Customer risk prediction method and system based on transaction behavior characteristics

Abstract

The application discloses a customer risk prediction method and system based on transaction behavior characteristics, relates to the technical field of transaction data processing and customer risk prediction, and the transaction platform server acquires transaction behavior events, product exposure data and market state data according to a prediction deadline, and forms a dynamic reference customer set based on the product exposure data and the market state data. The conditional expected transaction behavior characteristic is derived from the transaction behavior of customers under the condition of the same market state and similar product exposure, the actual transaction behavior characteristic of the target customer is compared with the conditional expected transaction behavior characteristic, a customer-specific transaction behavior deviation characteristic and a residual transaction behavior sequence are formed, the market common fluctuation and the product inherent fluctuation are first included in the reference benchmark, the risk prediction result can correspond to the transaction deviation of the target customer relative to the reference customer, and the probability that normal stop loss, passive portfolio adjustment or product switching are identified as abnormal transactions of the customer itself is reduced.

Description

Technical Field

[0001] This invention relates to the field of transaction data processing and customer risk prediction technology, and in particular to a customer risk prediction method and system based on transaction behavior characteristics. Background Technology

[0002] With the development of securities trading platforms, fund sales platforms, and internet wealth management businesses, platform servers typically continuously record clients' transaction behaviors such as subscriptions, redemptions, purchases, sales, order cancellations, order modifications, product conversions, and portfolio transfers. These records are then combined with clients' asset size, risk tolerance, historical trading frequency, changes in portfolio value, and product exposure risk levels to predict risks such as concentrated capital outflows, trading interruptions, or declining trading demand. Some processing methods also input clients' historical trading records into statistical models, ensemble learning models, or time-series prediction models to output client churn probability, redemption risk level, or a list of clients requiring intervention.

[0003] In the above-mentioned approaches, client risk is typically determined directly by characteristics such as the client's historical trading volume, trading frequency, profit changes, asset volatility, and portfolio concentration. For clients holding funds, stock portfolios, wealth management products, or margin trading products, a rapid decline in market indices, drawdowns in industry-specific assets, or fluctuations in the net asset value of a single product can simultaneously affect the floating profits and losses, redemption ratios, and trading activity of multiple clients. If the absolute losses, redemption ratios, or turnover rates of different clients are directly compared, differences in client holding stages, product categories, entry times, leverage levels, and account lifecycles will interfere with the risk assessment process, conflating normal stop-loss orders, passive portfolio adjustments, product maturity switching, or short-term observation behaviors with the client's own diminishing trading needs.

[0004] Other approaches establish customer relationship graphs based on jointly held products, similar trading directions, or similar return changes, using transaction changes of related customers as risk propagation signals. During market downturns or rapid fluctuations in the net asset value of popular products, a single product may simultaneously connect to a large number of customers. Transaction changes resulting from shared product exposure will repeatedly appear in product characteristics, customer profit and loss characteristics, and customer relationship propagation characteristics. This causes risk signals to spread along highly connected product nodes, leading to frequent switching of customer risk lists and causing the projected capital outflow and transaction demand changes at the product level to deviate from the reasons for abnormal customer transactions. Therefore, a customer risk prediction method and system based on transaction behavior characteristics is urgently needed to address these issues. Summary of the Invention

[0005] This application provides a customer risk prediction method and system based on transaction behavior characteristics, which solves the problem that abnormal customer transactions are difficult to separate and are repeatedly propagated under the common impact of the market.

[0006] In a first aspect, embodiments of the present invention provide a customer risk prediction method based on transaction behavior characteristics, executed by a transaction platform server, comprising: Obtain transaction behavior event sets, product exposure data, account status data, and market status data of target customers and candidate customers up to the prediction deadline; Reference matching conditions are generated from the product exposure data, the account status data, and the market status data, and a dynamic reference customer set is formed after excluding target customers. Based on the dynamic reference customer set, conditional expected transaction behavior characteristics are generated. The actual transaction behavior characteristics of the target customer and related candidate customers are compared with the corresponding conditional expected transaction behavior characteristics according to the event time period to form a multi-customer residual transaction behavior sequence. Candidate customer relationships with relationship source tags are generated based on the similarity of original transaction behavior, common product exposure relationship and the time period of the transaction behavior, and the weight component corresponding to the common product exposure relationship is marked as the common impact weight component; Residual behavior is reviewed for candidate customer relationships that meet the criteria for participation in the cumulative process. When both ends of the customer relationship maintain the same non-zero deviation state, or when the non-zero deviation state of one end of the customer relationship leads that of the other end of the customer relationship and continues to reach the preset time period quantity, the corresponding weight component is retained and a directed relationship weight is generated. If the residual behavior review criteria are not met, the common impact weight component is deleted. If the criteria for participation in the cumulative process are not met, a relationship anomaly record is generated. Based on the retained weight component, the directed relationship weight, and the relationship anomaly record, de-common impact customer relationship data is generated. The time-series coding of multi-customer residual transaction behavior sequences is performed, and the time-series coding results of related customers pointing to the target customer are normalized and aggregated according to the de-common impact customer relationship data to generate a joint customer representation; Based on the combined client representation, the output focuses on the risk of concentrated capital outflow and the risk of declining transaction demand, and aggregates and outputs the expected changes in capital outflow demand and transaction demand.

[0007] In some embodiments, the transaction behavior event set is obtained by standardizing the original transaction events from different transaction channels. The standardized transaction behavior event includes at least customer identifier, product identifier, event type, event time, time of awareness, transaction direction, transaction amount, processing status, original order identifier, cancellation association identifier, and correction association identifier. The transaction platform server associates cancellation events and modification events with the corresponding original transaction events based on the original order identifier, cancellation association identifier, or correction association identifier, and generates the actual transaction behavior characteristics based on the associated transaction behavior events.

[0008] In some embodiments, the product exposure data includes product category, holding stage, holding concentration status and leverage status, and the account status data includes account opening duration, historical valid trading days and the most recent valid trading time; The market status data includes market volatility, product price or net asset value changes, and market trading activity. The reference matching conditions include product matching relationships, holding stage relationships, holding concentration relationships, leverage relationships, account lifecycle relationships, and market status relationships.

[0009] In some embodiments, when the dynamic reference customer set is formed, the transaction platform server restores the product exposure status, account status and transaction behavior status of the candidate customer that has been formed and known to the transaction platform server at the same prediction cutoff time as the target customer. The trading platform server uses market status correspondence, leverage status correspondence, and position concentration status correspondence as fixed matching fields, and product matching correspondence, position holding stage correspondence, and account lifecycle correspondence as flexible matching fields. In the benchmark matching version, the product matching correspondence is the correspondence of the same product identifier, the position holding stage correspondence is the correspondence of the same position holding days range, and the account lifecycle correspondence is the correspondence of the same account lifecycle range.

