Prediction method for recovery of retail unfavorable assets

Abstract

The invention discloses a method for predicting recovery of retail unhealthy assets, which comprises the following steps of: S1, selecting asset data of a plurality of retail assets, such as assets in 2-7 years before a packaging day when credit card loan becomes unhealthy assets, as sample data, and inputting the sample data into a static pool, counting the recovery amount and the cumulative recovery rate CRR of each month after each asset in the static pool becomes an unhealthy asset; s2, acquiring an expansion variable according to the asset data; relates to the technical field of non-performing asset assessment, and solves the technical problem of loss of issuers and investors caused by inaccurate prediction of non-performing assets at present. A historical asset pool of a bank where assets are located is divided into a plurality of asset pools with homogeneity, and the average recovery rate of each sub-pool in each period in the divided asset pools is counted to obtain a grouping mapping basis; under the condition that the external environment does not change remarkably, the recycling condition of the same type of assets in the bank where the assets are located does not change remarkably, and high accuracy is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of non-performing asset evaluation, and in particular is a prediction of retail non-performing asset recovery. Background Art

[0002] ABS business (asset securitization business) refers to converting assets (underlying assets) that lack liquidity but can generate foreseeable cash flows into securities that can be sold and circulated in the financial market. Generally speaking, each of the underlying assets used is an asset that can estimate the planned cash flow according to the asset contract, and the dynamic pool or static pool operation mode is adopted according to the matching of the asset term and the ABS product term.

[0003] However, for non-performing assets resulting from credit card fraud, once they have become non-performing assets, there is a certain probability that repayments cannot be made as expected in the contract, so a method needs to be found to solve the problem of predicting cash flow for non-performing assets.

[0004] Since the history of disposing of non-performing assets is still short and the disposal methods of non-performing assets are constantly changing, the research on the distribution of historical recovery data is limited. In this case, the recovery prediction methods of non-performing assets are mostly simple group mapping or simulation based on Beta and other distributions. Simple group mapping lacks the necessary basis; and simulation based on Beta and other distributions is based on distribution assumptions and will bring certain systemic risks. No matter which method is used, it may lead to extremely inaccurate predictions. If the prediction is too small, a large amount of income from the static pool will flow into the hands of secondary investors, damaging the interests of the issuer; if the prediction is too large, it will lead to an over-high securities rating, damaging the interests of investors and affecting the credibility of the issuer. Therefore, the prediction of retail non-performing asset recovery is estimated. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a prediction of the recovery of retail non-performing assets, which is used to solve the technical problem that the current prediction of non-performing assets is not accurate enough, resulting in losses for issuers and investors. The present invention solves the above-mentioned problem by selecting the same grouping mapping method as the static pool when processing the asset pool, that is, corresponding each asset in the issued asset pool to the historical static pool grouping, obtaining the recovery status of each non-performing asset, and summarizing to obtain the average collection rate of the asset pool.

[0006] To achieve the above object, the first aspect of the present invention provides a forecast of retail non-performing asset recovery, comprising the following steps:

[0007] Step 1: Target variable construction: Select the asset data of several credit cards that became non-performing assets between 2 and 7 years before the package closing date as sample data and input them into the static pool. Count the monthly recovery amount and cumulative recovery rate CRR of each asset in the static pool after it became a non-performing asset.

[0008] Step 2: Independent variable expansion: Get the expanded variables based on the asset data;

[0009] It should be noted that the asset data include: the total amount of recovery, the monthly recovery amount of non-performing assets, the current recovery amount of non-performing assets, the time when it becomes a non-performing asset, the annual income of the credit card holder, the outstanding principal and interest balance when it becomes a non-performing asset, the outstanding principal balance when it becomes a non-performing asset, the outstanding principal and interest balance when it becomes a non-performing asset, the personal information of the credit card holder and the time of each fund recovery and other related information; among them, the current recovery amount of non-performing assets is the amount returned within the fixed repayment period of the credit card, and the personal information of the credit card holder includes personal occupation, industry, region, monthly income and other information, and the personal information is obtained with the customer's permission.

[0010] Step 3: Select the independent variable in the asset data of the net pool and the extended variable, and select the cumulative recovery rate CRR as the target variable. Pre-process the static pool according to the independent variable, target variable and the asset data corresponding to the target variable;

[0011] It should be noted that the selection of independent variables is that the user selects the independent variables corresponding to the factors that may be related to the recycling behavior within the scope allowed by the current static pool template based on business experience; and the selection of target variables is based on the characteristics of the static pool recycling time distribution and the term of the ABS product, and the cumulative recovery rate CRR of the specified term is selected as the target variable;

[0012] Step 4: Analyze the pre-processed static pool data to obtain analysis results, and screen significant variables based on the analysis results;

[0013] Step 5: Group the significant variables to analyze the indicator range, and mark the variable set as the grouping dimension; group the discrete variables in the independent variables to obtain the analysis interval;

[0014] The above-mentioned index range for grouping and analyzing significant variables is adjusted by technicians in this field according to factors such as missing rate, data reliability, acquisition cost and business experience.

