A new case repayment prediction model construction method based on historical repayment

Abstract

This invention belongs to the field of financial data processing technology and discloses a method for constructing a new case repayment prediction model based on historical repayments. The method includes the following steps: S1, acquiring multi-source data of historical cases and fusing the multi-source data to generate an initial unified feature vector for the case; S2, inputting the initial unified feature vector into a shared feature extraction network, and extracting a stable feature subspace representing the debtor's inherent repayment ability and a dynamic feature subspace representing the debtor's sensitivity to collection strategies through two parallel feature decoupling subnetworks; S3, constructing a dual-branch ensemble model containing an inherent repayment prediction branch and a strategy gain prediction branch; S4, using a course learning strategy to train the dual-branch ensemble model in stages; S5, deploying the trained dual-branch ensemble model as a prediction service, simultaneously outputting the basic repayment prediction value and the expected repayment gain value of each candidate strategy for new cases, and recommending the strategy with the highest overall expected value.

Description

Technical Field

[0001] This invention relates to the field of financial data processing technology, specifically to a method for constructing a new case repayment prediction model based on historical repayments. Background Technology

[0002] In the fields of financial institutions, non-performing asset management, and debt collection, accurately predicting the future recovery probability and amount of newly accepted cases is crucial for achieving precise resource allocation, scientific performance evaluation, and effective risk management. Currently, the industry's commonly used forecasting methods mainly rely on existing installment repayment schedules for cases, estimating the remaining recoverable amount through linear extrapolation. However, observations from actual business data show that a large number of repayments do not strictly follow installment plans but rather originate from debtors' single full or large-amount repayments. This "one-time repayment" pattern has significant non-linear and sporadic characteristics, making it difficult to effectively capture and predict using simple extrapolation methods based on schedules.

[0003] More significantly, existing forecasting methods generally suffer from a core flaw: they fail to effectively decouple the debtor's inherent repayment ability from the recovery gains brought about by external collection strategies. The models cannot distinguish whether the recovery effect is determined by the debtor's own creditworthiness or driven by external collection strategies, thus failing to quantify the actual improvement effect of different strategies on recovery. These methods can only output single, static recovery forecasts, often exhibiting a "black box" structure, lacking interpretability and strategy recommendation capabilities, and unable to provide directly executable optimal collection strategy guidance for business operations. Furthermore, the models do not support adaptive iterative optimization, making it difficult to adapt to dynamic changes in the market environment, debtor behavior, and collection strategies, ultimately leading to a significant disconnect between forecast results and actual business needs. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a new case repayment prediction model based on historical repayments, in order to solve the problem that existing technologies cannot distinguish between the debtor's inherent repayment ability and the gains from collection strategies, and thus cannot achieve accurate strategy recommendations and continuous model optimization.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for constructing a new loan repayment prediction model based on historical repayments includes the following steps: S1. Obtain multi-source data of historical cases and fuse the multi-source data to generate an initial unified feature vector for the cases; S2. Input the initial unified feature vector into the shared feature extraction network. Through two parallel feature decoupling subnetworks, extract the stable feature subspace of the debtor's inherent repayment ability and the dynamic feature subspace of the debtor's sensitivity to collection strategies, respectively. Make the two subspace representations independent of each other through adversarial training constraints. S3. Construct a dual-branch ensemble model that includes an inherent repayment prediction branch and a strategy gain prediction branch. The inherent repayment prediction branch takes a stable feature subspace as input and outputs the basic repayment prediction value without intervention. The strategy gain prediction branch takes a dynamic feature subspace and candidate strategy encoding as input and outputs the expected repayment gain value of each candidate strategy. S4. The dual-branch ensemble model is trained in stages using a course learning strategy. S5. Deploy the trained dual-branch ensemble model as a prediction service, simultaneously output the basic repayment prediction value and the expected repayment gain value of each candidate strategy for new cases, and recommend the strategy with the highest comprehensive expected value.

[0006] Preferably, the multi-source data includes static attributes of historical cases, time series of historical repayment behaviors, and collection operation records, wherein each time step of the time series of historical repayment behaviors includes a normalized repayment amount, an identifier indicating whether the repayment behavior is overdue, and the number of overdue days.

[0007] Preferably, the fusion process of the initial unified feature vector in step S1 is as follows: S11. Perform one-hot encoding on the categorical variables in the static attributes and map them into dense vectors; standardize the continuous variables. S12. Input the historical repayment behavior time sequence into the gated recurrent unit network and extract the sequence feature vector representing the repayment behavior pattern; S13. The collection operation record is converted into statistical features, and then spliced ​​and fused with the dense vector, the sequence feature vector and the standardized continuous features to form the initial unified feature vector.