[0010] In some embodiments, when the number of customers written is lower than the threshold of the set size, the trading platform server generates the next reference set version and gradually relaxes the restrictions in the order of product matching correspondence, holding stage correspondence and account lifecycle correspondence. The product matching relationship is broadened from matching the same product identifier to matching the same product subclass and then matching the same product exposure risk level. The correspondence between holding periods has been broadened from the same holding period interval to the correspondence between adjacent holding period intervals; The correspondence of account lifecycles is relaxed from the correspondence of the same account lifecycle interval to the correspondence of adjacent account lifecycle intervals; after each relaxation, the trading platform server records the prediction cutoff time, status version, relaxed matching field, matching caliber before relaxation, matching caliber after relaxation, and number of written customers in the reference set version table.

[0011] In some embodiments, when generating the multi-customer residual transaction behavior sequence, the transaction platform server generates residual transaction behavior sequences for the target customer and the associated candidate customer respectively according to the unified market state version, the unified event time period boundary and the unified residual state coding rule; When there is no market state change point that reaches the preset change intensity within the observation period, the start time of the observation period shall be used as the relative time starting point. When the dynamic reference customer set has a low sample count, the corresponding expected transaction behavior characteristics are written into the low reliability level.

[0012] In some embodiments, the criteria for participation in the accumulation include that both ends of the candidate customer relationship have valid residual transaction behavior data for no less than a preset number of consecutive event periods under the same transaction behavior deviation characteristics; When performing residual behavior review on the weighted components corresponding to the common product exposure relationship, exclude the other end customer from the dynamic reference customer set of one end customer. When the common impact weight component lacks a period of time that can participate in the cumulative event, the trading platform server generates a relationship anomaly record, and the corresponding candidate customer relationship does not participate in the aggregation of the time-series coding results of related customers in this prediction.

[0013] In some embodiments, when generating the joint client representation, the trading platform server inputs the multi-client residual transaction behavior sequence into a time encoder to obtain the time deviation representation of each client; The temporal deviation representations of related customers are weighted, normalized, and aggregated according to the directed relation weights pointing to the target customer to obtain the relation deviation representation; The time-series deviation representation of the target customer, the relationship deviation representation, and the product exposure data are concatenated to form the joint customer representation; When there are no available associated customers for the target customer, the relationship deviation is represented by a zero vector, and a neighbor shortage flag is written into the customer risk result record.

[0014] Secondly, embodiments of the present invention provide a customer risk prediction system based on transaction behavior characteristics, including a data storage device and a processor connected to the data storage device; The data storage device stores a transaction behavior event table, a product exposure status table, an account status table, a market status table, a reference set version table, a residual behavior sequence table, and a customer relationship correction table. The processor reads data from the data storage and executes a customer risk prediction method based on transaction behavior characteristics; The customer relationship correction table records the relationship source marker, common impact weight component, deletion status, relationship anomaly records, and directed relationship weight of candidate customer relationships.

[0015] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the customer risk prediction method based on transaction behavior characteristics as described in the first aspect of the present invention.

[0016] Through the above technical solution, the present invention can achieve at least the following beneficial effects: The trading platform server of this invention acquires trading behavior events, product exposure data, and market status data according to the predicted cutoff time, and forms a dynamic reference customer set based on the product exposure data and market status data. The conditionally expected trading behavior characteristics are derived from customer trading behavior under the same market conditions and similar product exposure conditions. The actual trading behavior characteristics of the target customer are then compared with these conditionally expected trading behavior characteristics to form customer-specific trading behavior deviation characteristics and residual trading behavior sequences. Therefore, common market fluctuations and inherent product fluctuations are first incorporated into the reference benchmark. The risk prediction results can correspond to the target customer's trading deviation relative to the reference customer, reducing the probability that normal stop-loss, passive portfolio adjustments, or product switching are identified as abnormal customer trading.

[0017] For customer relationship data, the trading platform server marks candidate customer relationships with shared product exposure and changes in trading behavior occurring within the same market state period as common shock candidate relationships. These relationships are then verified using the number of similar non-zero deviation states, sequential changes, and durations in the residual trading behavior sequence. Customer relationships formed solely by shared product exposure or a shared market state but not maintained as a connection in the residual trading behavior sequence are deleted. The remaining relationship source weights are determined by residual behavior fragments and durations, and participate in normalized aggregation using directed relationship weights. Therefore, when popular or broad-based products connect a large number of customers, the risk signals generated by shared product nodes are corrected before relationship aggregation, reducing the probability of the same market shock being repeatedly included in product characteristics, customer profit and loss characteristics, and customer relationship propagation characteristics.

[0018] The dynamic reference client set is generated using a combination of fixed and relaxed matching fields. Market status, leverage status, and position concentration status correspondences remain fixed within the same reference set generation round. Product category, position stage, and account lifecycle correspondences are relaxed progressively with each reference set version. Each relaxation records the prediction cutoff time, status version, relaxed fields, and the number of clients included. Therefore, even when the number of reference clients is insufficient, the reference set still maintains key conditions related to the target client's risk exposure, and the expected trading behavior characteristics have clear data sources and applicable versions.

[0019] The residual transaction behavior sequence is time-series encoded to form a time-series deviation representation, which is then used to aggregate the time-series encoded results of related customers by jointly impacting the relationship weights in the customer relationship data. These two representations, along with product exposure data, form a joint customer representation. The results of centralized capital outflow risk and transaction demand decay risk are first generated at the customer level and then aggregated into expected capital outflow demand and transaction demand changes according to the exposure correspondence between customers and financial products. Therefore, the product-level demand forecast records can distinguish between the demand portion corresponding to the common market state and the customer-specific demand change portion, allowing product redemption demand and net transaction demand to correspond to changes in transaction behavior from different sources. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0021] Figure 1 This is a flowchart of the customer risk prediction method based on transaction behavior characteristics in the embodiment. Figure 2 This is a framework diagram of the customer risk prediction system based on transaction behavior characteristics in the embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way. Example 1

[0024] like Figure 1 As shown, this embodiment employs a customer risk prediction method based on transaction behavior characteristics, executed by the trading platform server, to predict customer risk for financial products, including: Step S1: Obtain the transaction behavior event set, product exposure data, account status data, and market status data of the target customers and candidate customers up to the prediction deadline; In this embodiment, the prediction cutoff time is the latest recorded moment when the trading platform server is allowed to read trading events, product exposure status, account status, and market status when generating risk prediction results. The trading event's awareness time is the moment the trading platform server receives and completes status confirmation; the product exposure status's awareness time is the moment the product exposure status table is written to the new version; the account status's awareness time is the moment the account status table is written to the new version; and the market status's awareness time is the moment the market status table is written to the new version. For the same prediction cutoff time, the trading platform server truncates the trading events, product exposure status, account status, and market status according to the awareness time. Records formed or only become known after the prediction cutoff time are not included in this risk prediction.