[0015] Step 6: Adjust the grouping dimension and grouping interval based on the analysis results, group the static pool according to the adjusted grouping dimension and grouping interval, and calculate the current average recovery rate and current cumulative recovery rate based on the adjusted grouping dimension and grouping method;

[0016] Step 7: Group the asset pools according to the static pool grouping method and then predict the recovery amount.

[0017] It should be noted that the asset pool is composed of data on a number of non-performing assets that require analysis and estimated recovery amounts.

[0018] Preferably, the cumulative recovery rate is calculated as: CRRt = ∑(RM1,…,RMt) / X; wherein X is the outstanding principal and interest balance when the asset becomes non-performing, RMi is the recovery amount of the non-performing asset in the i-th period, wherein i∈[1,t], and t is the total number of non-performing assets.

[0019] It should be noted that: since some samples become bad closer to the sample data extraction time, that is, the "data time" in the imported template, the available recovery data covers a small number of periods. In order to ensure that the data is meaningful, the CRR of the longer period is blanked. The specific way to blank CRR is:

[0020] set up M refers to the number of months between becoming a non-performing asset and data extraction, that is, there are M months to obtain real recovery data; among them, is the rounding-down symbol; if there is missing recovery amount in the recovery amount statistics of each period before M, the missing value will be treated as 0; if the number of periods is greater than M, the CRR of the corresponding period will be directly treated as a null value.

[0021] Preferably, the extended variables include: bad balance income ratio, time to become bad, outstanding principal ratio when becoming bad, number of days from the first time of becoming bad to the first time of swiping the card / withdrawing cash / installment payment;

[0022] Bad balance to income ratio = annual income of the credit card holder / outstanding principal and interest balance when it becomes bad;

[0023] It should be noted that: if the annual income of the non-performing asset or the outstanding principal and interest balance when it becomes non-performing is empty or 0, the calculated value is empty;

[0024] The time of becoming non-performing is the month when the assets generated by the corresponding credit card become non-performing assets;

[0025] The proportion of outstanding principal when the debt becomes NPL = the outstanding principal balance when the debt becomes NPL / the outstanding principal and interest balance when the debt becomes NPL;

[0026] It should be noted that: when the numerator is empty, the denominator is empty or 0, the calculated value is empty.

[0027] Preferably, the preprocessing of the static pool according to the independent variable, the target variable and the asset data corresponding to the target variable comprises the following steps:

[0028] S1: Eliminate the sample data with null cumulative recovery rate (CRR) of some non-performing assets in the static pool;

[0029] S2: Determine whether the sub-data missing rate of asset data in non-performing assets is greater than 30%; if yes, jump to S3; if no, remove the sample data corresponding to the non-performing assets;

[0030] S3: Determine whether the sub-data of the asset data in the non-performing assets in several consecutive repayment cycles is a discrete variable; if yes, fill the sub-data with a null value and jump to S5; if no, fill the discrete variable with the average value of the sub-data of this type in the asset data in the non-performing assets;

[0031] S4: Determine whether the value of the independent variable is single in several consecutive repayment cycles; if yes, remove the sample data corresponding to the independent variable; if no, keep the sample data corresponding to the independent variable in the fund pool;

[0032] S5: Convert the discrete variable into a numerical variable of the sample data set by sum coding; wherein the numerical variable is the difference between the mean of the target variable and the overall mean.

[0033] Preferably, the method for analyzing the preprocessed static pool data includes univariate analysis and multivariate analysis; the analysis result corresponding to the univariate analysis is the P value of each variable, and the analysis result corresponding to the multivariate analysis is the P value of the independent variable and the variance inflation factor of the independent variable; wherein, univariate analysis includes linear regression and nonlinear regression; multivariate analysis includes correlation coefficient and multiple regression analysis.

[0034] It should be noted that the smaller the P value, the higher the statistical significance of the independent variable, and the variance inflation factor is <10, so there is no problem of collinearity;

[0035] Preferably, screening significant variables according to the analysis results includes:

[0036] A1: Determine whether the P value is less than 0.05. If yes, jump to A2; if no, eliminate the independent variable corresponding to the P value.

[0037] A2: Determine whether the variance inflation factor corresponding to the independent variable is less than 10; if so, mark the independent variable corresponding to the P value as a significant variable; if not, eliminate the independent variable corresponding to the P value;

[0038] If not, the independent variable corresponding to the P value will be eliminated.

