CN110599336B

CN110599336B - Financial product purchase prediction method and system

Info

Publication number: CN110599336B
Application number: CN201810607787.6A
Authority: CN
Inventors: 路宏琦; 刘国华; 张帆; 刘军; 张宇
Original assignee: Beijing Zetyun Tech Co ltd
Current assignee: Beijing Zetyun Tech Co ltd
Priority date: 2018-06-13
Filing date: 2018-06-13
Publication date: 2020-12-15
Anticipated expiration: 2038-06-13
Also published as: CN110599336A

Abstract

The invention provides a financial product purchase prediction method and a system, wherein the prediction method comprises the following steps: acquiring customer data of a customer to be forecasted who does not hold a financial product at a first moment; inputting customer data of each customer to be predicted into a purchasing probability prediction model of the financial product, and predicting the probability of purchasing the financial product by each customer to be predicted at a second moment, wherein the second moment is later than the first moment; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers. In the invention, the purchasing probability prediction model of the financial product is obtained by training the real historical customer data of a large number of customers, so that the probability of purchasing the financial product by the customer to be predicted can be accurately predicted, and the financial product can be accurately recommended to the customer.

Description

Financial product purchase prediction method and system

Technical Field

The invention relates to the technical field of data processing, in particular to a financial product purchase prediction method and system.

Background

With the rapid growth of the number and variety of bank customers and bank financial products, how to effectively recommend financial products to potential customers is one of the main targets for improving marketing effect. The traditional financial product recommendation method cannot accurately predict the probability of purchasing financial products by customers, so that the recommended customers cannot be accurately selected, and the recommendation success rate is not high.

Disclosure of Invention

In view of the above, the present invention provides a method and a system for predicting the purchase of a financial product, which can accurately predict the probability of a customer purchasing the financial product.

To solve the above technical problem, the present invention provides a financial product purchase prediction method, including:

acquiring customer data of a customer to be forecasted who does not hold a financial product at a first moment;

inputting customer data of each customer to be predicted into a purchasing probability prediction model of the financial product, and predicting the probability of purchasing the financial product by each customer to be predicted at a second moment, wherein the second moment is later than the first moment; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers.

Preferably, the customer data of the customer to be predicted is customer data between the first time and a third time, and the third time is earlier than the first time.

Preferably, each financial product corresponds to a purchasing probability prediction model; or

Each financial product corresponds to 12 purchasing probability prediction models, and the 12 purchasing probability prediction models respectively correspond to 1-12 months.

Preferably, when there are 12 purchasing probability prediction models corresponding to each financial product, the step of inputting the customer data of each customer to be predicted into the purchasing probability prediction models of the financial products comprises:

determining the month in which the second moment is located;

and selecting a purchase probability prediction model corresponding to the month in which the second moment is positioned, and predicting the probability of purchasing the financial product by each customer to be predicted at the second moment.

Preferably, before the step of predicting the probability of purchasing the financial product at the second time by each customer to be predicted, the method further comprises:

acquiring a training sample set, wherein the training sample set is a set of historical customer data of a plurality of training customers before a fourth moment;

obtaining N algorithm models, wherein N is a positive integer greater than or equal to 1;

setting parameters of each algorithm model, inputting the training sample set into each algorithm model, and predicting the probability that the customer for training purchases the financial product at a fifth moment, wherein the fifth moment is later than the fourth moment;

adjusting parameters of corresponding algorithm models according to the predicted probability that the customer for training purchases the financial product at the fifth moment to obtain N trained algorithm models;

and determining a purchase probability prediction model of the financial product according to the trained N algorithm models.