[0025] For example, the observation period can be 5 trading days prior to the predicted cutoff time, and the event period can be divided into 30 minutes, 1 hour, or 1 trading day. In other implementations, the observation period can be 1 to 30 trading days, and the length of the event period can be adjusted according to the product trading frequency and the market status update frequency. When the event period is too short, occasional order cancellations or small redemptions by individual customers are prone to forming isolated deviations; when the event period is too long, trading behaviors before and after market status changes will be merged into the same period.

[0026] Step S2 generates reference matching conditions from product exposure data, account status data, and market status data, and forms a dynamic reference customer set after excluding target customers; Candidate customers are those who have already engaged in transaction activities and have product exposure and account status records before the prediction deadline. The trading platform server truncates the candidate customers' transaction activities, product exposure, and account status according to the prediction deadline, excludes target customers from the candidate customers, and then performs reference matching.

[0027] When the market status version, product exposure status version, or account status version read at the end of this forecast is inconsistent with the corresponding version associated with the previous reference version, the trading platform server updates the current reference version and regenerates the reference matching conditions and dynamic reference customer set.

[0028] Step S3: Based on the dynamic reference customer set, generate conditional expected transaction behavior characteristics, compare the actual transaction behavior characteristics of the target customer and related candidate customers with the corresponding conditional expected transaction behavior characteristics according to the event time period, and form a multi-customer residual transaction behavior sequence. When generating multi-customer residual transaction behavior sequences, the transaction platform server generates residual transaction behavior sequences for target customers and related candidate customers respectively, according to a unified market status version, a unified event time period boundary, and a unified residual status coding rule. Step S4: Generate candidate customer relationships with relationship source tags based on the similarity of original transaction behaviors, common product exposure relationships, and the time period of the transaction behaviors, and mark the weight components corresponding to the common product exposure relationships as common impact weight components; Residual behavior verification is performed on candidate customer relationships that meet the conditions for participation in the accumulation phase. When both ends of a candidate customer relationship maintain the same non-zero deviation state, or when the non-zero deviation state of one end of the customer relationship leads that of the other end of the customer relationship and continues to reach the preset time period quantity, the corresponding weight component is retained and a directed relationship weight is generated. If the residual behavior verification conditions are not met, the common impact weight component is deleted. If the conditions for participation in the accumulation phase are not met, a relationship anomaly record is generated. Based on the retained weight component, the directed relationship weight, and the relationship anomaly record, de-common impact customer relationship data is generated.

[0029] When the same candidate customer relationship simultaneously exhibits similarity in original transaction behavior and shared product exposure, the transaction platform server records the weight components corresponding to the similarity in original transaction behavior and the weight components corresponding to the shared product exposure, respectively. When the shared impact weight component is deleted, the weight component corresponding to the similarity in original transaction behavior participates in the generation of subsequent directed relationship weights. Step S5: Perform time-series coding on the multi-customer residual transaction behavior sequence, and normalize and aggregate the time-series coding results of related customers pointing to the target customer according to the de-common impact customer relationship data to generate a joint customer representation; During normalized aggregation, the trading platform server uses the weights of directed relationships pointing to the target customer as the weights of the incoming edges, weights the temporal deviation representations of related customers, and uses the sum of the incoming edge weights as the normalization factor. When the sum of the incoming edge weights is 0, the relationship deviation representation adopts a zero vector, and a neighbor deficiency flag is written into the customer risk result record; Step S6: Based on the joint client representation, output the results of concentrated capital outflow risk and transaction demand decay risk, and aggregate and output the expected capital outflow demand and transaction demand change.

[0030] The centralized outflow risk result is used to characterize the outflow of funds from clients during the future forecast window due to redemptions, sales, liquidation, or position transfers. The change in trading demand is used to characterize the change in the effective trading amount or number of effective transactions from clients during the future forecast window relative to the historical benchmark.

[0031] Furthermore, the transaction behavior event set is obtained by standardizing the original transaction events from different transaction channels. The standardized transaction behavior event includes at least customer identifier, product identifier, event type, event time, time of awareness, transaction direction, transaction amount, processing status, original order identifier, cancellation association identifier, and correction association identifier. The transaction platform server associates the order cancellation event and order modification event with the corresponding original transaction event based on the original order identifier, cancellation association identifier, or correction association identifier. Based on the associated transaction behavior event, it determines at least one of the following as the actual transaction behavior characteristics: net subscription behavior, net redemption behavior, buy / sell direction switching behavior, order cancellation behavior, product switching behavior, position concentration status change behavior, and transaction interval behavior.

[0032] The original transaction behavior similarity is generated by the trading platform server based on at least one of the following during the observation period: the proportion of fund outflows, the proportion of changes in transaction amount, the proportion of changes in the number of transactions, the order cancellation rate, the product switching rate, and the number of times the transaction direction was switched. For any two customers, the trading platform server first aligns the above transaction behavior characteristics according to the same event period, then calculates the differences or the number of identical states between the corresponding transaction behavior characteristics, and records customer pairs whose differences are below the similarity threshold or whose number of identical states reaches the similarity threshold as having original transaction behavior similarity. Original transaction behavior similarity is used to generate candidate customer relationships and is not used as a retention condition for jointly impacting candidate relationships.

[0033] Furthermore, product exposure data includes product category, holding stage, holding concentration status, and leverage status; account status data includes account opening duration, historical valid trading days, and the time of the most recent valid trading; market status data includes market volatility status, product price or net asset value changes, and market trading activity status; and the reference matching conditions include product matching relationships, holding stage matching relationships, holding concentration status matching relationships, leverage status matching relationships, account lifecycle matching relationships, and market status matching relationships.

[0034] Specifically, the product category is determined by the product type, investment scope, and risk stratification fields in the product basic information table; the holding phase is determined by the interval between the moment when the target client first forms an effective position and the prediction cutoff time; the position concentration status is determined by the proportion of the target client's holding market value in a single financial product to the total holding market value of their financial products; the leverage status is determined by the margin balance, short selling balance, margin usage amount, or leveraged share identifier. The market volatility status is jointly determined by the market index return volatility, the product's industry index volatility, and the target financial product's net asset value volatility; the product price or net asset value change status is determined by the price change rate or net asset value change rate of the target financial product during the observation period; the market trading activity status is determined by the changes in transaction amount, number of transactions, and number of subscriptions and redemptions relative to the historical benchmark.