[0039] Preferably, the grouping analysis of discrete variables in the independent variables includes:

[0040] Perform hierarchical clustering from bottom to top according to the average distance between grouping dimensions, and output the clustering results; process the clustering results of each discrete variable to obtain the information value;

[0041] Through the regression decision tree, the continuous variables are grouped and analyzed one by one, and the maximum depth d is set. The selected grouping dimension is input to obtain the decision tree with the smallest mean square error MES in the regression result and output it; where 2^d = the number of proposed classifications; thus, the grouping is obtained when the recovery rate fitting effect is considered. Similarly, the grouping results of the decision tree are calculated to facilitate comparison with the variables involved in clustering;

[0042] The grouping interval is confirmed based on the decision tree with the smallest grouping dimension and mean square error (MES).

[0043] The larger the information value in the above, the better the grouping effect of the discrete variable is, and the stronger the ability of the grouping to distinguish the target variable is.

[0044] Preferably, the current average recovery rate in step 5 = the sum of the current recovery amounts of each asset in each group in the static pool at the same bad time / the total principal and interest of the bad assets in the group;

[0045] The current period cumulative recovery rate described in step 5 is the sum of several current period average recovery rates.

[0046] Preferably, the forecasting of the recovery amount after grouping the asset pool according to the static pool grouping method includes:

[0047] The asset pool is grouped according to the grouping dimension and grouping method of the static pool to obtain several groups of small assets. The cumulative recovery rate of each group in each period in the static pool is assigned to several groups of small assets. According to the asset pool packaging date, the distance from the packaging date to the bad account period of each non-performing asset on the packaging date is calculated, and the number of months of the bad account period on the packaging date is accumulated; among which, the bad account period on the packaging date = (packaging date-time of becoming bad) / 30;

[0048] Calculate the cumulative recovery rate of each non-performing asset at the beginning of each bad account period on the date of package closing and at the beginning of each subsequent month; the calculation method is: round down the number of bad periods at the beginning of the month, cumulative recovery rate = cumulative recovery rate corresponding to the integer month + recovery rate of the next month × the decimal part of the bad period after rounding off;

[0049] Calculate the monthly recovery rate and recovery amount of each non-performing asset on the date of sealing and in each subsequent month: the cumulative recovery rate at the beginning of the next month - the cumulative recovery rate at the beginning of the current month = the monthly recovery rate; the monthly recovery amount = the monthly recovery rate * the balance of non-performing principal, interest and fees;

[0050] Calculate the predicted recovery rate of a single non-performing asset: the total amount of recovery in the 36 months after the date of each non-performing asset packaging / the outstanding principal and interest on the date of each asset packaging × 100% = predicted recovery rate of a single asset.

[0051] Compared with the prior art, the beneficial effects of the present invention are: dividing the historical asset pool of the original credit card holder's assets into several homogeneous asset pools, and obtaining the average recovery rate of each sub-pool in each period in the asset pool after statistical division to obtain the basis for group mapping; in the case that the external environment does not change significantly, the recovery of the same type of assets belonging to the credit card holder should not change significantly, and it has greater accuracy. Therefore, by selecting the same group mapping method as the static pool when processing the asset pool, that is, corresponding each asset in the issued asset pool to the historical static pool grouping, obtaining the recovery situation of each non-performing asset, and summarizing to obtain the average collection rate of the asset pool, the problem of distribution assumptions can be avoided, and at the same time, it can also meet the requirement that ABS products only need to analyze and calculate the entire asset pool, thereby ensuring the interests of issuers and investors. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 It is a schematic diagram of the process of the present invention;

[0054] Figure 2 Outputting a heat map for correlation coefficient analysis in an embodiment of the present invention;

[0055] Figure 3 is a schematic diagram of corresponding clustering results in an embodiment of the present invention;

[0056] Figure 4 Schematic diagram of decision results of a regression decision tree in an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of the monthly recovery rate in an embodiment of the present invention;

[0058] Figure 6 Schematic diagram of cumulative recovery rate in an embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of the recovery ratio of the asset pool in each month in an embodiment of the present invention;

[0060] Figure 8 It is a schematic diagram of the non-performing asset pool in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] See also Figure 1-Figure 8 The first embodiment of the present invention provides a prediction of retail non-performing asset recovery, comprising the following steps:

[0063] Step 1: Target variable construction: Select the asset data of several credit cards that became non-performing assets between 2 and 7 years before the package closing date as sample data and input them into the static pool. Count the monthly recovery amount and cumulative recovery rate CRR of each asset in the static pool after it became a non-performing asset.

