process.

[0153] Figure 5 This is a schematic diagram of the architecture of a strategy adjustment engine provided in an embodiment of this application. Please refer to... Figure 5 When evaluating the loan amount output by the evaluation model, the system receives the predicted loan amount output by the integrated model and, combined with the user's contextual characteristics (such as income coverage ratio and safety cushion days), conducts a preliminary quantitative assessment of the reasonableness of the amount. Based on preset risk thresholds, it determines whether the predicted loan amount is reasonable: It judges whether the current predicted amount is reasonable based on preset risk thresholds (such as amount deviation rate and performance risk score); if it is determined to be "unreasonable," it enters the "optimize loan amount" stage, calibrating the predicted amount based on real-time strategy adjustment logic (such as adjusting the amount correction coefficient); if it is determined to be "reasonable," it enters the automatic approval stage.

[0154] Input a reasonable predicted amount into the automatic approval rules (such as matching the loan product's limit or the user's risk level): If the automatic approval "passes", the final amount is returned to the limit decision engine to complete this round of strategy process; if the automatic approval "failes", the "feedback adjustment" step is triggered, and the adjustment signal is sent to the model / feature step (such as strengthening the weight of a certain type of risk feature) based on the reasons for the failure to approve (such as the amount exceeding the product limit or the user's qualifications not matching).

[0155] This creates a closed-loop optimization process. The optimized loan amount or the adjusted result will be re-input into the loan amount output of the evaluation model, forming a closed loop of evaluation, decision-making, adjustment, and re-evaluation, which continuously improves the rationality and suitability of the loan amount decision.

[0156] Figure 6 This is a schematic diagram illustrating the architecture of a loan amount prediction method provided in an embodiment of this application. Please refer to [link / reference]. Figure 6The system collects multi-source time-series behavioral data of target users in real time (such as bank statements and third-party payment records) and performs data quality checks and standardization. For the collected multi-source time-series data, an improved STL or adaptive Prophet model is used for dynamic time-series decomposition and optimization to obtain trend, periodic, holiday, and residual components. Simultaneously, a sliding window optimizer is used to dynamically adjust the window size and parameters of the decomposition model based on recent predictive performance, ensuring that the time-series components accurately reflect the user's current behavioral patterns.

[0157] By constructing scenario-based features, the decomposed time-series components are transformed into income feature vectors (such as short-term growth slope and residual volatility), and loan domain knowledge (such as safety cushion days and dynamic debt-to-income ratio) is integrated to generate scenario feature vectors. Finally, the two types of vectors are concatenated into scenario-based feature vectors to form the input data of the integrated model.

[0158] The ensemble model prediction adopts a hierarchical ensemble architecture. The first intermediate result is obtained through the LightGBM / XGBoost model (to handle high-dimensional heterogeneous features), and the second intermediate result is obtained through the LSTM / TCN temporal neural network model (to capture temporal dynamic patterns). Then, the two results are dynamically weighted and fused based on a multimodal attention mechanism to output the preliminary predicted loan amount.

[0159] The credit limit decision engine inputs the predicted loan amount into the automatic approval process on the one hand, and links with the loan strategy optimization module (i.e., the real-time strategy adjustment engine) on the other hand to continuously monitor the difference between the predicted amount and the actual repayment performance, dynamically adjust the credit limit correction coefficient or trigger the manual review process, and realize the closed-loop iteration of the strategy.

[0160] If the automatic approval is successful, an electronic contract will be generated directly, and the loan amount will be finalized. If the automatic approval fails, the system will send a loan rejection notice and trigger a manual review process. The manual approval desk will verify the user's characteristics, model prediction logic, etc., to ensure that the risk is controllable.

[0161] Figure 7 This is a schematic diagram of a dynamic timing decomposition and optimization architecture provided in an embodiment of this application. Please refer to... Figure 7 Multi-source time-series behavioral data can be decomposed using STL / Prophet, outputting four core time-series components: trend component, periodic component, holiday component, and residual component. In deep optimization and feature extraction, based on the decomposed time-series components, targeted features are extracted—including the short-term growth slope of the trend component, the amplitude stability of the periodic component, the volatility of the residual component, and fusion features such as the trend-period coupling factor.

[0162] In the real-time feedback adjustment strategy determination, the deeply optimized features and subsequent prediction results are monitored in real time to determine whether feedback adjustment is triggered: if the determination is "no", the decomposition and optimization process of the current round ends and the extracted time-series features are output to the scenario-based feature construction stage; if the determination is "yes", the closed-loop feedback optimization process is started.

[0163] In the closed-loop feedback optimization process, the core parameters of the decomposition model can be iteratively updated using an online learning module based on recent prediction errors, changes in feature importance, and other data. The sliding window optimizer dynamically adjusts the time window size of the STL / Prophet model based on changes in current user behavior patterns (e.g., income cycles shifting from bi-weekly to monthly). Credit limit recalculation is triggered, recalculating the user's loan amount based on the optimized decomposition model and features. Finally, manual review is initiated; if high-risk features (e.g., significant residual mutations) still exist after adjustments, a manual review process is triggered, with risk control personnel intervening to verify the information.

