A method and system for predicting effects for bank retail business

Abstract

The application discloses a bank retail business effect prediction method and system, relates to the technical field of intelligent marketing prediction, and comprises the following steps: acquiring multi-source business data, uniformly mapping and processing the multi-source business data, and constructing a unified data base; constructing a business income joint label based on the unified data base, and combining the settlement confirmation state and risk factors in the unified data base to perform delay feedback correction on the business income joint label; performing feature construction on the unified data base, generating a multi-dimensional joint feature representation, and constructing a multi-model prediction model based on the multi-dimensional joint feature representation and the business income joint label that is corrected through delay feedback, so that the customer behavior and business income are jointly predicted. The application constructs a unified data base and a delay feedback correction mechanism, fuses multi-task prediction and risk constraint optimization, significantly improves the accuracy and timeliness of bank retail business effect prediction, and realizes the generation of a high-precision and executable bank retail marketing strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent marketing forecasting technology, and in particular to a method and system for predicting the effectiveness of retail banking business. Background Technology

[0002] In recent years, with the rapid development of fintech, retail banking has gradually shifted from traditional offline services to online digital operations. Precision marketing and risk management have become key elements in improving customer experience and business revenue. Existing performance prediction methods mainly rely on historical data and statistical models, developing marketing strategies by analyzing customer transaction behavior, merchant operating conditions, and campaign effectiveness. However, existing methods often lack effective integration of multi-source heterogeneous data, making it difficult to comprehensively capture the dynamic changes in customer behavior and its complex relationships with merchants and campaigns. Due to the time lag in settlement confirmation and refund feedback, the predicted labels built based on currently observable data often have significant biases, leading to inaccurate evaluation of marketing strategy effectiveness. Most existing technologies employ static budgeting or simple prioritization, failing to achieve dynamic optimization and thus impacting overall resource utilization efficiency.

[0003] Existing retail banking operations generally suffer from the following limitations: First, insufficient data integration and feature construction capabilities. Existing methods often rely on single data sources and fail to effectively integrate heterogeneous data from multiple sources such as merchants, benefits, and payment transactions. This makes it difficult to form a unified cross-source data foundation. During feature extraction, there is a lack of comprehensive modeling of customer behavior time-series characteristics, merchant suitability characteristics, activity benefit characteristics, and risk factors, which limits the accuracy and robustness of the prediction model. Second, a disconnect between prediction and strategy generation, and a lack of closed-loop feedback mechanisms. Existing systems often use independent models to predict customer behavior and business revenue separately, without establishing a joint prediction framework based on the customer-merchant-activity ternary relationship. The strategy selection and resource allocation stages lack quantitative constraints on comprehensive risks, and the actual results after execution are not effectively fed back for model updates, making it difficult to adapt to changes in the market environment. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for predicting the effectiveness of retail banking business, which solves the problems of insufficient data integration and feature construction capabilities and the disconnect between prediction and strategy generation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting the effectiveness of retail banking business, comprising, Acquire multi-source business data and perform unified mapping processing on the multi-source business data to build a unified data foundation; A joint label for business revenue is constructed based on a unified data platform, and the joint label for business revenue is modified by delaying feedback based on the settlement confirmation status and risk factors in the unified data platform. Features are constructed on a unified data foundation to generate multidimensional joint feature representations. Based on the multidimensional joint feature representations and business revenue joint labels corrected by delayed feedback, a multi-model prediction model is constructed to jointly predict customer behavior and business revenue, and the joint effect prediction results are obtained. Based on the joint effect prediction results, a marketing strategy mix is ​​generated that takes into account the synergistic constraints of customers, merchants and rights, and the marketing strategy mix is ​​screened and evaluated in combination with risk constraints formed by risk factors. Under resource constraints built on a unified data foundation, the marketing strategy mix is ​​optimized and allocated to obtain strategic decision results; After the strategy decision is executed, business execution data is acquired, the execution result is determined based on the business execution data, and the model is updated based on the execution result.

[0007] As a preferred embodiment of the method for predicting the effectiveness of retail banking business as described in this invention, wherein: The process of acquiring multi-source business data and performing unified mapping processing on the multi-source business data to construct a unified data foundation includes: By employing multiple data access methods, multi-source business data is acquired from banks, merchants, rights and interests, and payment transaction data sources. The multi-source business data undergoes standardized preprocessing, and heterogeneous identifiers of customers, merchants, activities, accounts, and time are mapped to a unified coding space to form a cross-source unified mapping relationship. Based on the mapping relationship, the outreach, collection, payment, verification, settlement, and refund behaviors of the same customer within a single activity cycle are associated in chronological order to construct a business behavior trajectory represented by a sequence of state events. The behavior trajectories of all customer-merchant-activity triplets are aggregated to form a unified data foundation.

[0008] As a preferred embodiment of the method for predicting the effectiveness of retail banking business as described in this invention, the step of constructing a joint business revenue label based on a unified data foundation, and performing delayed feedback correction on the joint business revenue label in conjunction with the settlement confirmation status and risk factors in the unified data foundation, includes: Based on a unified data foundation, construct a joint tag for business revenue under the customer-merchant-activity ternary relationship; Based on the currently observable data, an initial value for the joint label of business revenue is generated, and the refund risk is calculated based on the historical behavior and business attribute feature vectors in the unified data foundation. Based on the current state value of the sample unit when the label is constructed, the settlement confirmation state factor is determined, and combined with the refund risk value, the settlement confirmation degree is constructed. Based on the settlement confirmation degree, the initial value of the business revenue joint label is weighted and corrected to obtain the corrected business revenue joint label.