[0064] Step 2: Independent variable expansion: Get the expanded variables based on the asset data;

[0065] It should be noted that the asset data include: the total amount of recovery, the monthly recovery amount of non-performing assets, the current recovery amount of non-performing assets, the time when it becomes a non-performing asset, the annual income of the credit card holder, the outstanding principal and interest balance when it becomes a non-performing asset, the outstanding principal balance when it becomes a non-performing asset, the outstanding principal and interest balance when it becomes a non-performing asset, the personal information of the credit card holder and the time of each fund recovery, etc.; among them, the current recovery amount of non-performing assets is the amount returned within the fixed repayment period of the credit card, and the personal information of the credit card holder includes personal occupation, industry, region, monthly income, etc.

[0066] Step 3: Select the independent variable in the asset data of the static pool and the extended variable, and select the cumulative recovery rate CRR as the target variable, and pre-process the static pool according to the independent variable, target variable and the asset data corresponding to the target variable;

[0067] In this embodiment, the selection of independent variables is that the user selects independent variables corresponding to factors that may be related to the recovery behavior within the scope allowed by the current static pool template based on business experience; for example, the independent variables may include: occupation, industry, region, income, outstanding principal balance when becoming non-performing, non-performing balance income ratio and other variables; and the selection of target variables is based on the characteristics of the static pool recovery time distribution and the term of the product, and the cumulative recovery rate CRR of the specified term is selected as the target variable;

[0068] Target variable selection: Users can select the cumulative recovery rate CRR of a suitable period as the target variable based on the characteristics of the static pool recovery time distribution and the product life. For example, historical data experience shows that most of the recovery occurs within 36 months after becoming defective. At the same time, the product life is 2 years. Therefore, the cumulative recovery rate of 36 months can be selected as the target variable.

[0069] Step 4: Analyze the pre-processed static pool data to obtain analysis results, and screen significant variables based on the analysis results;

[0070] Step 5: Group the significant variables to analyze the indicator range, and mark the variable set as the grouping dimension; group the discrete variables in the independent variables to obtain the analysis interval;

[0071] It should be noted that the index range for grouping and analyzing significant variables in step 5 is adjusted by technicians in this field based on factors such as missing rate, data reliability, acquisition cost and business experience.

[0072] Step 6: Adjust the grouping dimension and grouping interval based on the analysis results, group the static pool according to the adjusted grouping dimension and grouping interval, and calculate the current average recovery rate and current cumulative recovery rate based on the adjusted grouping dimension and grouping method;

[0073] Step 7: Group the asset pools according to the static pool grouping method and then predict the recovery amount.

[0074] It should be noted that the asset pool is composed of data on a number of non-performing assets that require analysis and estimated recovery amounts.

[0075] The cumulative recovery rate in this embodiment is calculated as follows: CRRt = ∑(RM1,…,RMt) / X; wherein X is the outstanding principal and interest balance when the asset becomes non-performing, RMi is the recovery amount of the non-performing asset in the i-th period, wherein i∈[1,t], and t is the total number of non-performing assets.

[0076] It is worth noting that: since some samples become bad closer to the sample data extraction time, that is, the "data time" in the imported template, the available recovery data covers fewer periods. In order to ensure that the data is meaningful, the CRR of the longer period is blanked. The specific way to blank CRR is:

[0077] set up in, is the rounding-down symbol; if there is missing recovery amount in the recovery amount statistics of each period before M, the missing value will be treated as 0; if the number of periods is greater than M, the CRR of the corresponding period will be directly treated as a null value.

[0078] Since most of the fields in the static pool are absolute values ​​such as amount and date, in variable analysis, proportional values ​​or relative values ​​are often more significant. Therefore, some commonly used recovery prediction indicators are constructed. Therefore, in this embodiment, the independent variables are expanded. The specific expanded variables include: bad balance income ratio, bad time, outstanding principal ratio when bad, and the number of days from the first bad time to the first card swipe / cash withdrawal / installment payment;

[0079] Bad balance to income ratio = annual income of the credit card holder / outstanding principal and interest balance when it becomes bad;

[0080] If the annual income of the non-performing asset or the outstanding principal and interest balance when it becomes non-performing is empty or 0, the calculated value is empty;

[0081] The time of becoming non-performing is the month when the assets generated by the corresponding credit card become non-performing assets;

[0082] The proportion of outstanding principal when the debt becomes NPL = the outstanding principal balance when the debt becomes NPL / the outstanding principal and interest balance when the debt becomes NPL;

[0083] And when the numerator is empty, the denominator is empty or 0, the calculated value is empty.

[0084] To ensure that the results of the variable analysis are meaningful and can proceed smoothly, the static pool is preprocessed according to the independent variable, the target variable and the asset data corresponding to the target variable in this embodiment, including the following steps:

[0085] S1: Since some samples may have a short history of becoming bad and do not have sufficient recovery data, the recovery data for a longer period will be recorded as empty. Samples with empty target variables will be removed from the sample set. For example, if the cumulative recovery rate of 36 periods (CRR36) is selected as the target variable, samples with empty CRR36 will not be included in the analysis sample set.