Preferably, the step of adjusting parameters of the corresponding algorithm models according to the predicted probability of purchasing the financial product by the training customer at the fifth moment to obtain N trained algorithm models includes:

adjusting parameters of corresponding algorithm models according to the predicted probability of the training customer purchasing the financial product at the fifth moment and preset evaluation indexes to obtain N adjusted algorithm models;

inputting the training sample set into the adjusted algorithm model, and predicting the probability that the training customers in the training sample set purchase the financial products at the fifth moment; determining a first evaluation index value of each algorithm model according to the predicted probability of the training customer in the training sample set purchasing the financial product at the fifth moment and a preset evaluation index;

obtaining a test sample set, wherein the test sample set is a set of historical customer data of a plurality of training customers before a fourth time, and the test sample set is different from the training sample set; inputting the test sample set into the correspondingly adjusted algorithm model, and predicting the probability that the training customer in the test sample set purchases the financial product at the fifth moment; determining a second evaluation index value of each algorithm model according to the predicted probability of the training customer in the test sample set purchasing the financial product at the fifth moment and a preset evaluation index;

comparing the first evaluation index value with the second evaluation index value to obtain the fitting degree of each algorithm model; and adjusting parameters of the corresponding algorithm models according to the fitting degree of each algorithm model to obtain the trained algorithm models.

Preferably, the preset evaluation index includes at least one of prediction accuracy, prediction precision, recall rate, AUC score and F1 score.

Preferably, the plurality of training clients includes: a first customer who already holds the financial product at the fourth time, and a second customer who does not hold the financial product at the fourth time.

Preferably, each financial product corresponds to a purchase probability prediction model, the historical customer data of the first customer includes customer data between a sixth time and a seventh time, the sixth time is a time point when the first customer purchases the financial product, and the seventh time is earlier than the sixth time.

Preferably, a time period between the sixth timing and the seventh timing is equal to a time period between the fifth timing and the fourth timing.

Preferably, each financial product corresponds to a purchase probability prediction model, and the historical customer data of the second customer includes customer data of any one month in a data extraction period before the fourth time.

Preferably, each of the financial products corresponds to 12 purchase probability prediction models, and for a purchase probability prediction model for a given month, the historical customer data of the first customer and the second customer includes customer data for a month before the given month in M consecutive years before the year at the fourth time, where M is a positive integer greater than or equal to 2.

Preferably, the N algorithm models are selected from the following algorithm models: the system comprises a gradient lifting algorithm model, a neural network algorithm model, a polynomial Bayesian algorithm model, a random forest algorithm model and a support vector machine algorithm model.

Preferably, N is a positive integer greater than or equal to 2, and the step of determining a purchase probability prediction model of the financial product according to the trained N algorithm models includes:

acquiring the probability of purchasing the financial product at the fifth moment of the training customer predicted by the trained N algorithm models;

inputting the probability of purchasing the financial product at the fifth moment of the training client predicted by the trained N algorithm models into a voting model to obtain the probability of purchasing the financial product at the fifth moment of the training client predicted by the purchasing probability prediction model of the financial product.

Preferably, the voting model is a majority voting model or a weighted voting model.

Preferably, the weight of the algorithm model is determined based on at least one of a prediction accuracy, a recall, an AUC score and an F1 score of the algorithm model.

Preferably, the weight of each of the algorithm models is the same.

Preferably, the step of obtaining a training sample set comprises:

and carrying out data preprocessing on the training sample set to be processed to obtain the training sample set after data preprocessing.

Preferably, the data preprocessing includes at least one of missing value calculation, outlier rejection, data transformation, non-dimensionalization, and normalization.

Preferably, the step of obtaining a training sample set comprises:

and processing the customer characteristics in the training sample set to be processed by adopting a characteristic selection model, determining the selected customer characteristics, and screening the training sample set by adopting the selected customer characteristics.

Preferably, the step of processing the client features in the training sample set to be processed by using the feature selection model, and determining the selected client features includes:

screening the client characteristics in the training sample set to be processed by adopting at least two screening methods of mutual information, chi-square test and F verification;

aggregating the client characteristics in the training sample set to be processed screened by adopting different screening methods to obtain the aggregated client characteristics;

and screening the client characteristics after the polymerization treatment by adopting at least one of a recursive characteristic elimination method and a characteristic model elimination method to obtain the selected client characteristics.

Preferably, the step of determining the selected customer characteristic further comprises:

and performing dimension reduction processing on the client characteristics by adopting at least one of characteristic orthogonalization, characteristic principal component analysis and matrix decomposition.

Preferably, after the step of determining the purchasing probability prediction model of the financial product, the method further includes:

calculating importance information of the selected customer characteristics through a purchase probability prediction model of the financial product;

and adjusting the adopted client characteristics according to the importance information of the client characteristics.