[0035] Market status version, product exposure status version, and account status version are recorded using version number, version effective time, and version applicable range, respectively. When the market status version, product exposure status version, or account status version read at the current prediction deadline is inconsistent with the corresponding version associated with the previous reference version, the trading platform server updates the current reference version and regenerates the reference matching conditions and dynamic reference customer set. When the current prediction deadline falls within the applicable range of a certain version, the trading platform server reads the data corresponding to that version. When the applicable ranges of multiple versions overlap, the version with the latest effective time and whose knowledge time is not later than the prediction deadline is used; when there is a break in the applicable range, the trading platform server uses the most recently known market status version, product exposure status version, and account status version before the breakpoint to generate reference matching conditions and records a missing mark for the combined status applicable range in the reference set version table.

[0036] Furthermore, when generating the residual trading behavior sequence, the trading platform server identifies the time points of market state changes based on market state data, and divides multiple event periods using the time points of market state changes as the relative time starting point; it generates conditional expected trading behavior characteristics based on the capital outflow status, trading activity status, order cancellation status, position concentration status, and product switching status of the dynamic reference customer set within each event period; it subtracts the actual trading behavior characteristics of the target customer within the same event period from the conditional expected trading behavior characteristics, performs a difference binning or a level offset, and generates the customer-specific trading behavior deviation characteristics.

[0037] The timing of market state changes is determined by the effective time of the market state version. When multiple market state changes occur within the same observation period, the trading platform server uses the market state change closest to the prediction cutoff time and reaching the preset change intensity as the relative time starting point. If no market state change reaching the preset change intensity is detected, the trading platform server uses the start time of the observation period as the relative time starting point. The preset change intensity is used to limit the minimum change range of market volatility, product price or net asset value changes, and market trading activity; the preset change intensity is determined through a validation set based on historical data volume and risk identification error within the future prediction window.

[0038] Conditional expected trading behavior characteristics are generated according to the type of deviation from the expected trading behavior. For outflow status, the trading platform server calculates the ratio of net redemption amount to market value of holdings for the dynamic reference client set within the same event period, and uses the median or truncated mean as the conditional expected value. For active trading status, the changes in the number of transactions, transaction amount, and effective trading days relative to the historical benchmark are calculated. For order cancellation status, the ratio of cancelled orders to submitted orders is calculated. For concentrated holdings status, the change in the holding ratio of a single product is calculated. For product switching status, the percentage of the amount switched from the target financial product to other financial products is calculated. The truncation ratio of the truncated mean can be 5% to 20%. If the truncation ratio is too small, extreme client behavior will affect the conditional expected value; if the truncation ratio is too large, the true concentrated behavior in the dynamic reference client set will be weakened.

[0039] When the dynamic reference customer set is marked with a low sample size, the conditional expected transaction behavior feature is simultaneously written to a low reliability level. The low reliability level indicates that the conditional expected transaction behavior feature was generated from customer samples below a set size threshold. The low reliability level participates in subsequent risk outcome recording, but the conditional expected transaction behavior feature corresponding to the low reliability level is not used as a sample source for updating the judgment threshold.

[0040] Furthermore, when generating the joint customer representation, the trading platform server inputs the multi-customer residual transaction behavior sequence into the time-series encoder to obtain the time-series deviation representation of each customer; according to the directed relation weights in the de-settling customer relationship data, the time-series deviation representations of related customers pointing to the target customer are weighted, normalized, and aggregated to obtain the relation deviation representation; the time-series deviation representation of the target customer, the relation deviation representation, and the product exposure data are concatenated into the joint customer representation, and the results of concentrated capital outflow risk are generated through the classification head, and the results of transaction demand decay risk are generated through the regression head.

[0041] The input to the time-series encoder is a sequence of residual trading behaviors arranged by event time periods. Each event time period corresponds to at least one residual feature from the following: fund outflow status, trading activity status, order cancellation status, position concentration status, and product switching status. The time-series encoder can employ a gated recurrent network, a temporal convolutional network, or an attention-based sequence coding network. The training samples consist of residual trading behavior sequences corresponding to historical prediction cutoff times, de-impact customer relationship data, and product exposure data. The label for the concentrated fund outflow status classification is determined by whether the proportion of customer fund outflows within the future prediction window reaches the concentrated fund outflow threshold. The label for the transaction demand change prediction is determined by the decrease in the customer's effective transaction amount or number of effective transactions within the future prediction window relative to the historical benchmark. The classification head is trained using cross-entropy loss, and the regression head is trained using absolute error loss or squared error loss. The training samples are divided into a training set and a validation set according to the prediction cutoff time, with the prediction cutoff time in the validation set being later than that in the training set. The trading platform server determines the time encoder parameters, classification head parameters, regression head parameters, and threshold calibration parameters based on the identification error of the centralized capital outflow status and the prediction error of changes in trading demand on the validation set.

[0042] In one implementation, the time-series encoder employs a two-layer gated recurrent network. The input vector for each event period is obtained by concatenating the data on fund outflow status, trading activity status, order cancellation status, position concentration status, product switching status, and residual validity labels. The first and second gated recurrent networks are sequentially connected, and the output of the last event period of the second gated recurrent network serves as the time-series deviation representation. The classification head consists of a fully connected layer and a normalized output layer, while the regression head consists of a fully connected layer and a linear output layer. Training samples are divided into a training set and a validation set according to the prediction cutoff time, with the prediction cutoff time in the validation set being later than that in the training set. The trading platform server determines the network parameters and threshold calibration parameters based on the identification error of concentrated fund outflow status and the prediction error of changes in trading demand on the validation set.

[0043] The concentrated outflow threshold is determined by the ratio of net redemptions by clients within the future forecast window to the market value of their holdings at the forecast cutoff time. The concentrated outflow threshold can be calibrated based on the product's liquidity level, historical redemption distribution, and the length of the future forecast window, typically ranging from 10% to 40% in one implementation. The historical benchmark is determined by the average effective transaction amount, average number of effective transactions, or average number of effective trading days prior to the forecast cutoff time. The statistical period corresponding to the historical benchmark can be 20, 60, or 120 trading days.

[0044] After model deployment, the trading platform server generates time series deviation representations and relational deviation representations in batches according to the prediction cutoff time. When the number of event periods in the residual transaction behavior sequence where the residual cannot be generated exceeds half of the sequence length, the trading platform server generates a time series deviation representation with a low reliability level; when there are no available related customers for the target customer, the relational deviation representation uses a zero vector, and a neighbor shortage flag is written into the customer risk result record.

[0045] Furthermore, when generating the expected capital outflow demand and changes in trading demand, the trading platform server, according to customer identifiers, product identifiers, and future prediction windows, aggregates the concentrated capital outflow risk results and trading demand decay risk results of multiple customers with the customer's product holding shares, holding market value, or leverage balance, and generates a product demand forecast record. The product demand forecast record includes the demand portion corresponding to the common market state obtained based on the characteristics of expected trading behavior, and the customer-specific demand change portion obtained based on the deviation characteristics of the customer's specific trading behavior.