[0086] S2: Determine whether the sub-data missing rate of asset data in non-performing assets is greater than 30%; if yes, jump to S3; if no, remove the sample data corresponding to the non-performing assets;

[0087] S3: Determine whether the sub-data of the asset data in the non-performing assets in several consecutive repayment cycles is a discrete variable; if yes, fill the sub-data with a null value and jump to S5; if no, fill the discrete variable with the average value of the sub-data of this type in the asset data in the non-performing assets;

[0088] S4: If the value of the independent variable is single, the result of the regression analysis is meaningless and the corresponding P value cannot be calculated. Therefore, it is necessary to determine whether the value of the independent variable is single in several consecutive repayment cycles; if yes, the sample data corresponding to the independent variable is eliminated; if no, the sample data corresponding to the independent variable is retained in the fund pool;

[0089] S5: Convert discrete variables into numerical variables by summing and coding, that is, convert the values ​​of each group of discrete variables into the difference between the mean of the target variable of the group and the overall mean, which is convenient for subsequent analysis. Although this method is prone to overfitting in predictive models. However, in the analysis process here, the purpose of modeling is not to predict, but to screen significant variables, and overfitting does not affect the results of variable significance analysis. In addition, this method does not increase the sample dimension and can avoid the problem of too sparse sample matrix.

[0090] Please refer to Table 1. For example, the mean value of the sample target variable with the city variable value "Shenzhen" is 11.23, and the mean value of all sample target variables is 11.36. Then the new value of the city variable "Shenzhen" is -0.13.

[0091]

[0092]

[0093] Table 1

[0094] In this embodiment, the method of analyzing the pre-processed static pool data includes univariate analysis and multivariate analysis; the analysis result corresponding to the univariate analysis is the P value of each variable, and the analysis result corresponding to the multivariate analysis is the P value of the independent variable and the variance inflation factor of the independent variable; wherein, the univariate analysis includes linear regression and nonlinear regression; the multivariate analysis includes correlation coefficient and multiple regression analysis.

[0095] After the discrete independent variables are coded, they enter the variable analysis process together with other selected continuous independent variables. According to the distribution form of the independent variables and the target variable, linear or nonlinear regression analysis is used to test the significant relationship between the independent variables and the target variable in the simple model. Variables with P<0.05 or relatively significant are selected to enter the multivariate analysis process. If there is no variable with P<0.05, the 5 variables with the smallest P value are selected to enter the multivariate analysis.

[0096] For example, the results of univariate analysis are shown in Table 2:

[0097] Variable number Variable Name P-value 1 City 0.9322 2 age 0.0232 3 Credit Line 0.0002 4 Bad principal and interest <0.0001

[0098] Table 2

[0099] According to the P value, the variables "age", "credit limit" and "bad debt principal and interest" are selected to enter the multivariate analysis process.

[0100] The smaller the P value, the higher the statistical significance of the independent variable, and the variance inflation factor is <10, so there is no problem of collinearity;

[0101] Multivariate analysis includes two methods: correlation coefficient and multiple regression analysis. The correlation coefficient detects the correlation between variables, and multiple regression analysis can test the multicollinearity between variables by calculating VIF. In the next step of group analysis, variables with inflation coefficient VIF>10 are excluded by default; for correlation coefficients greater than 0.7, variables with lower P values ​​are retained. For the output heat map of correlation coefficient analysis, please refer to Figure 2 shown.

[0102] The output of multiple regression analysis is shown in Table 3:

[0103] Variable number Variable Name P-value Variance Inflation Factor 1 age 0.0258 1 2 Credit Line 0.0003 1.7 3 Bad principal and interest <0.0001 1.7

[0104] Table 3

[0105] Among them, the correlation coefficients between age (age), credit limit (limit) and bad principal and interest (dfbase) are -0.015, -0.043 and 0.58, and the variance inflation factor is <10, so there is no collinearity problem.

[0106] And it is still significant in the multivariate model, and all three variables are retained by default in the system.

[0107] In this embodiment, screening significant variables according to the analysis results includes:

[0108] A1: Determine whether the P value is less than 0.05. If yes, jump to A2; if no, eliminate the independent variable corresponding to the P value.

[0109] A2: Determine whether the variance inflation factor corresponding to the independent variable is less than 10; if so, mark the independent variable corresponding to the P value as a significant variable; if not, eliminate the independent variable corresponding to the P value;

[0110] If not, the independent variable corresponding to the P value will be eliminated.