Preferably, after the step of predicting the probability that the customer to be predicted will purchase the financial product at the second time, the method further comprises:

aiming at a financial product, sorting the predicted probability of purchasing the financial product of a plurality of clients to be predicted from high to low, selecting the first Y clients to be predicted, and recommending the financial product to the first Y clients to be predicted, wherein Y is a positive integer greater than or equal to 1; and/or

And aiming at one customer to be predicted, sorting the probability of purchasing a plurality of financial products at the second moment of the customer to be predicted from high to low, selecting the first Z financial products, and recommending the first Z financial products to the customer to be predicted, wherein Z is a positive integer greater than or equal to 1.

The present invention also provides a financial product purchase prediction system, comprising:

the system comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring customer data of a customer to be predicted who does not hold a financial product at a first moment;

the forecasting module is used for inputting the customer data of each customer to be forecasted into the purchasing probability forecasting model of the financial product and forecasting the probability that each customer to be forecasted purchases the financial product at a second moment, wherein the second moment is later than the first moment; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers.

Preferably, when each financial product corresponds to 12 purchasing probability prediction models, the prediction module is configured to determine the month in which the second time is located; and selecting a purchase probability prediction model corresponding to the month in which the second moment is positioned, and predicting the probability of purchasing the financial product by each customer to be predicted at the second moment.

Preferably, the financial product purchase prediction system further comprises:

a second obtaining module, configured to obtain a training sample set, where the training sample set is a set of historical customer data of a plurality of training customers before a fourth time;

the third acquisition module is used for acquiring N algorithm models, wherein N is a positive integer greater than or equal to 1;

the training module is used for setting parameters of each algorithm model, inputting the training sample set into each algorithm model and predicting the probability that the customer for training purchases the financial product at a fifth moment, wherein the fifth moment is later than the fourth moment;

the adjusting module is used for adjusting parameters of corresponding algorithm models according to the predicted probability that the training customer purchases the financial product at the fifth moment to obtain N trained algorithm models;

and the determining module is used for determining a purchase probability prediction model of the financial product according to the trained N algorithm models.

Preferably, the adjusting module is configured to adjust parameters of corresponding algorithm models according to the predicted probability that the training customer purchases the financial product at the fifth moment and a preset evaluation index, so as to obtain N adjusted algorithm models; inputting the training sample set into the adjusted algorithm model, and predicting the probability that the training customers in the training sample set purchase the financial products at the fifth moment; determining a first evaluation index value of each algorithm model according to the predicted probability of the training customer in the training sample set purchasing the financial product at the fifth moment and a preset evaluation index; obtaining a test sample set, wherein the test sample set is a set of historical customer data of a plurality of training customers before a fourth time, and the test sample set is different from the training sample set; inputting the test sample set into the correspondingly adjusted algorithm model, and predicting the probability that the training customer in the test sample set purchases the financial product at the fifth moment; determining a second evaluation index value of each algorithm model according to the predicted probability of the training customer in the test sample set purchasing the financial product at the fifth moment and a preset evaluation index; comparing the first evaluation index value with the second evaluation index value to obtain the fitting degree of each algorithm model; and adjusting parameters of the corresponding algorithm models according to the fitting degree of each algorithm model to obtain the trained algorithm models.

Preferably, N is a positive integer greater than or equal to 2, and the determining module is configured to obtain a probability that the trained customers predicted by the trained N algorithm models purchase the financial product at the fifth time; inputting the probability of purchasing the financial product at the fifth moment of the training client predicted by the trained N algorithm models into a voting model to obtain the probability of purchasing the financial product at the fifth moment of the training client predicted by the purchasing probability prediction model of the financial product.

Preferably, the weight of each of the algorithm models is the same.

Preferably, the second obtaining module is configured to perform data preprocessing on the training sample set to be processed, so as to obtain the training sample set after data preprocessing.

Preferably, the second obtaining module is configured to process the client features in the training sample set to be processed by using a feature selection model, determine selected client features, and filter the training sample set by using the selected client features.