[0046] Product demand forecast records are established using product identifiers and future forecast windows as indexes. When the centralized capital outflow risk result is a graded result, the trading platform server aggregates the data according to the capital outflow ratio range corresponding to the grade and the client's product holding shares or market value. When the centralized capital outflow risk result is a probability result, the trading platform server aggregates the data according to the mapping relationship between probability values ​​and predicted capital outflow ratios. The mapping relationship between probability values ​​and predicted capital outflow ratios can be generated using a segmented mapping table. For example, when the probability value of the centralized capital outflow risk result is less than 0.3, the predicted capital outflow ratio is taken from the first capital outflow ratio range; when the probability value is greater than or equal to 0.3 and less than 0.7, the predicted capital outflow ratio is taken from the second capital outflow ratio range; and when the probability value is greater than or equal to 0.7, the predicted capital outflow ratio is taken from the third capital outflow ratio range. Each capital outflow ratio range is determined by the product liquidity level and historical redemption distribution corresponding to the product demand forecast record.

[0047] When the risk of declining transaction demand is reflected in the magnitude of the decline, the trading platform server applies this magnitude of decline to the customer's historical valid transaction amount or valid number of transactions to obtain the corresponding change in the customer's transaction demand.

[0048] The demand component corresponding to the common market state is derived from the changes in transaction demand corresponding to the expected transaction behavior characteristics. The customer-specific demand component is derived from the changes in transaction demand corresponding to the deviation characteristics of customer-specific transaction behavior. Both are written into the product demand forecast record and indexed and associated with the forecast cutoff time, reference set version, residual behavior sequence version, and customer relationship correction version. When there are records with low reliability levels, insufficient neighbor markers, or abnormal relationships, the corresponding markers are simultaneously written into the product demand forecast record.

[0049] In one embodiment, when the trading platform server performs residual behavior verification on the two clients marked as joint impact candidates, it uses the event time period as the smallest judgment unit and performs state-based processing on each type of trading behavior deviation feature in the states of fund outflow, trading activity, order cancellation, concentrated holdings, and product switching.

[0050] For any customer, any event period, and any deviation characteristic of a transaction behavior, the trading platform server converts the residual magnitude in the residual transaction behavior sequence into three states: positive deviation, no significant deviation, and negative deviation.

[0051] in, Indicates customer During the event period Internal targeting of trading behavior deviation characteristics Deviation state, Indicates the customer sequence number. Indicates the sequence number of the event period. This indicates that the transaction behavior deviates from the characteristic type. Indicates customer During the event period Internal targeting of trading behavior deviation characteristics The residual amplitude, Indicates deviation from trading behavior characteristics The corresponding positive deviation judgment threshold, Indicates deviation from trading behavior characteristics The corresponding negative deviation judgment threshold.

[0052] The residual magnitude is obtained by comparing the actual transaction behavior characteristics and the conditionally expected transaction behavior characteristics of the target customer within the same event period. When both the actual and conditionally expected transaction behavior characteristics are numerical ratios, the residual magnitude is the difference between the two; when both are rank states, the residual magnitude is the rank difference; when both are binned results, the residual magnitude is the binning rank difference. Positive and negative deviation states are non-zero deviation states. When two non-zero deviation states have the same value, the two non-zero deviation states belong to the same type of non-zero deviation state.

[0053] When a target customer lacks valid trading activity events during a certain event period, the trading platform server processes the data based on whether a valid open position exists during that event period. If a valid open position exists but no trading activity events occur, the trading activity characteristic for that event period is recorded as a "no trading" state. If a valid open position is missing, the event period is recorded as a "residual non-generable" state and written into the residual anomaly record. The residual non-generable state is not included in the accumulation of the number of unidirectional change segments, the number of sequentially related change segments, and the number of duration periods. The residual anomaly record is used to record the generation status of residual trading activity sequences within the event period, and the relationship anomaly record is used to record the cumulative status of candidate customer relationships.

[0054] The positive and negative deviation thresholds are calibrated based on the residual amplitude distribution of the dynamic reference customer set during the observation period. The calibration calibrator uses quantiles or standard deviation multiples under the same product exposure risk level, the same market state version, and the same type of trading behavior deviation characteristic. Both the positive and negative deviation thresholds are positive values ​​and are limited to the minimum and maximum thresholds of the corresponding trading behavior deviation characteristics, respectively. The minimum and maximum thresholds are both preset threshold boundaries greater than 0. The positive and negative deviation thresholds increase with increasing market volatility and decrease with decreasing market volatility. When the number of samples used for calibration is less than the threshold sample number, the threshold that was already in effect at the previous prediction cutoff time is used. If no available threshold exists at the previous prediction cutoff time, a preset conservative threshold is used.

[0055] In one implementation, the positive deviation and negative deviation thresholds are defined separately according to the type of deviation characteristic of the transaction behavior, and are generated using quantiles of the residual amplitude distribution of the dynamically referenced customer set. The quantiles can range from 60% to 95%, and are determined through a validation set based on the amount of historical data and the risk identification error within the future prediction window. If the threshold is too small, ordinary transaction fluctuations are easily marked as non-zero deviations; if the threshold is too large, abnormal transaction changes by the customer themselves may be classified as having no significant deviation.

[0056] Before threshold calibration, the trading platform server performs validity checks on the residual magnitudes participating in the calibration. Validity checks include field completeness checks, event period alignment checks, reference set version consistency checks, and extreme value checks. Samples with missing fields, inconsistent event periods, or inconsistent reference set versions are not included in threshold calibration; samples exceeding preset extreme value boundaries are written into the threshold anomaly record and processed according to the truncated standard.

[0057] For clients who jointly impact both ends of the candidate relationship and The trading platform server during continuous event periods to Internal judgments change in the same direction.

[0058] If both ends of the customer's account are in the same non-zero deviation state for no less than a preset number of consecutive event periods, then it is determined that both ends of the customer's account change in the same direction during those consecutive event periods.

[0059] in, Indicates customer With customers During the event period to Internal targeting of trading behavior deviation characteristics The result of the same-direction change determination, This indicates the customer ranking at one end of the candidate relationship that is jointly impacting the candidate. This indicates the customer ranking at the other end of the candidate relationship, which jointly impacts the candidate's position. Indicates the starting position of a continuous event period. Indicates the end sequence of a period of consecutive events. This represents the minimum number of consecutive event periods required for a change in the same direction. This is an indicator function that takes a value of 1 when the condition is met and a value of 0 when the condition is not met. Representing logical AND relation, , , and Refer to the aforementioned definition.