[0111] Preferably, the grouping analysis of discrete variables in the independent variables includes:

[0112] Perform hierarchical clustering from bottom to top according to the average distance between grouping dimensions, and output the clustering results; process the clustering results of each discrete variable to obtain the information value IV. The larger the information value, the better the grouping effect of the discrete variable, and the stronger the ability of the grouping to distinguish the target variable;

[0113] For example, cluster analysis is performed on "customer location city" and the corresponding clustering results are output as shown in Table 4 and Figure 3 As shown:

[0114]

[0115]

[0116] Table 4

[0117] Through the regression decision tree, the continuous variables are grouped and analyzed one by one, and the maximum depth d is set. The selected grouping dimension is input to obtain the decision tree with the smallest mean square error MES in the regression result and output it; where 2^d = the number of proposed classifications; thus, the grouping is obtained when the recovery rate fitting effect is considered. Similarly, the grouping results of the decision tree are calculated to facilitate comparison with the variables involved in clustering;

[0118] For example, input the credit limit variable, set the maximum depth to 2, and build a regression decision tree. The results are shown in Table 5 and Figure 4 As shown:

[0119]

[0120] Table 4

[0121] The grouping intervals are manually confirmed based on the grouping dimension and the decision tree with the smallest mean square error (MES), business experience, and actual conditions.

[0122] The current average recovery rate in this embodiment = the sum of the current recovery amount of each asset in each group in the static pool under the same bad time / the total principal and interest of the bad assets in this group;

[0123] The current period cumulative recovery rate in step 5 is the sum of several current period average recovery rates.

[0124] For example, assuming that the grouping dimension is determined to be "credit limit" based on the results of the static pool variable analysis, the grouping intervals are: 15,000 and below, 15,000 to 25,000, and above 25,000. By dividing the static pool assets into three groups, the results shown in Table 5 can be obtained:

[0125]

[0126] Table 5

[0127] In this embodiment, the asset pool is grouped according to the static pool grouping method and then the recovery amount is predicted, including:

[0128] The asset pool is grouped according to the grouping dimension and grouping method of the static pool to obtain several groups of small group assets. The cumulative recovery rate of each group in each period in the static pool is assigned to several groups of small group assets, that is, the cumulative recovery rate of the assets of the z-th group in the asset pool in the first i months after becoming non-performing is set as the static pool cumulative recovery rate CRR (Z, i), where z, i are both positive integers, and the estimated recovery rate for the current month = the cumulative recovery rate of this month - the cumulative recovery rate of last month;

[0129] According to the asset pool packaging date, the distance from the packaging date to the bad account period of each non-performing asset on the packaging date is calculated, and the number of months from the bad account period on the packaging date is accumulated. Specifically, the number of months + 1, + 2, until + 36, recorded as 37 bad account periods. In this embodiment, 24 months are taken as an example; where the bad account period on the packaging date = (packaging date - time to become bad) / 30;

[0130] Calculate the cumulative recovery rate of each non-performing asset at the beginning of each bad account period on the date of package closing and at the beginning of each subsequent month; the calculation method is: round down the number of bad periods at the beginning of the month, cumulative recovery rate = cumulative recovery rate corresponding to the integer month + recovery rate of the next month × the decimal part of the bad period after rounding off;

[0131] Calculate the monthly recovery rate and recovery amount of each NPL on the date of closing and in each subsequent month: cumulative recovery rate at the beginning of next month - cumulative recovery rate at the beginning of this month = monthly recovery rate; monthly recovery amount = monthly recovery rate * balance of NPL principal and interest fees; here refers to the NPL principal and interest fees when it becomes NPL, not the outstanding principal and interest fees at the time of closing.

[0132] Calculate the predicted recovery rate of a single non-performing asset: the total amount of recovery in the 36 months after the date of each non-performing asset packaging / the outstanding principal and interest on the date of each asset packaging × 100% = predicted recovery rate of a single asset.

[0133] For example, the asset pool is divided into groups according to the credit line of 15,000 or less, 15,000 to 25,000, and above 25,000, and after substituting the static pool cumulative recovery rate, the results are shown in Table 6:

[0134]

[0135] Table 6

[0136] Among them, the bad account period and cumulative recovery rate (%) of non-performing asset 688 on the date of package closing and at the beginning of each subsequent month are shown in Table 7:

[0137]

[0138] Table 7

[0139] Therefore, it can be calculated that the cumulative recovery rate (%) of the non-performing asset 688 at the beginning of the second month after the packaging date = CRR16 + 0.97 × (CRR17-CRR16) = 14.93 + 0.97 × (15.26-14.93) = 15.25. Among them, CRR16 is the cumulative recovery rate of the non-performing asset in the first 16 non-performing account periods, and CRR17 is the cumulative recovery rate of the first 17 non-performing account periods. The cumulative recovery rate (%) of the non-performing asset at the beginning of one month after the packaging date = 14.64 + 0.97 × (14.93-14.64) = 14.92. And so on, the cumulative recovery rate of the non-performing asset on the packaging date and at the beginning of each month thereafter is obtained.