[0008] Preferably, the adversarial training constraint in step S2 uses a discriminator, which takes the feature vector as input and outputs the probability that it comes from the dynamic feature subspace. During training, the discriminator is used to accurately distinguish the source of the feature. The shared feature extraction network and two feature decoupling subnetworks are used to make the source judgment probability of the discriminator approach the random guess probability. The discriminator loss function is: ,in, For cross-entropy loss, To determine the number of feature vectors in the training batch, For which eigenvector is it? For the first One input feature vector, For feature source tags, Predict probabilities for the discriminator; The adversarial loss function is: ,in, To combat the losses, To stabilize the feature subspace, For dynamic feature subspace, For feature vectors from a stable feature subspace, These are the feature vectors from the dynamic feature subspace.

[0009] Preferably, step S3 specifically includes: S31. Construct an inherent repayment prediction branch, taking the stable feature subspace as input, and output the basic repayment prediction value of the case without special intervention through the first fully connected neural network. S32. Construct a policy gain prediction branch and calculate the attention score between the query vector formed by the dynamic feature subspace and each key vector in the parameterized candidate policy encoding matrix: ,in, For query vectors, The key vector matrix, The dimension of each policy key vector, Key vector matrix The transpose of the matrix, its square root Used for dimensionality normalization of attention scores For normalized activation functions, This is the normalized attention score vector; S33. Normalize the attention score to obtain the attention weight of each candidate strategy. Based on the attention weight, perform a weighted summation on the value vectors in the parameterized candidate strategy encoding matrix to obtain the fused strategy context representation. S34. Input the strategy context representation into the second fully connected neural network and output the expected return gain value corresponding to each candidate strategy.

[0010] Preferably, the step-by-step training in step S4 specifically includes: S41. In the first stage, only the inherent repayment prediction branch and the corresponding shared feature extraction network and stable feature decoupling sub-network are trained, and the parameters of the strategy gain prediction branch are frozen; the training objective is to minimize the mean square error loss between the basic repayment prediction value and the historical real repayment value. S42. In the second stage, all model parameters are unfrozen, and end-to-end joint training is performed using a multi-task loss function. The calculation formula for the multi-task loss function is as follows: , , , ,in, For multi-task loss function, Based on the prediction of loss, For strategy gain loss, To decouple and maintain loss, , , These are the weighting coefficients of the three losses, and they satisfy... ; Mean square error, For cross-entropy loss, Based on the basic collection forecast, This represents the actual amount recovered for historical cases under the baseline strategy. For the first The expected return on investment for each candidate strategy. For the first The true gain value of each candidate strategy, where P is the prediction probability of the fixed discriminator for the feature source, and 0.5 is the target value for the random guess probability.

[0011] Preferably, the formula for calculating the true gain value corresponding to the strategy gain loss is: ,in, The actual gain value of historical case c after adopting strategy s is the strategy gain loss. The true label; The actual amount recovered after adopting strategy s in historical case c is derived from the real records in the time-series data of historical recovery behavior; c: A single historical case sample, whose stable feature subspace is ; s: a specific candidate collection strategy; c': Similar to case c in terms of stable characteristics The set of most similar historical cases, where c' does not use strategy s or only uses the baseline strategy; The actual amount recovered in a single case within case set c'; The average actual amount recovered in case set c', i.e., the total amount recovered by all cases within the set. The arithmetic mean.

[0012] Preferably, step S5 specifically includes: S51. Deploy the trained dual-branch ensemble model as a prediction service, integrating Monte Carlo Dropout layers into both the inherent repayment prediction branch and the strategy gain prediction branch. S52. During model inference, multiple forward propagation samples are used to calculate the predicted mean and predicted variance of the basic repayment prediction value and the expected repayment gain value of each candidate strategy. Results with a predicted variance higher than a preset threshold are marked as high uncertainty predictions, and confidence intervals are output or manual review suggestions are triggered. S53. For new cases, simultaneously output the basic recovery forecast value and the expected recovery gain value of each candidate strategy; S54. Using counterfactual reasoning, fix the stable feature subspace of the case and encode the virtual substitution strategy, and calculate the counterfactual prediction results of the key substitution strategy; S55. Generate an interpretable report that includes a comparison of the recommended strategy and the predictions of key alternative strategies, and recommend the strategy with the highest overall expected value.

[0013] Preferably, the formulas for calculating the predicted mean and predicted variance in step S52 are as follows: , ,in, To predict the mean, This represents the total number of forward propagation samples during model inference. For the first Second sampling, The prediction result for the t-th sampling is... To predict variance, when For a preset variance threshold, it is marked as a high-uncertainty prediction and a confidence interval or a suggestion to trigger manual review is provided.