[0060] For sequential changes, the trading platform server uses an event time period as a lag benchmark for judgment, and the continuous event time periods used for determining sequential changes start from the event time period with the preceding event time period. If the customer... The non-zero deviation that occurred in the previous event period, and the customer If a non-zero deviation state occurring in a subsequent event period belongs to the same type of non-zero deviation state, and this correspondence continues to occur within a number of consecutive event periods not less than a preset number, then the customer is determined to be... To customers There are sequential changes:

[0061] in, Indicates customer To customers During the event period to Internal targeting of trading behavior deviation characteristics The results of determining the sequential relationship changes, Indicated by the customer Deviation state leading the customer The direction of correlation formed by the deviation state This represents the minimum number of consecutive event periods required for sequential changes. Indicates customer During the event period Deviation characteristics of trading behavior during the previous event period Deviation state, Referring to the aforementioned definition, , , , and Refer to the aforementioned definition.

[0062] Trading platform servers to clients To customers The sequential changes are evaluated using the same reverse rules. These sequential changes form a directed relationship weight. (Customer) To customers The sequential changes correspond to the customer. Directed relation weights, clients To customers The sequential changes correspond to the customer. The weight of directed relationships. When the trading platform server aggregates relationships for target customers, it uses the weight of directed relationships pointing to the target customer, and does not use the weight of directed relationships pointing from the target customer to other customers.

[0063] When performing residual behavior verification on candidate customer relationships containing common impact weights, the trading platform server first determines whether the candidate customer relationship has a period of time eligible for cumulative events. If a candidate customer relationship containing common impact weights lacks a period of time eligible for cumulative events, the trading platform server generates an anomaly record and excludes the candidate customer relationship from the aggregation of time-series coding results for related customers in this prediction. If a common impact candidate relationship has a period of time eligible for cumulative events, the trading platform server determines whether it meets the criteria for same-direction change or sequential association change. If a common impact candidate relationship has at least one consecutive event period window that meets the same-direction change criterion under any transaction behavior deviation feature, or at least one consecutive event period window that meets the sequential association change criterion in any direction under any transaction behavior deviation feature, the trading platform server retains the common impact candidate relationship. If a common impact candidate relationship has a period of time eligible for cumulative events, and there are no consecutive event period windows that meet the same-direction change criterion under any transaction behavior deviation feature, and there are no consecutive event period windows that meet the sequential association change criterion in either direction, the trading platform server deletes the common impact candidate relationship.

[0064] For candidate customer relationships not marked as co-impact candidate relationships, the trading platform server determines the directed relationship weight based on the similarity of the original transaction behavior and writes the candidate customer relationship into the co-impact customer relationship data. For candidate customer relationships marked as co-impact candidate relationships but lacking a period of time that can participate in the cumulative event, the trading platform server writes a relationship anomaly record. The relationship anomaly record includes the candidate customer relationship identifier, the type of transaction behavior deviation feature that is missing a period of time that can participate in the cumulative event, the prediction cutoff time, and the corresponding reference set version. This candidate customer relationship will not participate in the aggregation of the time-series coding results of related customers in this prediction.

[0065] Relationship anomaly records are used to document the eligibility status of candidate customer relationships for cumulative aggregation. When a customer identifier is missing, a product identifier is missing, an event time is missing, a processing status is invalid, there is a gap in the event timeframe, a market status version breakpoint, a product exposure status version breakpoint, or the number of eligible event timeframes is lower than the corresponding minimum number of consecutive event timeframes, the trading platform server writes a relationship anomaly record and stops using the corresponding candidate customer relationship for this relationship aggregation. When the dynamically referenced customer set has a low-sample flag, the corresponding expected transaction behavior characteristics are written to a low reliability level, and relationship anomaly records are not triggered independently.

[0066] Minimum number of consecutive event periods for unidirectional changes and the minimum number of consecutive event periods used for sequential changes. Based on the length of the observation period, both are positive integers and are greater than or equal to the corresponding preset lower limit value; It is limited to the number of consecutive event periods that can participate in the accumulation of changes in the same direction. The number of consecutive event periods that can be accumulated based on sequential changes is limited. The minimum number of consecutive event periods is increased when the observation period is long, and a preset lower limit is used when the observation period is short, to avoid customer relationships being formed solely from occasional transactions within a single event period.

[0067] For candidate customer relationships that retain the common impact weight component, the trading platform server determines the directed relationship weight based on the number of unidirectional change segments, the number of sequentially related change segments, and the duration of the relationship:

[0068] in, Indicates customer Pointing to customers The directed relation weights, This indicates the upper limit of the directed relation weight. The underlying weights of the relationships that are preserved. This represents the set of deviation feature types for transaction behaviors that participate in the calculation of directed relation weights. Indicates deviation from trading behavior characteristics The corresponding weighting coefficients of segments changing in the same direction, Indicates customer With customers Targeting the characteristics of trading behavior deviation The number of consecutive segments that satisfy the criterion of same-direction change. Indicates deviation from trading behavior characteristics The corresponding weight coefficients of sequentially related change segments, Indicates customer With customers Directional deviation characteristics of trading behavior The number of consecutive segments that satisfy the criteria for sequential change. Indicates deviation from trading behavior characteristics The corresponding weight coefficients of sequentially related change segments, Indicates customer With customers Targeting the characteristics of trading behavior deviation The maximum number of consecutive event periods that meet the criteria for same-direction change or sequential related change; Customer To customers The weights of directional relationships are calculated separately using the same method.

[0069] The weight coefficients of the same-direction change segments, the weight coefficients of the sequentially related change segments, and the weight coefficients of the duration period are determined by the concentrated capital outflow events, transaction demand decay events, and ordinary transaction events that have been marked in the historical observation period, and the values ​​are greater than or equal to 0. The value of is greater than or equal to 0. The value is greater than or equal to When the number of segments changing in the same direction, the number of segments changing sequentially, or the number of duration periods increases, the weight of the directed relationship remains unchanged, and is further enhanced by... Implement saturation limits.

[0070] The update frequency of directed relation weights is consistent with the generation frequency of de-impact customer relationship data; when the new prediction cutoff time arrives and a new residual transaction behavior sequence is formed, the transaction platform server re-counts the number of the above segments and the number of duration periods.

[0071] The number of unidirectional change segments represents the number of segments where both clients continuously satisfy the unidirectional change criterion under the same non-zero deviation state; the number of sequentially related change segments represents the number of segments where one client's non-zero deviation state appears before the other client and continuously satisfies the sequentially related change criterion; the number of duration segments represents the longest continuous event duration that has satisfied either the unidirectional change criterion or the sequentially related change criterion. When counting the number of unidirectional change segments and sequentially related change segments, the trading platform server merges overlapping continuous event duration windows into the same continuous segment according to the event duration order, and the same continuous segment is counted once. The directed relationship weight is used to limit the contribution ratio of the temporal coding results of related clients in the relationship aggregation. For example, the upper limit of the directed relationship weight can be 1, and the base weight can be 0~0.2; the weight coefficients of unidirectional change segments, sequentially related change segments, and duration segments can be 0~0.5. When the weight coefficient is too small, the impact of residual behavior association on relationship aggregation is weakened; when the weight coefficient is too large, a small number of continuous deviation segments will make a significant contribution to the aggregation results of related clients.