[0140] Then, the current recovery rate is obtained by subtracting the cumulative recovery rate at the beginning of next month from the cumulative recovery rate at the beginning of this month. The recovery rate (%) of the month after the package date is (15.25-14.92) = 0.33. The balance of bad principal, interest and fees of non-performing asset 688 is 28949.09, and the monthly recovery amount of the non-performing asset 688 is 0.33 × 28949.09 / 100 = 95.532. And so on, the recovery amount of the non-performing asset in each month after the package date is obtained.

[0141] Finally, after obtaining the predicted value of the monthly recovery amount of each asset in the asset pool, add them up to get the monthly recovery amount of the entire asset pool and the final gross recovery rate. The gross recovery rate of the asset pool = the cumulative recovery amount within the calculation period (36 months) / the total amount of principal, interest and fees outstanding at the time of package closing. The recovery amount of each asset on the package closing date and in each subsequent month is added up to get the monthly recovery amount of each month. The monthly recovery amount of each month / the outstanding principal, interest and fees at the time of package closing is obtained to get the monthly recovery rate, such as Figure 5 shown.

[0142] The total amount of asset pool recovered in each month and before is calculated to get the cumulative amount of asset pool recovered in each month. The cumulative amount of recovery is divided by the outstanding principal and interest at the time of packaging to get the cumulative recovery rate, such as Figure 6 As shown;

[0143] The recovery amount of the asset pool in the current month is divided by the total recovery amount of 36 months to obtain the recovery ratio of the asset pool in each month, such as Figure 7 shown.

[0144] The predicted recovery rate of a single non-performing asset is divided into several intervals according to 2.5%. The total outstanding principal, interest and fees when the asset package is packaged with a single recovery rate falling in a specific interval is counted to obtain the proportion of outstanding principal, interest and fees in the recovery rate interval. Therefore, the structure of the asset pool can be understood from the perspective of recovery rate, such as Figure 8 shown.

[0145] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0146] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A forecast of retail non-performing asset recovery, characterized in that: The following steps are involved: S1: Select the asset data of several credit cards that became non-performing assets between 2 and 7 years before the package was sealed as the static pool of sample data, and count the monthly recovery amount and cumulative recovery rate CRR of each asset in the static pool after it became a non-performing asset; S2: Get extended variables based on asset data; S3: Select independent variables in the asset data of the static pool and the extended variables, and select the cumulative recovery rate CRR as the target variable, and pre-process the static pool according to the independent variables, target variables and the asset data corresponding to the target variables; S4: Analyze the pre-processed static pool data to obtain analysis results, and screen significant variables according to the analysis results; S5: Perform group analysis on the significant variables and mark the variable set as the grouping dimension; perform group analysis on the discrete variables in the independent variables and obtain the analysis interval; S6: adjusting the grouping dimension and grouping interval based on the analysis results, grouping the static pool according to the adjusted grouping dimension and grouping interval, and calculating the current average recovery rate and the current cumulative recovery rate according to the adjusted grouping dimension and grouping method; S7: Group the asset pools according to the static pool grouping method and make a forecast of the recovery amount.

2. A prediction of retail non-performing asset recovery according to claim 1, characterized in that: The cumulative recovery rate is calculated as follows: CRRt = ∑(RM1,…,RMt) / X; wherein X is the outstanding principal and interest balance when the asset becomes non-performing, RMi is the recovery amount of the non-performing asset in the i-th period, wherein i∈[1,t], and t is the total number of non-performing assets.

3. The prediction of retail non-performing asset recovery according to claim 1, characterized in that: The asset data include the total amount of recovery, the monthly recovery amount of non-performing assets, the current recovery amount of non-performing assets, the time when the asset becomes a non-performing asset, the annual income of the credit card holder, the outstanding principal and interest balance when the asset becomes a non-performing asset, the outstanding principal balance when the asset becomes a non-performing asset, the outstanding principal and interest balance when the asset becomes a non-performing asset, the personal information of the credit card holder and the time of each fund recovery; among them, the current recovery amount of non-performing assets is the amount returned within the fixed repayment period of the credit card, and the personal information of the credit card holder includes personal occupation, industry, region, and monthly income.

4. A prediction of retail non-performing asset recovery according to claim 3, characterized in that: The extended variables include: bad balance income ratio, time to become bad balance, outstanding principal ratio when becoming bad balance, number of days from the first bad balance to the first card swipe / cash withdrawal / installment payment; The bad balance income ratio = annual income of the credit card holder / outstanding principal and interest balance when the card becomes bad; The time of becoming non-performing is the month when the assets generated by the corresponding credit card become non-performing assets; The ratio of outstanding principal when becoming non-performing = the balance of outstanding principal when becoming non-performing / the balance of outstanding principal and interest when becoming non-performing.