[0009] As a preferred embodiment of the performance prediction method for retail banking business described in this invention, the step of constructing features on a unified data foundation to generate a multi-dimensional joint feature representation includes: Based on a unified data foundation, customer behavior time-series features, merchant compatibility features, activity benefits features, payment settlement status features, and refund risk features are extracted; and customer-merchant-activity collaborative interaction embedding is integrated to construct a multi-dimensional joint feature representation.

[0010] As a preferred embodiment of the method for predicting the effectiveness of retail banking business as described in this invention, wherein: The multi-model prediction model is constructed based on multi-dimensional joint feature representation and business revenue joint labels corrected by delayed feedback. This model jointly predicts customer behavior and business revenue, yielding joint effect prediction results including: A training sample set is constructed based on multidimensional joint feature representation and business revenue joint labels corrected by delayed feedback. The input features are divided into three parts: customer side, merchant side, and activity side. Sub-prediction models are constructed for each part and then fused to generate a unified latent vector. It also performs multi-task output based on a unified latent vector, simultaneously predicting customer behavior and business revenue to form a joint effect prediction result.

[0011] As a preferred embodiment of the effect prediction method for retail banking business described in this invention, the step of generating a marketing strategy combination based on the joint effect prediction results, which involves customer, merchant, and benefit synergy constraints, includes: Based on the joint effect prediction results, and combined with customer, merchant, activity, channel and timing information in the unified data platform, candidate strategy combinations are constructed. Based on the business revenue joint label and comprehensive risk item corrected by delayed feedback, the synergistic value score of each candidate strategy combination is calculated, and high-value marketing strategy combinations are generated based on the synergistic value score.

[0012] As a preferred embodiment of the method for predicting the effectiveness of retail banking business as described in this invention, wherein: The screening and evaluation of marketing strategy mix, which incorporates risk constraints formed by risk factors, includes: Risk constraints are constructed by combining risk factors, and high-value marketing strategy combinations are screened for risk compliance. The screened marketing strategy combinations are then ranked according to their synergistic value scores to form the final marketing strategy candidate set.

[0013] As a preferred embodiment of the method for predicting the effectiveness of retail banking business as described in this invention, wherein: The optimization and allocation of marketing strategy combinations under resource constraints based on a unified data foundation, resulting in strategic decision-making outcomes, include: Based on a unified data foundation, constraints on budget, equity inventory, channel capacity, and customer frequency resources are constructed. With collaborative value scoring as the optimization objective, the marketing strategy mix is ​​constrained and solved to obtain the optimal or near-optimal resource allocation results, and the final strategy decision results are determined.

[0014] As a preferred embodiment of the effect prediction method for retail banking business described in this invention, the step of acquiring business execution data after the strategy decision result is executed, determining the execution result based on the business execution data, and updating the model based on the execution result includes: After the strategy is executed, business execution data is acquired and mapped to a unified coding space, associated with the state event stream in the unified data foundation, and updated to form the post-execution state event stream; The actual execution results of the strategy are determined based on the updated state event stream and execution data, and compared with the joint effect prediction results to calculate the execution deviations in the dimensions of payment, write-off, handling fees and net income. When any key deviation is greater than or equal to a preset threshold, updated labels and incremental training samples are constructed based on the actual execution results, added to the historical training data pool, and the model is triggered to continuously iterate and optimize.

[0015] Secondly, this invention provides an effectiveness prediction system for retail banking business, comprising, The unified data foundation module is used to collect and clean customer, merchant, activity and transaction behavior data from multiple business systems, and to build a standardized state event flow foundation with the "customer-merchant-activity" triple as the core through unified identifier mapping. The label correction module is used to calculate the settlement confirmation degree by combining the settlement confirmation status and refund risk, and to perform weighted correction on the estimated revenue label to generate a joint net revenue label that takes into account both timeliness and accuracy. The joint prediction module is used to fuse multi-dimensional features, build a multi-task deep learning model, and simultaneously predict payment probability, write-off probability, transaction fees, and net income. The strategy screening module is used to generate candidate marketing strategy combinations based on prediction results, and to screen and sort them in combination with comprehensive risk assessment and business risk control rules, outputting a high-value and compliant strategy set; The optimization allocation module is used to solve for the optimal strategy allocation scheme under resource constraints, and after execution, it feeds back real data to trigger incremental model updates to achieve closed-loop optimization.

[0016] The beneficial effects of this invention are as follows: By constructing a unified data foundation, it achieves standardized integration and state event flow reconstruction of multi-source heterogeneous data from banks, merchants, rights and interests, and payments. It innovatively introduces a delayed feedback correction mechanism based on settlement confirmation status factors and refund risks, effectively alleviating label bias caused by settlement lag. It adopts a multi-task joint prediction model that integrates customer-merchant-activity multi-sided features, and simultaneously outputs multi-dimensional performance indicators of payment, write-off, handling fees, and net income, strengthening the comprehensiveness of strategy value assessment. It embeds comprehensive risk items into collaborative value scoring and optimizes allocation under multi-dimensional resource constraints to ensure the high value and executability of the strategy. Through an incremental learning mechanism driven by execution bias, it achieves closed-loop iteration of the model, significantly improving the system's adaptability and long-term operational efficiency in dynamic business environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an effect prediction method for retail banking business in Example 1.

[0019] Figure 2 This is a structural diagram of an effect prediction system for retail banking business in Example 1.

[0020] Figure 3 This is a schematic diagram of the neural network collaborative prediction model structure with three input paths: bank, merchant, and activity, as shown in Example 1.