[0014] Preferably, step S6 further includes: S61. Regularly monitor the actual repayment effect of the case recommendation strategy based on the model, and calculate the average absolute error for a single period: ,in, The mean absolute error, This represents the number of cases that employed the target strategy during that period. For the first time in the period One case, For the first The expected gain value for each case, For the first The actual gain value for each case; S62. When the average absolute error exceeds a preset threshold for multiple consecutive statistical periods, it is determined that there is a continuous deviation between the predicted gain and the actual payment collection effect, and targeted incremental learning is automatically triggered. The incremental learning adopts an elastic weight consolidation algorithm, and a regularization term is added to the loss function of the incremental learning. The regularization term is expressed as: ,in, For the model number The values ​​of the parameters to be updated in the current incremental learning task. For the model number The optimal values ​​of each parameter are saved after training for the historical task. For the model number The importance weight of each parameter The total loss is calculated by summing all parameters involved in the incremental learning of the model to form the regularization term.

[0015] By adopting the above technical solution, the present invention has the following advantages compared with the prior art: 1. This invention provides a method for constructing a new loan repayment prediction model based on historical repayment data. By building a dual-branch integrated model, the inherent repayment prediction branch provides the base value, while the strategy gain prediction branch quantifies the added value brought by different strategies through a cross-attention mechanism. The combined model can directly output a recommendation to adopt a certain strategy and predict the total repayment, seamlessly transforming the prediction results into immediately executable, differentiated, and tiered strategy suggestions, thus completely connecting the link from data insight to business action.

[0016] 2. This invention provides a method for constructing a new loan repayment prediction model based on historical repayment data. Through phased learning and multi-task joint training including decoupling maintenance loss, it ensures stable convergence of complex models while maintaining decoupling characteristics. After deployment, the system continuously monitors the deviation between predictions and actual results and automatically triggers targeted incremental learning based on algorithms such as elastic weight consolidation. This allows the model to continuously evolve with the iteration of business strategies and changes in the market environment, forming a business empowerment closed loop and avoiding the problem of performance decay in traditional static models.

[0017] 3. This invention provides a method for constructing a new case repayment prediction model based on historical repayments. It quantifies the repayment gain of each strategy based on a cross-attention mechanism, combines the basic repayment value with the strategy gain value, and directly outputs the collection strategy with the highest overall expectation. This opens up the decision-making link from prediction to business execution and improves the efficiency of resource allocation. Monte Carlo Dropout realizes the quantification of prediction uncertainty, marks and warns high-risk cases, and generates interpretable reports by combining counterfactual reasoning, making the model transform from a "black box" into an understandable and traceable decision-making tool.

[0018] 4. This invention provides a method for constructing a new case repayment prediction model based on historical repayment records. It introduces a task-driven adversarial training mechanism, decoupling case features into stable features reflecting inherent repayment ability and dynamic features reflecting strategy sensitivity. This not only makes the model easier to understand and train but also distinguishes whether poor repayment is due to debtor credit issues or strategy mismatch, thus supporting accurate root cause identification. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Example Please refer to Figure 1 As shown, this invention discloses a method for constructing a new loan repayment prediction model based on historical repayments, comprising the following steps: S1. Obtain multi-source data of historical cases and merge the multi-source data into an initial unified feature vector of the case. The multi-source data includes the static attributes of historical cases, the time sequence of historical repayment behavior, and collection operation records.

[0022] Static attributes include inherent information such as the debtor's age, occupation, credit score, number of past defaults, principal amount in the case, and duration of debt. These data can be divided into categorical variables (such as occupation type) and continuous variables (such as age).

[0023] Historical repayment behavior timeline: Records the sequence of repayment events for each case in chronological order. The feature vector for each time step includes the normalized repayment amount, an indicator (Boolean value) indicating whether the repayment behavior is overdue, and the number of overdue days (0 if not overdue).

[0024] The normalized formula for repayment amount is: ,in The amount is the normalized amount. This refers to the actual amount repaid in a single transaction. This represents the total outstanding principal amount in the case.

[0025] Collection Operation Records: Record all operations taken for the case, such as telephone contact (time, duration, result), SMS sending, issuance and feedback of mediation plans, and records of field visits.

[0026] The fusion process of the initial unified feature vector in step S1 is as follows: S11. Perform one-hot encoding on the categorical variables in the static attributes and map them to dense vectors through an embedding layer. Standardize the continuous variables using the Z-score standardization formula. ,in, The standardized value. The original value of the variable. This represents the mean of the variable across the historical dataset. This represents the standard deviation of the variable across the historical dataset.

[0027] S12. Input the historical repayment behavior time series data into a gated recurrent unit network (GRU). The GRU can effectively capture long-term dependencies, and the hidden state of its last time step is extracted as a sequence feature vector representing the entire historical repayment behavior pattern of the case.