Preferably, the second obtaining module is configured to filter the client features in the training sample set to be processed by using at least two filtering methods of mutual information, chi-square test, and F-test; aggregating the client characteristics in the training sample set to be processed screened by adopting different screening methods to obtain the aggregated client characteristics; and screening the client characteristics after the polymerization treatment by adopting at least one of a recursive characteristic elimination method and a characteristic model elimination method to obtain the selected client characteristics.

Preferably, the second obtaining module is configured to perform dimension reduction processing on the client feature by using at least one of feature orthogonalization, principal component analysis of the feature, and matrix decomposition.

Preferably, the second obtaining module is configured to calculate importance information of the selected customer characteristic through a purchase probability prediction model of the financial product; and adjusting the adopted client characteristics according to the importance information of the client characteristics.

Preferably, the financial product purchase prediction system further comprises:

the first recommending module is used for sorting the predicted probability of purchasing the financial products of the multiple to-be-predicted customers from high to low aiming at a financial product, selecting the front Y to-be-predicted customers and recommending the financial product to the front Y to-be-predicted customers, wherein Y is a positive integer greater than or equal to 1; and/or

And the second recommending module is used for sorting the probability that the customer to be predicted purchases a plurality of financial products at the second moment according to the sequence from high to low aiming at the customer to be predicted, selecting the first Z financial products and recommending the first Z financial products to the customer to be predicted, wherein Z is a positive integer greater than or equal to 1.

The invention also provides a financial product purchase prediction system comprising a memory, a processor and a computer program stored on the memory and executable on the processor; the processor implements the above-described financial product purchase prediction method when executing the program.

The present invention also provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the above-described financial product purchase prediction method.

The technical scheme of the invention has the following beneficial effects:

in the embodiment of the invention, the purchasing probability prediction model of the financial products is obtained by training the real historical customer data of a large number of customers, so that the probability of purchasing the financial products by the customer to be predicted can be accurately predicted, and the financial products can be accurately recommended to the customer.

Drawings

FIG. 1 is a flow chart illustrating a method for predicting a purchase of a financial product according to an embodiment of the present invention;

FIG. 2 is a diagram illustrating various time points during prediction according to an embodiment of the present invention;

FIG. 3 is a flowchart illustrating a method for training a purchase probability prediction model of a financial product according to an embodiment of the invention;

FIG. 4 is a diagram illustrating time points during model training according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a financial product purchase forecasting system according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of a financial product purchase forecasting system according to another embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the drawings of the embodiments of the present invention. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the described embodiments of the invention, are within the scope of the invention.

Referring to fig. 1, fig. 1 is a schematic flow chart illustrating a financial product purchase forecasting method according to an embodiment of the present invention, the method including:

step 11: acquiring customer data of a customer to be forecasted who does not hold a financial product at a first moment;

step 12: inputting customer data of each customer to be predicted into a purchasing probability prediction model of the financial product, and predicting the probability of purchasing the financial product by each customer to be predicted at a second moment, wherein the second moment is later than the first moment; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers.

In an embodiment of the present invention, the data output by the purchasing probability prediction model of the financial product may be a specific probability value, for example, the probability that a customer to be predicted purchases the financial product at the second time is 0.38, or may be an identifier indicating whether the customer to be predicted will purchase the financial product, for example, the identifier of the output customer to be predicted is 1 indicating that the customer to be predicted will purchase the financial product, and the identifier of the output customer to be predicted is 0 indicating that the customer to be predicted will not purchase the financial product.

In the embodiment of the present invention, the first time is a reference time, and may be a current time or another past time. The second time is a target time, namely the purchasing probability from the first time to the second time is predicted.

In the embodiment of the present invention, referring to fig. 2, the customer data of the customer to be predicted input to the purchase probability prediction model is the customer data between the first time Tt and the third time Tp. Preferably, the duration between the third time Tp and the first time Tt is a shortest period Tmin for obtaining customer data, which may be, for example, one hour, one day, one month, or three months, etc. The present invention is explained by taking the time period between the third time Tp and the first time Tt as an example of one month. For example, when predicting the probability of purchasing a financial product in 2018, month 1, customer data of a customer to be predicted in 2017, month 12 may be obtained. In fig. 2, Tf' is the second time. The third time is a past historical time with the current time as a reference.