[0021] Figure 4 This is a schematic diagram of the optimal resource allocation model with settlement risk constraints in Example 1. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides a method for predicting the effectiveness of retail banking business, including the following steps: S1. Acquire multi-source business data and perform unified mapping processing on the multi-source business data to build a unified data foundation.

[0026] S1.1: Obtain multi-source business data from banks, merchants, rights and interests, and payment transaction data sources through various data access methods.

[0027] Specifically, multi-source business data is obtained from bank-side data sources, merchant-side data sources, rights management data sources, and payment transaction data sources through data interface calls, message subscriptions, log collection, and batch data synchronization. The multi-source business data includes customer data, merchant activity data, rights and benefits distribution and collection data, channel reach data, payment transaction data, acquiring and settlement data, and refund or cancellation data.

[0028] S1.2: Standardize and preprocess multi-source business data, and map heterogeneous identifiers of customers, merchants, activities, accounts and time to a unified coding space to form a cross-source unified mapping relationship.

[0029] Specifically, the acquired multi-source business data is preprocessed, including field format standardization, missing value handling, outlier removal, time field unification, and identifier field normalization. Map customer identifiers, account identifiers, merchant identifiers, activity identifiers, and time identifiers from different sources to a unified coding space, and define a unified mapping function across data sources:

[0030] in, This represents the set of identifiers in the original data source. Indicates the original time identifier. , , and These respectively represent the unified customer identifier, merchant identifier, event identifier, and time identifier.

[0031] S1.3: Based on the mapping relationship, the actions of the same customer in a single activity cycle, including contact, receipt, payment, verification, settlement and refund, are associated in chronological order to construct a business behavior trajectory represented by a sequence of state events. The behavior trajectories of all customer-merchant-activity triplets are aggregated to form a unified data foundation.

[0032] Specifically, based on the mapping function, cross-data source unified mapping and association reconstruction are performed on multi-source business data, so that the outreach, receipt, payment, cancellation, settlement and refund behaviors of the same customer in the same activity period are uniformly associated in chronological order, and the business behavior trajectory under the three-element relationship of customer, merchant and activity is obtained. Encode various business behaviors in the business behavior trajectory into unified state events, and define the state transition sequence of a single customer within a single activity cycle. for:

[0033] in, Indicates that the message was not reached. This indicates that the message has been received. This indicates that the rights have been claimed. This indicates that payment has been made. This indicates that the transaction has been cancelled. This indicates that the settlement has been confirmed. Indicates a refund or cancellation; For any customer Merchants ,Activity At any moment Business behaviors, construct state event groups :

[0034] in, This indicates the state value corresponding to the event. This represents the feature vector of the business attributes corresponding to the event; The state event groups are assembled in chronological order to form the state event flow of the customer within the activity cycle. :

[0035] Based on state event stream And the mapped multi-source business data, to build a unified data foundation. : .

[0036] S2. Construct a joint business revenue label based on a unified data platform, and perform delayed feedback correction on the joint business revenue label in conjunction with the settlement confirmation status and risk factors in the unified data platform.

[0037] S2.1: Based on a unified data foundation, construct a joint tag for business revenue under the three-element relationship of customer-merchant-activity.

[0038] Specifically, based on a unified data foundation From state event stream Extract customers Merchants ,Activity Construct sample units from the complete behavioral sequence within the decision-making cycle. :

[0039] Based on sample units Extract various metrics used to characterize business revenue, including commission increments. Net change in customer assets Incremental payment transactions Activity subsidy costs Losses resulting from refunds, cancellations, or write-offs ; Based on revenue composition indicators, a joint tag for business revenue is constructed under the three-element relationship of customers, merchants, and activities. :

[0040] It should be noted that, The weight coefficients are non-negative real numbers and can be determined based on historical sample data using regression analysis, machine learning model training, or expert experience. They can be dynamically adjusted according to different business objectives using multi-objective optimization methods, with the optimal real numbers in the [0,1] interval being selected, and the sum of all weight coefficients being 1. The specific process is as follows: For the "consumption voucher activation and payment conversion improvement" activity in the retail banking business, a training dataset can be constructed based on historical marketing sample data. Indicators such as handling fee income, AUM changes, payment transaction increments, subsidy costs, and refund losses can be extracted. With actual operating income or ROI as the optimization objective, the parameters are fitted through regression analysis or machine learning models to obtain the initial weight coefficients. Business constraints are then introduced, and the weights are dynamically adjusted through multi-objective optimization methods to finally determine the weight combination that meets the current business strategy objectives. This allows for flexible control of the revenue structure and optimal balance of overall operating results under different marketing scenarios.

[0041] S2.2: Based on the currently observable data, generate the initial value of the joint label for business revenue, and calculate the refund risk based on the historical behavior and business attribute feature vectors in the unified data foundation.

[0042] Specifically, due to the time lag between settlement confirmation and refund return, a phased revenue interpolation method is adopted based on currently observable data. Depending on the current state of the strategy, each revenue and cost item in the business revenue joint label formula is filled with observed values ​​or historical statistical estimates to generate initial values ​​for complete but not yet finalized business revenue joint labels. ; The currently observable data includes customer identifiers, merchant identifiers, activity identifiers, current status values, payment, redemption, and outreach events that have occurred, records of rights and benefits issuance and receipt, historical behavioral statistical characteristics of customers and merchants, as well as risk factors and business attribute characteristics; Based on historical behavior and business attribute feature vectors in a unified data foundation Calculate refund risk :