[0028] GRU extracts sequence feature vectors representing repayment behavior patterns through the following core calculations: Reset door and update door calculations: , ,in, : Reset the gate output at all times. : Update the gate output constantly. Sigmoid activation function , Weight matrix : Hide your status at all times. : Input feature vector at any time, , : Bias term; Calculation of candidate hidden state and final hidden state: , ,in, Candidate hidden state Hyperbolic tangent activation function Weight matrix : Hadamaji, : Bias term; The hidden state of the GRU network at the last moment ( (where the time series length is 1) is the sequence feature vector representing the repayment behavior pattern; S13. Aggregate collection operation records and convert them into statistical features, such as: the number of contacts in the past 7 days, the total duration of all historical contacts, and the proportion of different contact results (such as promise to repay or loss of contact).

[0029] The dense vector, sequence feature vector, statistical features, and standardized continuous features obtained from the above steps are concatenated and fused along the feature dimension to form a high-dimensional, information-rich initial unified feature vector.

[0030] S2. Input the initial uniform feature vector into a shared feature extraction network, such as a multilayer perceptron. MLP Initial feature transformation is performed. Subsequently, two structurally identical but parameter-independent feature decoupling subnetworks (also known as...) are used. MLP The two subspaces are processed in parallel to output stable feature subspaces and dynamic feature subspaces, respectively.

[0031] The stable feature subspace aims to extract relatively stable features related to the debtor's inherent repayment ability and willingness, such as credit foundation, income stability, and historical habit of fulfilling obligations. The dynamic feature subspace aims to extract features that reflect the debtor's sensitivity to external collection strategies, such as the speed of their response to telephone reminders and changes in their acceptance of mediation plans.

[0032] To achieve effective decoupling, a discriminator D is introduced (the discriminator is a simple classification network, such as several fully connected layers with a sigmoid output). The input to the discriminator D is a feature vector from a stable feature subspace or a dynamic feature subspace, and its output is a scalar between 0 and 1, representing the probability that the vector comes from the dynamic feature subspace.

[0033] The training objective of the discriminator is to distinguish between input features originating from a stable feature subspace and a dynamic feature subspace as accurately as possible. Its loss function is: ,in, For cross-entropy loss, To determine the number of feature vectors in the training batch, For which eigenvector is it? For the first One input feature vector, For feature source tags, This indicates that it comes from the dynamic feature subspace. This indicates that it comes from a stable feature subspace. Predict probabilities for the discriminator; One of the training objectives of the shared feature extraction network and the two decoupled sub-networks is to jointly optimize the adversarial loss with the subsequent prediction task objective, thereby deceiving the discriminator into failing to distinguish the source of the features. The adversarial loss function is: ,in, To combat the losses, To stabilize the feature subspace, For dynamic feature subspace, For feature vectors from a stable feature subspace, These are the feature vectors from the dynamic feature subspace.

[0034] Ultimately, the discriminator's prediction probability of the feature source approaches 0.5 (random guess). Through the aforementioned adversarial game, the stable feature subspace and the dynamic feature subspace are forced to learn non-overlapping and independent information representations, thereby achieving feature decoupling.

[0035] S3. Construct a two-branch ensemble model that includes an inherent repayment prediction branch and a strategy gain prediction branch, and perform basic repayment prediction and strategy gain prediction respectively.

[0036] The inherent repayment prediction branch takes a stable feature subspace as input and outputs a scalar through a first fully connected neural network (which may contain multiple hidden layers and activation functions). This scalar is the basic repayment prediction value of the case under no special intervention or by using only the conventional benchmark strategy. The inherent repayment prediction branch assesses the debtor's fundamental situation.

[0037] Using the dynamic feature subspace as the query vector and the parameterized candidate strategy encoding matrix as the key and value, the matching weight between the strategy and the case is calculated through the cross-attention mechanism, and the expected repayment gain value corresponding to each candidate strategy is output through the second fully connected neural network. The digitized candidate strategy encoding matrix is ​​generated by the candidate collection strategy through standardized encoding and is matched with the feature vector dimension.

[0038] The strategy gain prediction branch is used to quantify the value of different collection strategies. First, a series of candidate collection strategies (such as high-frequency telephone collection, gentle SMS reminders, installment settlement plans, etc.) are parameterized and encoded to form a candidate strategy encoding matrix E, where each row represents the encoding vector of a strategy.