[0072] If any client lacks valid residual transaction behavior data within a certain event period, that event period will not participate in the accumulation of same-direction changes and sequential related changes under the corresponding transaction behavior deviation feature; if the number of event periods that can participate in the accumulation for same-direction changes under a certain transaction behavior deviation feature is lower than the minimum number of consecutive event periods required for same-direction changes, then that transaction behavior deviation feature will not participate in the determination of retention for same-direction changes; if the number of event periods that can participate in the accumulation for sequential related changes under a certain transaction behavior deviation feature is lower than the minimum number of consecutive event periods required for sequential related changes, then that transaction behavior deviation feature will not participate in the determination of retention for retention for sequential related changes. Example 2

[0073] Based on Example 1, this example provides a specific way to relax the product matching correspondence, holding stage correspondence, or account life cycle correspondence in a preset order in the customer risk prediction method based on transaction behavior characteristics. In one implementation, when the trading platform server forms a dynamic reference client set according to the predicted cutoff time, it uses the market status correspondence, leverage status correspondence, and position concentration status correspondence as fixed matching fields, which remain fixed during the same round of reference set generation; and uses the product matching correspondence, position stage correspondence, and account lifecycle correspondence as relaxed matching fields, which are relaxed step by step in the order of product matching correspondence, position stage correspondence and account lifecycle correspondence.

[0074] In the unrestricted benchmark matching version, product matching is based on the same product identifier, holding period matching is based on the same holding days range, and account lifecycle matching is based on the same account lifecycle range. After excluding target clients from the candidate clients, the trading platform server writes clients that meet the benchmark matching conditions according to the current benchmark set version and counts the number of clients written: , in, Indicates the first Number of clients writing to each reference set version Indicates the version order of the reference set. Indicates according to the first A dynamic reference customer set is obtained by relaxing the matching criteria and fixing the matching fields in the reference set version.

[0075] when When the number of matching fields falls below the threshold, the trading platform server generates the next reference set version and relaxes only one matching field corresponding to the current order. The relaxation order is as follows: the matching relationship between products is relaxed from the same product identifier to the same product subclass; the product subclass is jointly determined by the product type field, investment scope field, and risk stratification field in the product basic information table. For example, equity funds can be subcategorized according to broad-based indices, industry themes, actively managed equity, and quantitative equity, while bond funds can be subcategorized according to interest rate bonds, credit bonds, convertible bonds, and short-term bonds.

[0076] If the number of clients added is still below the threshold for the total number of clients, the product matching relationship will be further relaxed to correspondence within the same product exposure risk level. The product exposure risk level is pre-determined by the trading platform server based on the product category, product net asset value fluctuation range, leverage status, and preset risk stratification rules. If the number of clients added is still below the threshold for the total number of clients, while maintaining the product matching relationship as correspondence within the same product exposure risk level, the holding phase relationship will be relaxed from correspondence within the same holding days interval to correspondence within adjacent holding days intervals. Adjacent holding days intervals refer to holding days intervals whose interval order difference from the target client's holding days interval is one. If the number of clients added is still below the threshold for the total number of clients, while maintaining the market status relationship, leverage status relationship, and holding concentration status relationship, the account lifecycle relationship will be relaxed from correspondence within the same account lifecycle interval to correspondence within adjacent account lifecycle intervals. The account lifecycle interval is jointly defined by the account opening duration, the number of historical valid trading days, and the time of the most recent valid trading session.

[0077] Each time the restrictions are relaxed, the trading platform server generates a new reference set version and records the prediction cutoff time, market status version, product exposure status version, account status version, current reference set version sequence, relaxed matching fields, matching criteria before relaxation, matching criteria after relaxation, and the corresponding number of clients written to the reference set version table.

[0078] The update frequency of the reference set version is consistent with the prediction cutoff time. When the market status version or product exposure status version changes, the trading platform server restarts the relaxation process from the baseline matching version. If the number of writing clients obtained according to the above maximum relaxation caliber is still lower than the set quantity threshold, the dynamic reference client set formed under the maximum relaxation caliber is adopted, and a low sample mark is recorded for this reference set version. In abnormal situations where there are missing market status, missing leverage status, or missing position concentration status, the trading platform server no longer relaxes the fixed matching fields, but uses the reference set version corresponding to the most recent complete market status version, the most recent complete product exposure status version, and the most recent complete account status version as the conservative reference version. Example 3

[0079] Based on Embodiments 1 and 2, this embodiment provides a customer risk prediction system based on transaction behavior characteristics, such as... Figure 2 As shown, it includes a data memory and a processor connected to the data memory; The data storage device stores the transaction behavior event table, product exposure status table, account status table, market status table, reference set version table, residual behavior sequence table, and customer relationship correction table. The account status table records at least the customer identifier, account opening duration, number of historical valid transaction days, time of the most recent valid transaction, account lifecycle range, account status version, version effective time, time of notification, and version applicable range.

[0080] The processor reads data from the data storage and executes a customer risk prediction method based on transaction behavior characteristics; the reference set version table records the correlation between the dynamic reference customer set and the market status version, product exposure status version, and prediction cutoff time; the residual behavior sequence table records the correspondence between actual transaction behavior characteristics, conditional expected transaction behavior characteristics, customer-specific transaction behavior deviation characteristics, and event time periods; and the customer relationship correction table records the relationship source markers, common impact weight components, deletion status, relationship anomaly records, and directed relationship weights of candidate customer relationships.