[0043] in, Refers to the customer's historical refund rate. This refers to the merchant's historical refund rate. Refers to the refund rate of historical activities. Refers to the characteristic value of abnormal behavior. Refers to the Sigmoid function. Refers to the model bias term; The abnormal behavior feature value is an indicator that reflects potential abnormal operations or high-risk patterns. It is obtained by using the isolated forest model to score the behavioral characteristics of customers-merchants-activities in an unsupervised manner, and combined with rule detection and statistical analysis of the behavioral sequences of customers-merchants-activities in the unified data foundation. It should be noted that, The weight coefficients of each risk factor are non-negative real numbers, which can be obtained by training a logistic regression model, gradient boosting tree model, or neural network model based on historical sample data. They are preferably real numbers between [0,1], and the sum of each weight coefficient is 1. The basis for this preference is that the customer historical refund rate, merchant historical refund rate, activity historical refund cancellation rate, and order anomaly characteristics in this scheme are all probabilistic or intensity-based risk quantities. Using normalized weights in the [0,1] interval is more conducive to matching the refund risk probability output by the Sigmoid function, so as to minimize the error between the predicted refund risk and the actual refund probability, thereby improving the accuracy of risk assessment.

[0044] S2.3: Based on the current state value of the sample unit when the label is constructed, determine the settlement confirmation state factor, and combine it with the refund risk value to construct the settlement confirmation degree. Based on the settlement confirmation degree, the initial value of the business revenue joint label is weighted and corrected to obtain the corrected business revenue joint label.

[0045] Specifically, from the sample unit State event stream In the middle, get the status value of the latest event. :

[0046] Status value based on the latest event The settlement confirmation status factor is determined by a preset state-confidence mapping rule. :

[0047] Among them, if If so, the real label will be used directly, skipping this correction process; Further integrate refund risk values ​​to construct settlement confirmation levels. :

[0048] Based on settlement confirmation The initial value of the joint tag for business revenue Make corrections to obtain the revised joint label for business revenue. :

[0049] in, This refers to the actual revenue label value calculated based on complete observation results after subsequent clearing and settlement and refund return.

[0050] S3. Construct features on the unified data foundation to generate multi-dimensional joint feature representations. Based on the multi-dimensional joint feature representations and the business revenue joint labels corrected by delayed feedback, construct a multi-model prediction model to jointly predict customer behavior and business revenue, and obtain joint effect prediction results.

[0051] S3.1: Based on a unified data foundation, extract customer behavior time-series characteristics, merchant compatibility characteristics, activity benefits characteristics, payment settlement status characteristics, and refund risk characteristics.

[0052] Specifically, extract customer information. Merchants ,Activity Behavioral data and business attribute feature vectors within the decision-making cycle Construct the original feature set:

[0053] Based on the original feature sequence, time-decay aggregation of customer behavior is performed to obtain customer behavior and value features. :

[0054] in, This refers to the current time and the end time of the statistical window. Refers to the feature category index. Customer In time The Similar to original observations, Refers to the current moment With the moment when historical behavior occurred The time difference between them Refers to the time decay coefficient. Refers to the time decay weighting function; It should be noted that the time decay coefficient The value range is non-negative real numbers. The preferred value range is the positive real number interval after normalization of time units. It can be determined by training historical behavior data based on regression analysis, time series modeling or machine learning model training. The preferred value range is adaptively adjusted by parameter optimization method according to the model prediction effect or business objectives to achieve reasonable weighting of different time-sensitive behavioral characteristics. The original observations include, but are not limited to, transaction amount, number of transactions, number of mobile banking logins, number of marketing responses, number of payments, and number of write-offs. One or more of the following: balance, quantity of products held; Based on merchant and activity-related data in the unified data platform, merchant adaptation feature vectors are constructed respectively. and rights activity feature vector ; Based on the payment, reconciliation, and settlement status in the state event stream, payment and settlement features are extracted. :

[0055] in, This refers to a mapping function that extracts statistical features of payment and settlement based on state event streams; Constructing a risk feature vector based on refund risk :

[0056] S3.2: Integrate customer-merchant-activity collaborative interaction embedding to construct a multi-dimensional joint feature representation.

[0057] Specifically, based on customers, merchants, :

[0058] in, Interaction mapping functions are used to learn the non-linear interaction relationships between customers, merchants, and activities. Refers to the comprehensive feature vector of a customer; By fusing various features, a multidimensional joint feature representation is obtained: .

[0059] S3.3: Construct a training sample set based on multidimensional joint feature representation and business revenue joint labels corrected by delayed feedback.

[0060] Specifically, based on multidimensional joint feature representation and the joint label of business revenue after delayed feedback correction Construct a training sample set : .

[0061] S3.4: Divide the input features into three parts: customer side, merchant side, and activity side, construct sub-prediction models for each part, and then merge them to generate a unified latent vector.

[0062] Specifically, based on multidimensional joint feature representation The composition structure divides the input features into customer-side input features, merchant-side input features, and activity-side input features; Customer-side input features include customer behavior and value features. Merchant-side input features include merchant-adaptive feature vectors. The input features on the activity side include the equity activity feature vector. Payment and settlement characteristics and risk feature vector ; Based on customer-side input features Build a customer-side sub-prediction model :

[0063] in, This represents the client-side feature mapping function; Based on merchant-side input features Construct a merchant-side prediction model:

[0064] in, This represents the feature mapping function on the merchant side; Based on activity-side input features Construct an active side prediction model:

[0065] in, Refers to channel identification. Refers to channel characteristics, Refers to timing characteristics, Refers to the mapping function of a deep neural network; Features are fused to obtain a unified latent vector. :

[0066] S3.5: Based on a unified latent vector, multi-task output is performed to simultaneously predict customer behavior and business revenue, forming a joint effect prediction result.