[0039] The policy gain prediction branch uses the dynamic feature subspace Zd as the query vector (Q) and the policy encoding matrix E as the key (K) and value (V). It normalizes the attention scores to obtain the attention weights for each candidate policy. The specific process is as follows: 1) Calculate the attention score between the query vector formed by the dynamic feature subspace and each key vector in the parameterized candidate policy encoding matrix, and then calculate the attention score: ,in, (Query, query vector): composed of dynamic feature subspace Directly constituted, with dimensions as , It is a feature dimension of the dynamic feature subspace, used to characterize the sensitivity of the current case to collection strategies; (Key, key vector matrix): This refers to the key portion of the parameterized candidate policy encoding matrix, with dimensions of... , The total number of candidate collection strategies; : The dimension of each policy key vector (satisfying) To ensure the effectiveness of matrix multiplication, and to characterize the core features of each candidate strategy; : key vector matrix The transpose of the matrix, with dimension Used with query vectors Perform matrix multiplication to calculate similarity; : Dimensions of query vector If they are equal, their square roots This is used to normalize the attention score to avoid numerical overflow or gradient vanishing problems caused by excessive dimensionality. Normalized activation function: used to transform the raw score after matrix multiplication into... The probability distribution within the interval ensures that the sum of the attention scores of each candidate strategy is 1, and the output dimension is... ; The normalized attention score vector, with dimension 1. , of which Each element represents the degree of match between the i-th candidate strategy and the current case.

[0040] 2) The value vector V (i.e. the policy encoding itself) in the parameterized candidate policy encoding matrix is ​​weighted and summed according to the attention weight to obtain the policy context representation C that integrates the dynamic features of the case and the policy information; 3) Input the policy context representation C into a second fully connected neural network, which outputs a vector where each element corresponds to a candidate policy. Expected return on investment The gain value represents the additional amount that can be recovered by using this candidate strategy compared to the baseline strategy.

[0041] 4) For a new case, a candidate strategy is adopted. The overall expected return is: ,in, The current case adopts the first The comprehensive expected return value (in yuan) after each candidate strategy is one of the core indicators of the final output of the model. (Base Repayment Forecast): Output from the inherent repayment forecast branch, representing the expected repayment amount for the case without special intervention or by using only the baseline strategy; (No. The expected return gain of the first strategy): output by the strategy gain prediction branch, representing the return gain compared to the benchmark strategy. The additional expected return amount after each candidate strategy can be positive or negative (a negative gain indicates that the strategy is inferior to the benchmark strategy).

[0042] S4. A course learning strategy from easy to difficult is adopted to distribute the training of the dual-branch ensemble model to ensure stable convergence of the model. In the first stage, only the inherent repayment prediction branch and the related shared feature extraction network and stable feature decoupling sub-network are trained. The training objective is to minimize the loss between the basic repayment prediction value and the historical real repayment value. In the second stage, all model parameters are unfrozen, and the training objective is defined by the multi-task loss function and end-to-end joint training is performed.

[0043] S41. The first stage is the pre-training stage, where only the inherent repayment prediction branch is enabled, and all parameters of the policy gain prediction branch are frozen. Simultaneously, the feature decoupling part retains only the path leading to Zs, and the training objective is to minimize the base repayment prediction value Y.base Compared with the actual amount recovered in historical cases Y true The loss function L between base (e.g., mean squared error, MSE). This stage enables the shared feature extraction network and the stable feature decoupling sub-network to initially learn to extract effective features related to inherent repayment ability.

[0044] After the first phase of training is completed, the discriminator used for adversarial training also reaches a relatively stable state, and can better distinguish the features of initial decoupling.

[0045] S42. The second stage is the joint training stage, where all model parameters are unfrozen, including the policy gain prediction branch and the dynamic feature decoupling subnetwork, for end-to-end joint training. The training objective of the second stage is defined by a multi-task loss function. Based on predicted loss Strategy gain and loss and decoupling maintenance loss Weighted sum: ,in, , , These are the weight coefficients for the three loss terms, dynamically adjusted based on the model's training convergence performance, with values ​​ranging from [0, 1] and satisfying... This makes the present embodiment more feasible.

[0046] Basic predicted loss The loss term used to optimize the basic collection forecast is calculated using the mean squared error of MSE, i.e. , This represents the actual amount recovered for historical cases under the benchmark strategy.

[0047] Strategy gain loss The loss term used to optimize the predicted strategy gain is calculated using the mean squared error (MSE). , For the first The true gain value of each strategy.

[0048] in, Mean squared error (MSE) is used to measure the squared mean of the error between the predicted and the actual values. Based on the basic collection forecast, This represents the actual amount recovered for historical cases under the baseline strategy. For the first The expected return on investment for each candidate strategy. Let be the true gain value of the i-th candidate strategy.