In the embodiment of the invention, the financial products can be, for example, financial products such as a large deposit receipt, a periodic deposit in one year or more, and the like.

In the embodiment of the invention, the holding financial products: the financial product is held in the current month when Tp is the reference time Tt (for example, Tmin is one month), Tt is the current time, and the held financial product is held in the current month, including that the history is not held and is held in the current month, and the history is held in the current month; not holding financial products: the term "not held" means that the financial product is not held in the time period from Tp to Tt (e.g. Tmin), for example, Tmin is one month, Tt is the current time, and the non-held financial product is not held in the current month, including that the history is not held and not held in the current month, the history is once held and not held in the current month.

The customer data of the customer to be predicted input to the purchase probability prediction model of the financial product may be directly acquired customer data, or may be customer data obtained by performing data preprocessing on the acquired customer data, where the data preprocessing may include at least one of missing value calculation, outlier exclusion, data transformation, dimensionless processing, and normalization processing. The dimensionless method includes normalization, interval scaling, etc., the missing value calculation includes missing value filling, etc., and the data transformation includes polynomial data conversion, etc. For example, customer data is deleted or filled in, where the filling may be to complement a default value such as zero or an average value.

In some embodiments of the invention, each of the financial products may correspond to a purchase probability prediction model.

In some other embodiments of the present invention, each of the financial products may correspond to 12 purchasing probability prediction models, and the 12 purchasing probability prediction models correspond to 1-12 months, respectively. In this way, the prediction result is more accurate because each financial product corresponds to one purchasing probability prediction model every month.

In an embodiment of the present invention, when each financial product corresponds to 12 purchasing probability prediction models, the step 12 is to input customer data of each customer to be predicted into the purchasing probability prediction models of the financial products, and the step of predicting the probability that each customer to be predicted purchases the financial product at the second time includes:

step 121: determining the month in which the second moment is located;

step 122: and selecting a purchase probability prediction model corresponding to the month in which the second moment is positioned, and predicting the probability of purchasing the financial product by each customer to be predicted at the second moment.

(1) Model training

In an embodiment of the present invention, before the step 12 (inputting the customer data of each customer to be predicted into the purchase probability prediction model of the financial product, and predicting the probability that each customer to be predicted purchases the financial product at the second time), the method further includes: the purchase probability prediction model of the financial product is trained, and a method of training the purchase probability prediction model of the financial product is explained below.

In the embodiment of the invention, the purchase probability prediction model of each financial product is obtained by independently training.

Referring to fig. 3, fig. 3 is a flowchart illustrating a method for training a purchase probability prediction model of a financial product according to an embodiment of the present invention, the method including:

step 31: acquiring a training sample set, wherein the training sample set is a set of historical customer data of a plurality of training customers before a fourth moment; the fourth time is earlier than the first time; the fourth time is a past historical time with the current time as a reference.

Step 32: obtaining N algorithm models, wherein N is a positive integer greater than or equal to 1;

step 33: setting parameters of each algorithm model, inputting the training sample set into each algorithm model, and predicting the probability that the customer for training purchases the financial product at a fifth moment, wherein the fifth moment is later than the fourth moment;

step 34: adjusting parameters of corresponding algorithm models according to the predicted probability that the customer for training purchases the financial product at the fifth moment to obtain N trained algorithm models;

step 35: and determining a purchase probability prediction model of the financial product according to the trained N algorithm models.

In the embodiment of the invention, the purchasing probability prediction model of the financial products is obtained by training the real historical customer data of a large number of customers, so that the probability of purchasing the financial products by the customer to be predicted can be accurately predicted. In addition, the purchase probability prediction model is obtained according to the N algorithm models, and when N is larger than 1, the performance of the purchase probability prediction model can be improved, and the prediction accuracy rate is improved.