[0067] Specifically, based on the unified latent vector Construct multi-task prediction outputs respectively:

[0068] in, This refers to the probability of a payment occurring. This refers to the probability of rights write-off. This refers to the projected increase in transaction fees. Refers to the revised net income forecast after settlement. , , and This refers to the bias parameters corresponding to each prediction task. Refers to the Sigmoid activation function; It should be noted that, The weight parameters corresponding to different prediction tasks take values ​​in the real number domain, preferably in a finite real number range after normalization or regularization constraints. They can be determined by iteratively training the multi-layer neural network model using the gradient descent algorithm based on training sample data, or by optimizing the solution using existing backpropagation algorithms, stochastic gradient descent algorithms and their improved algorithms, so as to minimize the loss function between the model prediction output and the actual label. Will Joint feature representation Input the trained model to obtain customer behavior prediction results, business revenue prediction results, and generate a joint effect prediction result. :

[0069] The multi-model prediction model includes an input layer, a unit layer, a fusion layer, and an output layer; The input layer is used to receive multidimensional joint feature representations and split them into customer-side input, merchant-side input, and activity-side input according to their source. The unit layer is defined as follows: mapping customer-side input, merchant-side input, and activity-side input respectively to obtain corresponding hidden layer representations. , and ; The fusion layer is defined as follows: it fuses the three hidden layer representations to output a unified hidden vector. ; The output layer is based on a unified latent vector. Multiple prediction output units are constructed separately to obtain the joint effect prediction results. ; Model training steps: Construct a training sample set based on the multi-dimensional joint feature representation and the business revenue joint label corrected by delayed feedback; input the training samples into the multi-model prediction model; and calculate the multi-task joint loss function. The model parameters are iteratively updated through backpropagation to obtain the trained prediction model. When the model is used, the multi-dimensional joint feature representation of the sample to be predicted is input into the trained prediction model, and the joint effect prediction result is output.

[0070] S4. Generate a marketing strategy mix based on the joint effect prediction results, taking into account the synergistic constraints of customers, merchants and rights, and screen and evaluate the marketing strategy mix in combination with the risk constraints formed by risk factors.

[0071] S4.1: Based on the joint effect prediction results, and combined with customer, merchant, activity, channel and timing information in the unified data foundation, construct candidate strategy combinations.

[0072] Specifically, based on the joint effect prediction results Extract customers In merchants ,Activity ,channel and time The multidimensional prediction results under the given conditions are used to construct the input data for the policy generation stage.

[0073] The corresponding quintuple is only used if the multidimensional prediction result exists and is valid. Include in the candidate strategy space; Based on joint effect prediction results and unified data foundation Use customer, merchant, event, channel, and timing data to build candidate strategy combinations. :

[0074] S4.2: Calculate the synergistic value score of each candidate strategy combination based on the business revenue joint label and comprehensive risk item after delayed feedback correction, and select and generate high-value marketing strategy combinations based on the synergistic value score.

[0075] Specifically, for any combination of candidate strategies Based on the joint label of business revenue and comprehensive risk items after delayed feedback correction, the synergistic value score of the candidate strategy combination is calculated:

[0076] It should be noted that, , , These are the business value weights for adjusting net revenue, payment conversion probability, and equity write-off probability, respectively. The risk penalty weights, all of which are non-negative real numbers, can be determined through regression analysis, machine learning fitting, or expert calibration of historical marketing campaign revenue-conversion-risk multidimensional effect data. Specifically, the process involves collecting a large number of executed bank retail marketing strategy samples, extracting actual observed values ​​for adjusted net income, payment conversion rate, equity write-off rate, and comprehensive risk items, and using overall ROI or operating net income as the optimization objective. A multi-objective optimization method is employed to select a balance between maximizing revenue and controlling risk, with the optimal value being... , , , In scenarios targeting "consumption voucher activation and payment conversion improvement", the optimal value was determined after multiple rounds of testing, strategy backtracking verification and business KPI alignment evaluation for different customer groups, different merchant types and different benefit cost structures. The comprehensive risk item is as follows:

[0077] in, This refers to the risk of refunds. This refers to the risk of abnormal write-offs. This refers to the risk of merchants failing to fulfill their obligations. Indicating the intensity of resource consumption; The aforementioned refund risk It is derived by weighting the calculation based on the historical refund frequency statistics of the customer, merchant and activity triplet and the anomaly detection features of the current behavior sequence; The risk of abnormal write-off It is derived by using rule engine detection or isolated forest model to detect the time interval, operation frequency and amount matching characteristics of rights claim and cancellation behavior in the state event stream; Merchant performance risk It is derived by constructing a merchant risk scoring card based on the merchant's historical settlement success rate, customer complaint records, and the risk level of the industry to which the merchant belongs; resource occupancy intensity It is derived by using a multi-dimensional business rule weighted evaluation method based on the current channel load rate, the percentage of remaining activity budget, and whether the time window is during the peak business period; It should be noted that the weighting coefficients , , , This is the weighting coefficient for each risk factor in the comprehensive risk item, and all values ​​are non-negative real numbers, preferably satisfying the following conditions: This can be determined through statistical modeling or machine learning training of the correlation data between risk events and business losses in historical marketing activities. The specific process involves collecting a large number of refund records, abnormal reimbursement behaviors, merchant performance complaints, and channel resource overrun events corresponding to executed strategies, and quantifying them accordingly. , , , Four risk indicators are used to construct regression or classification targets based on actual net return deviations. Logistic regression, gradient boosting trees, or neural network models are employed to fit the contribution of each risk factor to the overall loss. A balance is selected between risk sensitivity and strategy coverage, with the optimal value being... In joint marketing campaigns targeting retail banking that combine "payment activation and rights redemption," the optimal value was determined after three empirical tests: historical backtesting, stress disturbance testing, and risk control-return Pareto trade-off analysis. This value covers typical scenarios such as customers with a high tendency to request refunds, newly joined small and medium-sized merchants, high-concurrency holiday activities, and multi-channel SMS outreach. Based on collaborative value scoring, a greedy resource allocation algorithm with multiple constraints is used to screen candidate strategy combinations, resulting in high-value marketing strategy combinations: .