[0081] The transaction event table records at least the customer identifier, product identifier, event type, event time, time of awareness, transaction direction, transaction amount, processing status, original order identifier, cancellation association identifier, and correction association identifier. The product exposure status table records at least the customer identifier, product identifier, shareholding, market value of holdings, product category, holding stage, concentration status of holdings, leverage status, version effective time, time of awareness, and version applicable range. The market status table records at least the product identifier, market volatility status, product price or net asset value change status, market trading activity status, market status version, version effective time, time of awareness, and version applicable range. The reference set version table, residual behavior sequence table, and customer relationship correction table all record the prediction cutoff time, market status version, product exposure status version, and account status version, enabling the same risk outcome to be located to the corresponding data version and intermediate processing result. Example 4

[0082] Based on Embodiments 1 to 3, this embodiment provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by the processor of a trading platform server, causes the processor to perform a customer risk prediction method based on trading behavior characteristics.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0084] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A customer risk prediction method based on transaction behavior characteristics, executed by a trading platform server, characterized in that: include: Obtain transaction behavior event sets, product exposure data, account status data, and market status data of target customers and candidate customers up to the prediction deadline; Reference matching conditions are generated from the product exposure data, the account status data, and the market status data, and a dynamic reference customer set is formed after excluding target customers. Based on the dynamic reference customer set, conditional expected transaction behavior characteristics are generated. The actual transaction behavior characteristics of the target customer and related candidate customers are compared with the corresponding conditional expected transaction behavior characteristics according to the event time period to form a multi-customer residual transaction behavior sequence. Candidate customer relationships with relationship source tags are generated based on the similarity of original transaction behavior, common product exposure relationship and the time period of the transaction behavior, and the weight component corresponding to the common product exposure relationship is marked as the common impact weight component; Perform residual behavior review on candidate customer relationships that meet the criteria for participation in the cumulative process; When both ends of the customer's non-zero deviation state remains the same, or when the non-zero deviation state of one end of the customer's non-zero deviation state leads that of the other end of the customer's non-zero deviation state and continues to reach the preset time period quantity, the corresponding weight component is retained and a directed relation weight is generated. If the residual behavior review conditions are not met, delete the common impact weight component; if the conditions for participation in the accumulation are not met, generate an abnormal relationship record. Generate customer relationship data that jointly impacts the data by retaining weighted components, directed relationship weights, and relationship anomaly records; The time-series coding of multi-customer residual transaction behavior sequences is performed, and the time-series coding results of related customers pointing to the target customer are normalized and aggregated according to the de-common impact customer relationship data to generate a joint customer representation; Based on the combined client representation, the output focuses on the risk of concentrated capital outflow and the risk of declining transaction demand, and aggregates and outputs the expected changes in capital outflow demand and transaction demand.

2. The customer risk prediction method based on transaction behavior characteristics according to claim 1, characterized in that, The transaction behavior event set is obtained by standardizing the original transaction events from different transaction channels. The standardized transaction behavior event includes at least customer identifier, product identifier, event type, event time, time of awareness, transaction direction, transaction amount, processing status, original order identifier, cancellation association identifier, and correction association identifier. The transaction platform server associates cancellation events and modification events with the corresponding original transaction events based on the original order identifier, cancellation association identifier, or correction association identifier, and generates the actual transaction behavior characteristics based on the associated transaction behavior events.

3. The customer risk prediction method based on transaction behavior characteristics according to claim 1, characterized in that, The product exposure data includes product category, holding stage, holding concentration status and leverage status, and the account status data includes account opening duration, historical valid trading days and the most recent valid trading time. The market status data includes market volatility, product price or net asset value changes, and market trading activity. The reference matching conditions include product matching relationships, holding stage relationships, holding concentration relationships, leverage relationships, account lifecycle relationships, and market status relationships.

4. The customer risk prediction method based on transaction behavior characteristics according to claim 1, characterized in that, When the dynamic reference customer set is formed, the transaction platform server restores the product exposure status, account status and transaction behavior status of the candidate customers that have been formed and known to the transaction platform server at the same prediction cutoff time as the target customers. The trading platform server uses market status correspondence, leverage status correspondence, and position concentration status correspondence as fixed matching fields, and product matching correspondence, position holding stage correspondence, and account lifecycle correspondence as flexible matching fields. In the benchmark matching version, the product matching correspondence is the correspondence of the same product identifier, the position holding stage correspondence is the correspondence of the same position holding days range, and the account lifecycle correspondence is the correspondence of the same account lifecycle range.

5. The customer risk prediction method based on transaction behavior characteristics according to claim 4, characterized in that, When the number of customers written is lower than the threshold of the set, the trading platform server generates the next reference set version and gradually relaxes the restrictions in the order of product matching relationship, holding stage relationship and account life cycle relationship. The product matching relationship is broadened from matching the same product identifier to matching the same product subclass and then matching the same product exposure risk level. The correspondence between holding periods has been broadened from the same holding period interval to the correspondence between adjacent holding period intervals; The correspondence of account lifecycles is relaxed from the correspondence of the same account lifecycle interval to the correspondence of adjacent account lifecycle intervals; after each relaxation, the trading platform server records the prediction cutoff time, status version, relaxed matching field, matching caliber before relaxation, matching caliber after relaxation, and number of written customers in the reference set version table.

6. The customer risk prediction method based on transaction behavior characteristics according to claim 1, characterized in that, When generating the multi-customer residual transaction behavior sequence, the transaction platform server generates residual transaction behavior sequences for the target customer and related candidate customers respectively according to the unified market state version, unified event time period boundary and unified residual state coding rule; When there is no market state change point that reaches the preset change intensity within the observation period, the start time of the observation period shall be used as the relative time starting point. When the dynamic reference customer set has a low sample count, the corresponding expected transaction behavior characteristics are written into the low reliability level.

7. The customer risk prediction method based on transaction behavior characteristics according to claim 1, characterized in that, The criteria for participation in the cumulative process include that both ends of the candidate customer relationship have valid residual transaction behavior data for no less than a preset number of consecutive event periods under the same transaction behavior deviation characteristics. When performing residual behavior review on the weighted components corresponding to the common product exposure relationship, exclude the other end customer from the dynamic reference customer set of one end customer. When the common impact weight component lacks a period of time that can participate in the cumulative event, the trading platform server generates a relationship anomaly record, and the corresponding candidate customer relationship does not participate in the aggregation of the time-series coding results of related customers in this prediction.

8. The customer risk prediction method based on transaction behavior characteristics according to claim 1, characterized in that, When generating the joint customer representation, the trading platform server inputs the multi-customer residual transaction behavior sequence into the time encoder to obtain the time deviation representation of each customer; The temporal deviation representations of related customers are weighted, normalized, and aggregated according to the directed relation weights pointing to the target customer to obtain the relation deviation representation; The time-series deviation representation of the target customer, the relationship deviation representation, and the product exposure data are concatenated to form the joint customer representation; When there are no available associated customers for the target customer, the relationship deviation is represented by a zero vector, and a neighbor shortage flag is written into the customer risk result record.

9. A customer risk prediction system based on transaction behavior characteristics, characterized in that, Includes a data storage device and a processor connected to the data storage device; The data storage device stores a transaction behavior event table, a product exposure status table, an account status table, a market status table, a reference set version table, a residual behavior sequence table, and a customer relationship correction table. The processor reads data from the data storage and executes the customer risk prediction method based on transaction behavior characteristics as described in any one of claims 1 to 8; The customer relationship correction table records the relationship source marker, common impact weight component, deletion status, relationship anomaly records, and directed relationship weight of candidate customer relationships.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the customer risk prediction method based on transaction behavior characteristics as described in any one of claims 1 to 8.

Images