[0049] For example, the true gain value of a certain historical data case is calculated as follows: ,in, The actual gain value of historical case c after adopting strategy s is the strategy gain loss. The true label; The actual amount recovered after adopting strategy s in historical case c is derived from the real records in the time-series data of historical recovery behavior; c: A single historical case sample, whose stable feature subspace is ; s: A specific candidate collection strategy (such as high-frequency telephone collection, installment settlement plan, etc.); c': Similar to case c in terms of stable characteristics The set of most similar historical cases (similarity criteria are Euclidean distance less than a preset threshold, or Top-K most similar samples, where K ranges from 5 to 20), and c' does not use strategy s or only uses the baseline strategy; The actual amount recovered in a single case within case set c'; The average actual amount recovered in case set c', i.e., the total amount recovered by all cases within the set. The arithmetic mean; Decoupling maintenance loss Used to maintain a stable feature subspace With dynamic feature subspace The independence loss term is calculated using cross-entropy loss, i.e. ,in, denoted as cross-entropy loss, used to measure the difference in probability distributions. P is the probability predicted by the discriminator for the source of the feature, and 0.5 is the target value for random guess probability, used to constrain the discriminator from being unable to distinguish the source of the feature.

[0050] The parameters of the discriminator D trained in the first stage are fixed. In the second stage of training, features from both the stable and dynamic feature subspaces are input into this fixed discriminator D to obtain the predicted probability P. Decoupling maintenance loss is then applied. Designed to make the discriminator D more difficult to judge, for example, defined as cross-entropy loss, but with the goal of making the predicted probability P close to 1 for stable feature subspace inputs (or close to 0 for dynamic feature subspace inputs), or directly minimizing the discriminator's output entropy. Decoupling maintenance loss. This ensures that the independence of the stable feature subspace and the dynamic feature subspace is not compromised when jointly training complex tasks.

[0051] S5. Deploy the trained dual-branch ensemble model as a prediction service. Based on the prediction service, output the basic repayment prediction value and the expected repayment gain value of each candidate strategy for new cases simultaneously, and recommend the strategy with the highest comprehensive expected value.

[0052] To enhance the reliability and transparency of decision-making, this implementation integrates functions for quantifying forecast uncertainty and generating interpretable reports.

[0053] Prediction uncertainty quantification: Monte Carlo Dropout is integrated into the fully connected layers of both prediction branches, and multiple forward propagation samplings are performed during model inference (number of samplings). (e.g., 50 times) The predicted mean and predicted variance of the base collection forecast and the expected collection gain of each candidate strategy are calculated using the following formula: Predicted mean: Prediction variance: ,in, To predict the mean, This represents the total number of forward propagation samples during model inference. For the first Second sampling, The prediction result for the t-th sampling is... To predict variance, when , If a preset variance threshold is set, such as the 90th percentile of the historical prediction variance, it is marked as a high uncertainty prediction, and a confidence interval is provided or a manual review suggestion is triggered.

[0054] Interpretability Report Generation: For the optimal strategy A recommended by the model, the system performs counterfactual reasoning. Specifically, it fixes all features of the current case, virtually replaces the strategy code with the code of another key strategy B (such as a suboptimal strategy or a historically common strategy), and reruns the strategy gain prediction branch to obtain the gain value and combined value for "assuming strategy B is adopted." Subsequently, the system automatically generates an interpretable report. The core content of the report is: "Strategy A is recommended for you, with an estimated return of [X] yuan. If strategy B is adopted, the estimated return will increase or decrease by [Y] yuan." This allows business personnel to intuitively understand the basis for the recommendation and make a final decision based on experience.

[0055] S6. Monitor the actual repayment effect of the recommended strategy. When there is a persistent deviation between the predicted gain and the actual repayment effect, automatically trigger targeted incremental learning of the strategy gain prediction branch. The quantitative criterion for persistent deviation is: within three consecutive statistical periods (e.g., three weeks), the average absolute error between the predicted gain and the actual gain in each period exceeds a preset threshold. The specific formula is: Mean absolute error per single period: ,in, Mean absolute error The number of cases employing the target strategy during this period. For the first time in the period One case, : No. The expected gain value for each case, : No. The actual gain value of each case ( ); Persistent Deviation Determination: ,in, Current statistical period : No. Mean absolute error of the period : Preset threshold coefficient, value range , The historical average actual gain of this strategy; For a large number of cases where the recommended strategy was implemented, the system periodically (e.g., weekly) calculates the average predicted gain and average actual gain (actual revenue - benchmark revenue for similar cases) of the strategy. When a persistent and statistically significant deviation is found between the predicted and actual values ​​of a certain strategy or strategy, the system determines that the model's gain prediction for that strategy may no longer be suitable for the latest business environment.