[0078] S4.3: Combine risk factors to construct risk constraints, screen high-value marketing strategy combinations for risk compliance, and sort the screened marketing strategy combinations according to their synergistic value scores to form the final marketing strategy candidate set.

[0079] Specifically, based on business rules and risk control requirements, a set of risk constraints is constructed. :

[0080] It should be noted that, , , and The preset risk control threshold in the risk constraints takes values ​​that are all real numbers in the range [0,1]. It can be determined by statistical analysis of the distribution characteristics and business tolerance of various risk indicators in historical marketing activities. The specific process is as follows: collect a large number of samples of the actual refund rate, abnormal resale rate, merchant default rate, and channel resource occupancy intensity corresponding to the executed strategies, and construct them respectively. , , , The empirical cumulative distribution function is used to select a balance point between risk interception rate and strategy investability rate, with the optimal value being [value missing]. In the joint marketing scenario of "consumption vouchers + payment cashback" for bank retail, it covers a typical business environment of highly active urban customer groups, mixed merchant ecosystem of chain supermarkets and small and micro catering, high-concurrency reach during holidays and multi-channel collaboration of SMS / APP / WeChat. At the same time, it faces a complex risk factor of user mis-claiming, merchant order fraud, over-issuance of inventory, and frequent channel control conflicts. The optimal value was determined after three empirical studies: historical backtesting stress test, anti-disturbance simulation and risk control-benefit Pareto boundary analysis. Meet risk constraints Under this premise, select high-value marketing strategy combinations:

[0081] The selected marketing strategy mix was ranked according to its synergy value score:

[0082] The sorted strategy combinations will be used as the final marketing strategy candidate set: .

[0083] S5. Under the resource constraints of a unified data platform, optimize the allocation of marketing strategy mix to obtain strategic decision results.

[0084] S5.1: Based on a unified data foundation, construct resource constraints for budget, equity inventory, channel capacity, and customer frequency. With collaborative value scoring as the optimization objective, solve the constraints of the marketing strategy mix to obtain the optimal or near-optimal resource allocation results and determine the final strategy decision results for the campaign.

[0085] Specifically, based on the final marketing strategy candidate set Extract the collaborative value score corresponding to each candidate strategy combination. Comprehensive risk items and a unified data foundation The corresponding budget consumption, equity inventory usage, channel capacity usage, and customer reach frequency information are used to construct the input dataset for the optimization allocation phase; Based on the input dataset, define the policy assignment decision variables. :

[0086] in, This indicates that candidate strategy combinations will be assigned to the customer. , This indicates that no candidate strategy combination will be assigned. Based on the collaborative value score of candidate strategy combinations, a resource optimization allocation objective function is constructed:

[0087] The objective function is used to maximize the overall collaborative value of the allocated strategy combination under resource constraints. Based on a unified data foundation Based on the recorded activity budget information, construct budget constraints:

[0088] in, This refers to the budget consumption corresponding to the combination of candidate strategies. This refers to the upper limit of the total budget available within the current decision-making cycle; Based on a unified data foundation Based on the equity inventory information recorded in the database, construct equity inventory constraints:

[0089] in, This refers to the equity allocation corresponding to the candidate strategy combination. Refers to activities The upper limit of available inventory corresponding to the equity; Based on a unified data foundation Based on the channel capacity information recorded in the database, channel capacity constraints are constructed:

[0090] in, This refers to the channel resource ownership corresponding to the candidate strategy combination. Channels At any moment The maximum available capacity; Based on a unified data foundation Based on the recorded customer history and frequency control information, customer frequency constraints are constructed:

[0091] in, This refers to the maximum frequency control limit per passenger. Under the constraints of budget, equity inventory, channel capacity, and customer frequency, a greedy approximation algorithm based on value scoring and ranking is used to evaluate the decision variables. Solve the problem to obtain the optimal or near-optimal resource allocation results:

[0092] Strategic decision results : .

[0093] S6. After the strategy decision is executed, obtain the business execution data, determine the execution result based on the business execution data, and update the model based on the execution result.

[0094] S6.1: After the strategy is executed, the business execution data is obtained and mapped to the unified coding space, associated with the state event stream in the unified data foundation, and updated to form the state event stream after execution.

[0095] Specifically, for the executed strategy combinations, obtain business execution data. Furthermore, it maps business execution data back to a unified coding space corresponding to customer identifiers, merchant identifiers, activity identifiers, and time identifiers, in order to integrate with a unified data foundation. Associate the original state event streams in the data; The business execution data includes outreach result data, rights and benefits claim data, payment transaction data, verification result data, clearing and settlement result data, and refund, cancellation, or offset data; Based on business execution data, the state event groups of the corresponding strategy combinations are supplemented and updated to form a post-execution state event stream. :

[0096] in, This refers to the new status event group added after the strategy is executed.

[0097] S6.2: Determine the actual execution results of the strategy based on the updated state event stream and execution data, and compare them with the joint effect prediction results to calculate the execution deviations in dimensions such as payment, write-off, handling fees and net income.