In an embodiment of the present invention, the step of adjusting parameters of corresponding algorithm models according to the predicted probability that the training customer purchases the financial product at the fifth time to obtain N trained algorithm models may include:

obtaining a test sample set, wherein the test sample set is a set of historical customer data of a plurality of training customers before a fourth time, the test sample set is different from the training sample set, the test sample set can be obtained by randomly splitting customer data of the training customers into the training sample set and the test sample set, and further the data splitting is performed before training an algorithm model; inputting the test sample set into the correspondingly adjusted algorithm model, and predicting the probability that the training customer in the test sample set purchases the financial product at the fifth moment; determining a second evaluation index value of each algorithm model according to the predicted probability of the training customer in the test sample set purchasing the financial product at the fifth moment and a preset evaluation index;

In some preferred embodiments of the present invention, the preset evaluation index includes at least one of prediction accuracy, prediction precision, recall, area under the curve (AUC) score and F1 score.

In the embodiment of the invention, the method for determining the trained algorithm model comprises two optimization steps, wherein one step is to optimize the algorithm model according to the preset evaluation value in the training process, and the other step is to further optimize the algorithm model according to the comparison with the prediction result of the test sample set after the training is finished, so that the performance of the algorithm model at the training position is better.

In an embodiment of the present invention, the training client includes: a first customer who already holds the financial product at the fourth time, and a second customer who does not hold the financial product at the fourth time.

In this embodiment of the present invention, the fourth time T0 is the end time of the data extraction period Te. For example, the historical customer data of the extracted training customers may be data in a period from 2016, 1/2017, 12/31/2017, and the fourth time is 2017, 12/31/2017.

1) Training model for whole year

In some embodiments of the invention, each financial product corresponds to one purchase probability prediction model, the financial products include a plurality of products, each product is predicted separately, and a corresponding purchase rate probability model is trained for each product. The purchase probability prediction model of the present embodiment is a model for all years, and in this case, the historical customer data of the training customer may be: referring to fig. 4, the historical customer data of the first customer (the customer who has purchased the financial product at time T0) may include customer data between a sixth time Tb and a seventh time T-2, the sixth time Tb being a time point when the first customer purchases the financial product, and the customer data may further include an identifier 1 that the first customer holds the financial product. The seventh time T-2 is earlier than the sixth time Tb. Because the first customer data is not yet available to purchase the financial product, it is only available as training data. Preferably, the time duration between the sixth time Tb and the seventh time T-2 is equal to the time duration between the fifth time Tf and the fourth time T0. It will of course be appreciated that the sixth time Tb and the seventh time T-2 of different first clients will generally be different from each other. In addition, if the first customer purchases the financial product a plurality of times within the data extraction period Te, the time of first purchasing the financial product may be selected as the sixth time Tb. Of course, the time of purchasing the financial product at any time may be selected as the sixth time Tb.

In an embodiment of the present invention, the historical customer data of the second customer (the customer not holding the financial product at time T0) may include customer data of any one month in a data extraction period Te (for example, 24 months) before the fourth time T0, and at the same time, the customer data may further include an identifier 0 that the second customer does not hold the financial product. Wherein, the months corresponding to different customers are the same or different. Because the number of clients is generally a large magnitude, large data is randomly extracted, and data of each month in a year can be extracted, so that the extracted data can cover each month, and the accuracy of the trained model is ensured.

With respect to such a model, in the prediction, customer data of customers to be predicted who do not hold a financial product at the first time Tt may be extracted as customer data input to the purchase probability prediction model, and the probability that each of the customers to be predicted purchases the financial product at the second time Tf' is predicted, for example, the customer data includes customer data between the third time Tp and the first time Tt (for example, the time duration is one month). In the embodiment of the present invention, it is preferable that the time length from the second time Tf' to the first time Tt used in the prediction is equal to the time length from the fifth time Tf to the fourth time T0 used in the training.

2) Training month model

In some embodiments of the present invention, there are 12 purchase probability prediction models for each financial product, i.e., one purchase probability prediction model for each month. At this time, for the purchase probability prediction model for the specified month, the historical customer data of the first customer and the second customer includes customer data of a month before the specified month in M consecutive years before the year at the fourth time, where M is a positive integer greater than or equal to 2. The customer data may include target variables and characteristic variables for training a purchase probability prediction model, the target variables being 1 (held or purchased) or 0 (not held or purchased), the target variables being taken from a specified month, the target variables being 1, i.e. the customer holds or purchases a financial product in the specified month, and the target variables being 0, i.e. the customer does not hold or purchase a financial product in the specified month; the characteristic variable is taken from a month before the specified month, and may specifically include customer attribute information, customer transaction information, and the like, and in other embodiments, the medium-term information or the long-term information, such as quarterly information, semiannuity information, and the like, may also be calculated by the characteristic variable one month before the specified month.