[0164] Figure 8 This application provides a schematic diagram of the structure of a loan amount prediction device. The device 800 includes an acquisition module 801, a generation module 802, a determination module 803, and a prediction processing module 803.

[0165] The acquisition module 801 is used to acquire multi-source time-series behavioral data of the target user based on the target user's loan request. The multi-source time-series behavioral data is a collection of multi-dimensional time-series data generated by the target user in the financial business scenario.

[0166] The generation module 802 is used to decompose multi-source time-series behavioral data and generate a time-series feature vector corresponding to the target user. The time-series feature vector includes multiple component feature vectors.

[0167] The determination module 803 is used to determine a scenario-based feature vector based on time-series feature vectors and loan domain knowledge. The scenario-based feature vector is used to indicate the repayment ability and risk characteristics of the target user.

[0168] The prediction processing module 803 is used to perform prediction processing on the scenario-based feature vector through a preset prediction model to obtain the predicted loan amount corresponding to the target user.

[0169] In some possible embodiments, the generation module 802 is specifically used for:

[0170] The preset parameters of the dynamic time series decomposition model are updated to the target parameters using the sliding window optimizer.

[0171] By using a dynamic temporal decomposition model, multi-source temporal behavioral data is decomposed based on target parameters to generate temporal feature vectors.

[0172] In some possible embodiments, the preset parameters include a preset window length and preset decomposition parameters; the target parameters include a target window length and a target decomposition parameter; the generation module 802 is specifically used for:

[0173] By using a sliding window optimizer, based on the prediction performance of the feature vectors of each component within the recent window, the preset window length is adjusted to the target window length, and the preset decomposition parameters are adjusted to the target decomposition parameters.

[0174] Specifically, when the preset decomposition parameter is the preset smoothing coefficient, the target decomposition parameter is the preset smoothing coefficient; when the preset decomposition parameter is the preset trend flexibility parameter, the target decomposition parameter is the preset trend flexibility parameter.

[0175] In some possible embodiments, the determining module 803 is specifically used for:

[0176] Determine the income feature vector based on the time-series feature vector;

[0177] Based on the temporal feature vector and knowledge of the loan domain, determine the scenario feature vector;

[0178] The income feature vector and the scene feature vector are concatenated to obtain the scene-specific feature vector.

[0179] In some possible embodiments, the income feature vector includes multiple income sub-vectors; for any given income sub-vector, the determining module 803 is specifically used for:

[0180] Among multiple component feature vectors, determine at least one component feature vector corresponding to the income sub-vector;

[0181] Data processing is performed on at least one component feature vector to obtain the input sub-vector.

[0182] In some possible embodiments, the scene feature vector includes a trend cycle coupling factor and an income coverage ratio; the determination module 803 is specifically used for:

[0183] Based on knowledge of the loan domain, multiple scenario sub-vectors are determined in the time-series feature vector;

[0184] Multiple scene sub-vectors are concatenated to obtain a scene feature vector;

[0185] Among them, several scenario sub-vectors include income coverage ratio, safety cushion days, and scenario adaptation flexibility.

[0186] In some possible embodiments, the preset prediction model includes a decision tree model and a temporal neural network model; the prediction processing module 803 is specifically used for:

[0187] The first intermediate result is obtained by predicting the scenario-based feature vectors using a decision tree model.

[0188] By using a temporal neural network model to predict scene-specific feature vectors, a second intermediate result is obtained.

[0189] The first and second intermediate results are weighted to obtain the predicted loan amount.

[0190] In some possible embodiments, the prediction processing module 803 is specifically used for:

[0191] Based on the multimodal attention mechanism, the first weight corresponding to the decision tree model and the second weight corresponding to the temporal neural network model are determined according to the scenario-based feature vector.

[0192] The sum of the product of the first intermediate result and the first weight, and the product of the second intermediate result and the second weight, is determined as the predicted loan amount.

[0193] In some possible embodiments, the apparatus further includes an adjustment module, which is used to:

[0194] The real-time strategy adjustment engine adjusts the loan amount correction coefficient, model hyperparameters, or triggers a manual review process based on the quantitative indicators of the difference between the predicted loan amount and the actual repayment performance.

[0195] The loan amount prediction device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0196] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Please refer to... Figure 9 The electronic device 900 may include a processor 901 and a memory 902. Exemplarily, the processor 901 and the memory 902 are interconnected via a bus 903.

[0197] Memory 902 stores instructions executed by the computer;

[0198] The processor 901 executes computer execution instructions stored in the memory 902, causing the processor 901 to perform the loan amount prediction method as shown in the above method embodiment.

[0199] Accordingly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the loan amount prediction method of the above method embodiments.

[0200] Accordingly, embodiments of this application may also provide a computer program product, including a computer program, which, when executed by a processor, can implement the loan amount prediction method shown in the above method embodiments.