[0056] At this point, the system automatically triggers targeted incremental learning, using only newly generated data relevant to the biased policy to fine-tune the policy gain prediction branch (or the further associated dynamic feature decoupling subnetwork). To prevent the forgetting of previously learned knowledge when adapting to new data, this implementation employs an elastic weight consolidation algorithm. This algorithm calculates the importance weights of each parameter in the network in historical tasks (approximately using the Fischer information matrix) and adds a regularization term to the loss function of incremental learning. ,in, Model 1 The values ​​to be updated (i.e., new parameters) of each parameter in the current incremental learning task; Model 1 The optimal values ​​of each parameter (i.e., old parameters) are saved after training of the historical task is completed. Model 1 The importance weights of each parameter are approximated by the Fischer information matrix, and the values ​​range from [0, +∞). The larger the value, the more significant the impact of the parameter on the performance of historical tasks (such as the first stage of pre-training and the early joint training). The total loss of the regularization term is formed by summing all parameters of the model involved in incremental learning (mainly the parameters of the policy gain prediction branch and the dynamic feature decoupling sub-network).

[0057] When performing incremental learning on new data, for each parameter If its importance weight If the value is very large (meaning this parameter is important for remembering historical knowledge), then we will deviate it from its historical optimum. Punishment will be based on the degree of severity. Punishment items This will be added to the loss function. In this way, while the model adapts to new data, it is flexibly constrained from making drastic changes to important old parameters, thus effectively mitigating catastrophic forgetting.

[0058] This closed-loop model can continuously evolve with market changes, shifts in debtor behavior, and iterations in collection strategies, maintaining the long-term effectiveness of its predictive capabilities and its value in guiding business.

[0059] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a new loan repayment prediction model based on historical repayments, characterized in that: Includes the following steps: S1. Obtain multi-source data of historical cases and fuse the multi-source data to generate an initial unified feature vector for the cases; S2. Input the initial unified feature vector into the shared feature extraction network. Through two parallel feature decoupling subnetworks, extract the stable feature subspace of the debtor's inherent repayment ability and the dynamic feature subspace of the debtor's sensitivity to collection strategies, respectively. Make the two subspace representations independent of each other through adversarial training constraints. S3. Construct a dual-branch ensemble model that includes an inherent repayment prediction branch and a strategy gain prediction branch. The inherent repayment prediction branch takes a stable feature subspace as input and outputs the basic repayment prediction value without intervention. The strategy gain prediction branch takes a dynamic feature subspace and candidate strategy encoding as input and outputs the expected repayment gain value of each candidate strategy. S4. The dual-branch ensemble model is trained in stages using a course learning strategy. S5. Deploy the trained dual-branch ensemble model as a prediction service, simultaneously output the basic repayment prediction value and the expected repayment gain value of each candidate strategy for new cases, and recommend the strategy with the highest comprehensive expected value.

2. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 1, characterized in that: The multi-source data includes static attributes of historical cases, time series of historical repayment behaviors, and collection operation records. Each time step of the time series of historical repayment behaviors includes a normalized repayment amount, an indicator of whether the repayment behavior is overdue, and the number of overdue days.

3. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 2, characterized in that, The fusion process of the initial unified feature vector in step S1 is as follows: S11. Perform one-hot encoding on the categorical variables in the static attributes and map them into dense vectors; standardize the continuous variables. S12. Input the historical repayment behavior time sequence into the gated recurrent unit network and extract the sequence feature vector representing the repayment behavior pattern; S13. The collection operation record is converted into statistical features, and then spliced ​​and fused with the dense vector, the sequence feature vector and the standardized continuous features to form the initial unified feature vector.

4. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 1, characterized in that: The adversarial training constraint mentioned in step S2 uses a discriminator, which takes the feature vector as input and outputs the probability that it comes from the dynamic feature subspace. During training, the discriminator is used to accurately distinguish the source of the feature. The shared feature extraction network and two feature decoupling subnetworks are used to make the discriminator's source judgment probability approach the random guess probability. The discriminator loss function is: ,in, For cross-entropy loss, To determine the number of feature vectors in the training batch, For which eigenvector is it? For the first One input feature vector, For feature source tags, Predict probabilities for the discriminator; The adversarial loss function is: ,in, To combat the losses, To stabilize the feature subspace, For dynamic feature subspace, For feature vectors from a stable feature subspace, These are the feature vectors from the dynamic feature subspace.

5. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 1, characterized in that, Step S3 is as follows: S31. Construct an inherent repayment prediction branch, taking the stable feature subspace as input, and output the basic repayment prediction value of the case without special intervention through the first fully connected neural network. S32. Construct a policy gain prediction branch and calculate the attention score between the query vector formed by the dynamic feature subspace and each key vector in the parameterized candidate policy encoding matrix: ,in, For query vectors, The key vector matrix, The dimension of each policy key vector, Key vector matrix The transpose of the matrix, its square root Used for dimensionality normalization of attention scores For normalized activation functions, This is the normalized attention score vector; S33. Normalize the attention score to obtain the attention weight of each candidate strategy. Based on the attention weight, perform a weighted summation on the value vectors in the parameterized candidate strategy encoding matrix to obtain the fused strategy context representation. S34. Input the strategy context representation into the second fully connected neural network and output the expected return gain value corresponding to each candidate strategy.

6. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 1, characterized in that, The step-by-step training described in step S4 specifically includes: S41. In the first stage, only the inherent repayment prediction branch and the corresponding shared feature extraction network and stable feature decoupling sub-network are trained, and the parameters of the strategy gain prediction branch are frozen; the training objective is to minimize the mean square error loss between the basic repayment prediction value and the historical real repayment value. S42. In the second stage, all model parameters are unfrozen, and end-to-end joint training is performed using a multi-task loss function. The calculation formula for the multi-task loss function is as follows: , , , ,in, For multi-task loss function, Based on the prediction of loss, For strategy gain loss, To decouple and maintain loss, , , These are the weighting coefficients of the three losses, and they satisfy... ; Mean square error, For cross-entropy loss, Based on the basic collection forecast, This represents the actual amount recovered for historical cases under the baseline strategy. For the first The expected return on investment for each candidate strategy. For the first The true gain value of each candidate strategy, where P is the prediction probability of the fixed discriminator for the feature source, and 0.5 is the target value for the random guess probability.

7. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 6, characterized in that: The formula for calculating the true gain value corresponding to the strategy gain loss is as follows: ,in, The actual gain value of historical case c after adopting strategy s is the strategy gain loss. The true label; The actual amount recovered after adopting strategy s in historical case c is derived from the real records in the time-series data of historical recovery behavior; c: A single historical case sample, whose stable feature subspace is ; s: a specific candidate collection strategy; c': Similar to case c in terms of stable characteristics The set of most similar historical cases, where c' does not use strategy s or only uses the baseline strategy; The actual amount recovered in a single case within case set c'; The average actual amount recovered in case set c', i.e., the total amount recovered by all cases within the set. The arithmetic mean.

8. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 1, characterized in that, Step S5 is as follows: S51. Deploy the trained dual-branch ensemble model as a prediction service, integrating Monte Carlo Dropout layers into both the inherent repayment prediction branch and the strategy gain prediction branch. S52. During model inference, multiple forward propagation samples are used to calculate the predicted mean and predicted variance of the basic repayment prediction value and the expected repayment gain value of each candidate strategy. Results with a predicted variance higher than a preset threshold are marked as high uncertainty predictions, and confidence intervals are output or manual review suggestions are triggered. S53. For new cases, simultaneously output the basic recovery forecast value and the expected recovery gain value of each candidate strategy; S54. Using counterfactual reasoning, fix the stable feature subspace of the case and encode the virtual substitution strategy, and calculate the counterfactual prediction results of the key substitution strategy; S55. Generate an interpretable report that includes a comparison of the recommended strategy and the predictions of key alternative strategies, and recommend the strategy with the highest overall expected value.

9. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 8, characterized in that, The formulas for calculating the predicted mean and predicted variance in step S52 are as follows: , ,in, To predict the mean, This represents the total number of forward propagation samples during model inference. For the first Second sampling, The prediction result for the t-th sampling is... To predict variance, when For a preset variance threshold, it is marked as a high-uncertainty prediction and a confidence interval or a suggestion to trigger manual review is provided.

10. The method for constructing a new loan repayment prediction model based on historical repayments as described in claim 1, characterized in that: It also includes step S6, which specifically includes: S61. Regularly monitor the actual repayment effect of the case recommendation strategy based on the model, and calculate the average absolute error for a single period: ,in, The mean absolute error, This represents the number of cases that employed the target strategy during that period. For the first time in the period One case, For the first The expected gain value for each case, For the first The actual gain value for each case; S62. When the average absolute error exceeds a preset threshold for multiple consecutive statistical periods, it is determined that there is a continuous deviation between the predicted gain and the actual payment collection effect, and targeted incremental learning is automatically triggered. The incremental learning adopts an elastic weight consolidation algorithm, and a regularization term is added to the loss function of the incremental learning. The regularization term is expressed as: ,in, For the model number The values ​​of the parameters to be updated in the current incremental learning task. For the model number The optimal values ​​of each parameter are saved after training for the historical task. For the model number The importance weight of each parameter The total loss is calculated by summing all parameters involved in the incremental learning of the model to form the regularization term.

Images