[0098] Specifically, based on the updated state event stream and business execution data Determine the execution result corresponding to the strategy combination. :

[0099] in, Indicates the actual payment result. This indicates the actual write-off result. This indicates the actual transaction fee result. This indicates the actual net profit result after combining settlement and refund repatriation; Based on the joint effect prediction results With execution results Calculate execution deviation :

[0100] The execution deviation includes payment deviation. Verification deviation Transaction fee deviation and net profit deviation; Net income deviation can be expressed as: .

[0101] S6.3: When any key deviation is greater than or equal to the preset threshold, update the labels and incremental training samples based on the actual execution results, add them to the historical training data pool, and trigger continuous iterative optimization of the model.

[0102] Specifically, based on the execution results and the updated state event stream, the joint feature representation of the corresponding strategy combination is re-extracted. ; And build updated tags based on actual net income results. :

[0103] Constructing an incremental training sample set :

[0104] Incremental training sample set Add to the historical training data pool to achieve continuous iterative optimization of predictive capabilities; Based on the deviation in the execution results, determine whether to trigger a model update. Or satisfy Or satisfy When this happens, the model is updated. in, , and These represent the net income deviation threshold, payment deviation threshold, and write-off deviation threshold, respectively. It should be noted that the threshold , and The values ​​are all non-negative real numbers and can be determined through regression analysis based on historical data, machine learning model training, or expert experience. The specific process is as follows: Regression analysis is performed on net income, payment behavior, and write-off behavior based on historical data to determine reasonable thresholds for each deviation, so that model updates are triggered when significant deviations occur. Multiple machine learning models are trained, and the triggering conditions for model updates are predicted based on historical sample data to further optimize the threshold values. The preferred value is... In retail banking, banks improve payment conversion rates and rights redemption rates through promotional activities and interactions with merchants and customers. By analyzing historical business execution data and combining the actual changes in net revenue deviation, payment deviation, and redemption deviation, regression analysis or machine learning methods are used to fit the optimal threshold range. In practical applications, the reasonableness of the deviation and the stability of the model need to be considered when setting the threshold. By selecting an appropriate threshold range, it is possible to avoid triggering model updates too frequently, while ensuring the accuracy and stability of the model.

[0105] This embodiment also provides an effect prediction system for retail banking business, including: The unified data foundation module is used to collect and clean customer, merchant, activity and transaction behavior data from multiple business systems, and to build a standardized state event flow foundation with the "customer-merchant-activity" triple as the core through unified identifier mapping. The label correction module is used to calculate the settlement confirmation degree by combining the settlement confirmation status and refund risk, and to perform weighted correction on the estimated revenue label to generate a joint net revenue label that takes into account both timeliness and accuracy. The joint prediction module is used to fuse multi-dimensional features, build a multi-task deep learning model, and simultaneously predict payment probability, write-off probability, transaction fees, and net income. The strategy screening module is used to generate candidate marketing strategy combinations based on prediction results, and to screen and sort them in combination with comprehensive risk assessment and business risk control rules, outputting a high-value and compliant strategy set; The optimization allocation module is used to solve for the optimal strategy allocation scheme under resource constraints, and after execution, it feeds back real data to trigger incremental model updates to achieve closed-loop optimization.

[0106] In summary, this invention achieves standardized integration and state event flow reconstruction of heterogeneous data from multiple sources, including banks, merchants, rights and interests, and payments, by constructing a unified data foundation. It innovatively introduces a delayed feedback correction mechanism based on settlement confirmation status factors and refund risks, effectively mitigating label bias caused by settlement lags. Furthermore, it employs a multi-task joint prediction model that integrates customer-merchant-activity multi-sided features, simultaneously outputting multi-dimensional performance indicators for payment, write-off, handling fees, and net revenue, enhancing the comprehensiveness of strategy value assessment. It embeds comprehensive risk items into the collaborative value scoring, optimizing allocation under multi-dimensional resource constraints to ensure the high value and executability of the strategy. Finally, it achieves closed-loop iteration of the model through an execution bias-driven incremental learning mechanism, significantly improving the system's adaptability and long-term operational efficiency in dynamic business environments.

[0107] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the effectiveness of retail banking business, characterized in that: include, Acquire multi-source business data and perform unified mapping processing on the multi-source business data to build a unified data foundation; A joint tag for business revenue is constructed based on a unified data platform, and the joint tag for business revenue is modified by delaying feedback based on the settlement confirmation status and risk factors in the unified data platform. Features are constructed on a unified data foundation to generate multidimensional joint feature representations. Based on the multidimensional joint feature representations and business revenue joint labels corrected by delayed feedback, a multi-model prediction model is constructed to jointly predict customer behavior and business revenue, and the joint effect prediction results are obtained. Based on the joint effect prediction results, a marketing strategy mix is ​​generated that takes into account the synergistic constraints of customers, merchants and rights, and the marketing strategy mix is ​​screened and evaluated in combination with risk constraints formed by risk factors. Under resource constraints built on a unified data platform, the marketing strategy mix is ​​optimized and allocated to obtain strategic decision results; After the strategy decision is executed, business execution data is acquired, the execution result is determined based on the business execution data, and the model is updated based on the execution result.