[0201] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0202] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0203] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0204] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0205] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0206] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0207] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0208] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0209] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting loan amounts, characterized in that, include: Based on the target user's loan request, obtain the target user's multi-source time-series behavioral data, which is a multi-dimensional time-series data set generated by the target user in financial business scenarios. The multi-source time-series behavioral data is decomposed to generate a time-series feature vector corresponding to the target user, and the time-series feature vector includes multiple component feature vectors. Based on the time-series feature vector and loan domain knowledge, a scenario-based feature vector is determined, which is used to indicate the repayment ability and risk characteristics of the target user. By using a preset prediction model, the scenario-based feature vector is processed to predict the loan amount corresponding to the target user.

2. The method according to claim 1, characterized in that, The multi-source time-series behavioral data is decomposed to generate a time-series feature vector corresponding to the target user, including: The preset parameters of the dynamic time series decomposition model are updated to the target parameters using the sliding window optimizer. The dynamic temporal decomposition model decomposes the multi-source temporal behavior data based on the target parameters to generate temporal feature vectors.

3. The method according to claim 2, characterized in that, The preset parameters include a preset window length and preset decomposition parameters; the target parameters include a target window length and a target decomposition parameter. The sliding window optimizer updates the preset parameters of the dynamic time series decomposition model to the target parameters, including: The sliding window optimizer adjusts the preset window length to the target window length and the preset decomposition parameters to the target decomposition parameters based on the prediction performance of the feature vectors of each component within the recent window. Wherein, when the preset decomposition parameter is a preset smoothing coefficient, the target decomposition parameter is a preset smoothing coefficient; when the preset decomposition parameter is a preset trend flexibility parameter, the target decomposition parameter is a preset trend flexibility parameter.

4. The method according to claim 1, characterized in that, Based on the aforementioned time-series feature vectors and loan domain knowledge, scenario-based feature vectors are determined, including: Based on the time-series feature vector, determine the revenue feature vector; Based on the time-series feature vector and the loan domain knowledge, determine the scenario feature vector; The income feature vector and the scene feature vector are concatenated to obtain the scene-specific feature vector.

5. The method according to claim 4, characterized in that, The income feature vector includes multiple income sub-vectors; for any given income sub-vector, the income feature vector is determined based on the time-series feature vector, including: Among the plurality of component feature vectors, at least one component feature vector corresponding to the income sub-vector is determined; Data processing is performed on the at least one component feature vector to obtain the income sub-vector.

6. The method according to claim 4, characterized in that, The scenario feature vector includes a trend cycle coupling factor and an income coverage ratio; Based on the time-series feature vector and the loan domain knowledge, the scenario feature vector is determined, including: Based on the knowledge of the loan domain, multiple scenario sub-vectors are determined in the time-series feature vector; The multiple scene sub-vectors are concatenated to obtain the scene feature vector; The multiple scenario sub-vectors include income coverage ratio, safety cushion days, and scenario adaptation flexibility.

7. The method according to claim 1, characterized in that, The preset prediction model includes a decision tree model and a temporal neural network model; by using the preset prediction model to predict the scenario-based feature vector, the predicted loan amount corresponding to the target user is obtained, including: The decision tree model is used to predict the scenario-based feature vector to obtain a first intermediate result. The temporal neural network model is used to predict the scene-specific feature vector to obtain a second intermediate result. The first intermediate result and the second intermediate result are weighted to obtain the predicted loan amount.

8. The method according to claim 7, characterized in that, The predicted loan amount is obtained by weighting the first intermediate result and the second intermediate result, including: Based on the multimodal attention mechanism, the first weight corresponding to the decision tree model and the second weight corresponding to the temporal neural network model are determined according to the scenario-based feature vector. The sum of the product of the first intermediate result and the first weight, and the product of the second intermediate result and the second weight, is determined as the predicted loan amount.

9. The method according to claim 1, characterized in that, After obtaining the predicted loan amount for the target user by performing prediction processing on the scenario-based feature vector using a preset prediction model, the method further includes: The real-time strategy adjustment engine adjusts the loan amount correction coefficient, model hyperparameters, or triggers a manual review process based on the quantitative indicators of the difference between the predicted loan amount and the actual repayment performance.

10. A loan amount prediction device, characterized in that, It includes an acquisition module, a generation module, a determination module, and a prediction processing module: The acquisition module is used to acquire multi-source time-series behavioral data of the target user based on the target user's loan request. The multi-source time-series behavioral data is a set of multi-dimensional time-series data generated by the target user in financial business scenarios. The generation module is used to decompose the multi-source time-series behavioral data to generate a time-series feature vector corresponding to the target user, wherein the time-series feature vector includes multiple component feature vectors. The determining module is used to determine a scenario-based feature vector based on the time-series feature vector and loan domain knowledge. The scenario-based feature vector is used to indicate the repayment ability and risk characteristics of the target user. The prediction processing module is used to perform prediction processing on the scenario-based feature vector through a preset prediction model to obtain the predicted loan amount corresponding to the target user.

11. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 9.

Images