5. The prediction of retail non-performing asset recovery according to claim 1, characterized in that: The method of preprocessing the static pool according to the independent variable, the target variable and the asset data corresponding to the target variable comprises the following steps: S1: Eliminate the sample data with null cumulative recovery rate (CRR) of some non-performing assets in the static pool; S2: Determine whether the sub-data missing rate of asset data in non-performing assets is greater than 30%; if yes, jump to S3; if no, remove the sample data corresponding to the non-performing assets; S3: Determine whether the sub-data of the asset data in the non-performing assets in several consecutive repayment cycles is a discrete variable; if yes, fill the sub-data with a null value and jump to S5; if no, fill the discrete variable with the average value of the sub-data of this type in the asset data in the non-performing assets; S4: Determine whether the value of the independent variable is single in several consecutive repayment cycles; if yes, remove the sample data corresponding to the independent variable; if no, keep the sample data corresponding to the independent variable in the fund pool; S5: Convert the discrete variable into a numerical variable of the sample data set by sum coding; wherein the numerical variable is the difference between the mean of the target variable and the overall mean.

6. A forecast of retail non-performing asset recovery according to claim 1, characterized in that: The method of analyzing the pre-processed static pool data includes univariate analysis and multivariate analysis; the analysis result corresponding to the univariate analysis is the P value of each variable, and the analysis result corresponding to the multivariate analysis is the P value of the independent variable and the variance inflation factor of the independent variable; wherein, the univariate analysis includes linear regression and nonlinear regression; the multivariate analysis includes correlation coefficient and multivariate regression analysis.

7. A forecast of retail non-performing asset recovery according to claim 6, characterized in that: The screening of significant variables according to the analysis results includes: A1: Determine whether the P value is less than 0.

05. If yes, jump to A2; if no, eliminate the independent variable corresponding to the P value. A2: Determine whether the variance inflation factor corresponding to the independent variable is less than 10; If yes, the independent variable corresponding to the P value is marked as a significant variable; if no, the independent variable corresponding to the P value is eliminated; If not, the independent variable corresponding to the P value will be eliminated.

8. The prediction of retail non-performing asset recovery according to claim 1, characterized in that: The grouping analysis of discrete variables in the independent variables includes: Perform hierarchical clustering from bottom to top according to the average distance between grouping dimensions, and output the clustering results; process the clustering results of each discrete variable to obtain the information value; The continuous variables are grouped and analyzed one by one through the regression decision tree, and the maximum depth d is set. The selected grouping dimension is input to obtain the decision tree with the smallest mean square error MES in the regression result and output it; where 2^d = the number of intended classifications; The grouping interval is confirmed based on the decision tree with the smallest grouping dimension and mean square error (MES).

9. The prediction of retail non-performing asset recovery according to claim 1, characterized in that: The current average recovery rate mentioned in step 5 = the sum of the current recovery amounts of each asset in each group in the static pool under the same bad period / the total principal and interest of the bad assets in this group; The current period cumulative recovery rate described in step 5 is the sum of several current period average recovery rates.

10. The prediction of retail bad asset recovery according to claim 1, characterized in that: The method of grouping the asset pool according to the static pool grouping method and then predicting the recovery amount includes: The asset pool is grouped according to the grouping dimension and grouping method of the static pool to obtain several groups of small assets. The cumulative recovery rate of each group in each period in the static pool is assigned to several groups of small assets. According to the asset pool packaging date, the distance from the packaging date to the bad account period of each non-performing asset on the packaging date is calculated, and the number of months of the bad account period on the packaging date is accumulated; among which, the bad account period on the packaging date = (packaging date-time of becoming bad) / 30; Calculate the cumulative recovery rate of each non-performing asset at the beginning of each bad account period on the date of package closing and at the beginning of each subsequent month; the calculation method is: round down the number of bad periods at the beginning of the month, cumulative recovery rate = cumulative recovery rate corresponding to the integer month + recovery rate of the next month × the decimal part of the bad period after rounding off; Calculate the monthly recovery rate and recovery amount of each non-performing asset on the date of sealing and in each subsequent month: the cumulative recovery rate at the beginning of the next month - the cumulative recovery rate at the beginning of the current month = the monthly recovery rate; the monthly recovery amount = the monthly recovery rate * the balance of non-performing principal, interest and fees; Calculate the predicted recovery rate of a single non-performing asset: the total amount of recovery in the 36 months after the date of each non-performing asset packaging / the outstanding principal and interest on the date of each asset packaging × 100% = predicted recovery rate of a single asset.

Images