2. The method for predicting the effectiveness of retail banking business as described in claim 1, characterized in that: The process of acquiring multi-source business data and performing unified mapping processing on the multi-source business data to construct a unified data foundation includes: By employing multiple data access methods, multi-source business data is acquired from banks, merchants, rights and interests, and payment transaction data sources. The multi-source business data undergoes standardized preprocessing, and heterogeneous identifiers of customers, merchants, activities, accounts, and time are mapped to a unified coding space to form a cross-source unified mapping relationship. Based on the mapping relationship, the outreach, collection, payment, verification, settlement, and refund behaviors of the same customer within a single activity cycle are associated in chronological order to construct a business behavior trajectory represented by a sequence of state events. The behavior trajectories of all customer-merchant-activity triplets are aggregated to form a unified data foundation.

3. The method for predicting the effectiveness of retail banking business as described in claim 2, characterized in that: The process of constructing a joint business revenue label based on a unified data platform, and then performing delayed feedback correction on the joint business revenue label in conjunction with the settlement confirmation status and risk factors in the unified data platform, includes: Based on a unified data foundation, construct a joint tag for business revenue under the customer-merchant-activity ternary relationship; Based on the currently observable data, an initial value for the joint label of business revenue is generated, and the refund risk is calculated based on the historical behavior and business attribute feature vectors in the unified data foundation. Based on the current state value of the sample unit when the label is constructed, the settlement confirmation state factor is determined, and combined with the refund risk value, the settlement confirmation degree is constructed. Based on the settlement confirmation degree, the initial value of the business revenue joint label is weighted and corrected to obtain the corrected business revenue joint label.

4. The method for predicting the effectiveness of retail banking business as described in claim 3, characterized in that: The step of constructing features on a unified data foundation to generate a multidimensional joint feature representation includes: Based on a unified data foundation, customer behavior time-series features, merchant compatibility features, activity benefits features, payment settlement status features, and refund risk features are extracted; and customer-merchant-activity collaborative interaction embedding is integrated to construct a multi-dimensional joint feature representation.

5. The method for predicting the effectiveness of retail banking business as described in claim 4, characterized in that: The multi-model prediction model is constructed based on multi-dimensional joint feature representation and business revenue joint labels corrected by delayed feedback. This model jointly predicts customer behavior and business revenue, yielding joint effect prediction results including: A training sample set is constructed based on multidimensional joint feature representation and business revenue joint labels corrected by delayed feedback. The input features are divided into three parts: customer side, merchant side, and activity side. Sub-prediction models are constructed for each part and then fused to generate a unified latent vector. It also performs multi-task output based on a unified latent vector, simultaneously predicting customer behavior and business revenue to form a joint effect prediction result.

6. The method for predicting the effectiveness of retail banking business as described in claim 5, characterized in that: The marketing strategy mix generated based on the joint effect prediction results, which is oriented towards customer, merchant, and benefit synergy constraints, includes: Based on the joint effect prediction results, and combined with customer, merchant, activity, channel and timing information in the unified data platform, candidate strategy combinations are constructed. Based on the business revenue joint label and comprehensive risk item corrected by delayed feedback, the synergistic value score of each candidate strategy combination is calculated, and high-value marketing strategy combinations are generated based on the synergistic value score.

7. The method for predicting the effectiveness of retail banking business as described in claim 6, characterized in that: The screening and evaluation of marketing strategy mix, which incorporates risk constraints formed by risk factors, includes: Risk constraints are constructed by combining risk factors, and high-value marketing strategy combinations are screened for risk compliance. The screened marketing strategy combinations are then ranked according to their synergistic value scores to form the final marketing strategy candidate set.

8. The method for predicting the effectiveness of retail banking business as described in claim 7, characterized in that: The optimization and allocation of marketing strategy combinations under resource constraints based on a unified data foundation, resulting in strategic decision-making outcomes, include: Based on a unified data foundation, constraints on budget, equity inventory, channel capacity, and customer frequency resources are constructed. With collaborative value scoring as the optimization objective, the marketing strategy mix is ​​constrained and solved to obtain the optimal or near-optimal resource allocation results, and the final strategy decision results are determined.

9. The method for predicting the effectiveness of retail banking business as described in claim 8, characterized in that: The step of acquiring business execution data after the strategy decision result is executed, determining the execution result based on the business execution data, and updating the model based on the execution result includes: After the strategy is executed, business execution data is acquired and mapped to a unified coding space, associated with the state event stream in the unified data foundation, and updated to form the post-execution state event stream; The actual execution results of the strategy are determined based on the updated state event stream and execution data, and compared with the joint effect prediction results to calculate the execution deviation in the dimensions of payment, write-off, handling fee and net income. When any key deviation is greater than or equal to a preset threshold, updated labels and incremental training samples are constructed based on the actual execution results, added to the historical training data pool, and the model is triggered to continuously iterate and optimize.

10. A performance prediction system for retail banking business, based on the performance prediction method for retail banking business according to any one of claims 1 to 9, characterized in that: include, The unified data foundation module is used to collect and clean customer, merchant, activity and transaction behavior data from multiple business systems, and to build a standardized state event flow foundation with the "customer-merchant-activity" triple as the core through unified identifier mapping. The label correction module is used to calculate the settlement confirmation degree by combining the settlement confirmation status and refund risk, and to perform weighted correction on the estimated revenue label to generate a joint net revenue label that takes into account both timeliness and accuracy. The joint prediction module is used to fuse multi-dimensional features, build a multi-task deep learning model, and simultaneously predict payment probability, write-off probability, transaction fees, and net income. The strategy screening module is used to generate candidate marketing strategy combinations based on prediction results, and to screen and sort them in combination with comprehensive risk assessment and business risk control rules, outputting a high-value and compliant strategy set; The optimization allocation module is used to solve for the optimal strategy allocation scheme under resource constraints, and after execution, it feeds back real data to trigger incremental model updates to achieve closed-loop optimization.

Images