For example, for a first customer (i.e., a customer identified as 1 who has purchased the financial product at time T0), take 1 month of customer data for 10 consecutive years per year to obtain a target variable, and 12 months of customer data for 10 consecutive years per year to obtain a characteristic variable; for a second customer (i.e., a customer who did not purchase the financial product at time T0, identified as 0), take 1 month of customer data per year for 10 consecutive years to obtain a target variable, and 12 months of customer data per year for 10 consecutive years to obtain a characteristic variable; training results in a predictive model of the probability of purchase for the next 1 month of the year for a financial product. For months 2-12, similarly, a predictive model of the probability of purchase for months 2-12 of a financial product may be trained. The customer data may include monthly information and characteristic variables, the characteristic variables include customer attribute information, customer transaction behavior information, and the like, and in other embodiments, medium-term and long-term information, such as quarterly information, semiannuity information, and the like, may also be obtained through the monthly information calculation.

For such a model, in the prediction, customer data of customers to be predicted who do not hold financial products at the first time Tt may be extracted as customer data input of the purchase probability prediction model, and the probability of purchasing the financial products by each customer to be predicted at the second time Tf' is predicted, where the customer data may include customer data between the third time Tp and the first time Tt (for example, the duration is one month), for example, the customer data of the customer to be predicted who do not hold financial products at the month 2017 is taken when the purchase probability of the month 1 of 2018 is predicted, the customer data of the customer to be predicted who do not hold financial products at the month 12 of 2017 is extracted, and the model of the month 1 is input to predict the probability of purchasing financial products by the customer to be predicted at the month 1 of 2018.

In the embodiment of the present invention, preferably, the N algorithm models are selected from the following algorithm models: the system comprises a gradient lifting algorithm model, a neural network algorithm model, a polynomial Bayesian algorithm model, a random forest algorithm model and a support vector machine algorithm model. The algorithm model is selected based on the characteristics of nonlinearity, occupied computing resources and the like.

The following briefly introduces each algorithm model.

gradient-Boosting trees are members of the Boosting family of ensemble learning, but are very different from traditional Adaboost. The gradient lifting tree is also iteration, a forward distribution algorithm is used, in the iteration of the gradient lifting tree, the strong learner obtained in the previous iteration is ft-1(x), the loss function is L (y, ft-1(x)), the objective of the iteration of the current round is to find a weak learner ht (x) of the CART regression tree model, and the loss L (y, ft (x)) of the current round is enabled to be the minimum (y, ft-1(x) + ht (x)). That is, the iteration of the present round finds the decision tree, so that the loss of the sample is as small as possible.

A multi-layer perceptual classifier (mlpclasifier), one of the neural network algorithms, implements a multi-layer perceptron (MLP) algorithm trained using Backpropagation and supports a cross-entropy loss function, which can be probability estimated by running the predict _ proba method. Meanwhile, back-propagated MLP training is used. More precisely, it uses a specified form of gradient descent training and computes the gradient using back propagation. For classification, it minimizes the cross entropy loss function, giving a vector of probability estimates for each sample. Mlpclasifier supports multi-class classification by applying Softmax as an output function. Furthermore, the model supports multi-label classification, where samples can belong to multiple classes.

Polynomial naive bayes (MultinomialNB) implements a naive bayes algorithm that obeys polynomial distribution data, also one of the two classical naive bayes algorithms for text classification (data in this field is often represented in word vectors, although in practice tf-idf vectors perform well when predicted). Distribution parameter is represented by theta of each type y_y＝(θ_y1，...，θ_yn) Vector determination, where n is the number of features (for text classification, the size of the vocabulary), and θ_yiIs the probability P (x) of the feature i in the sample belonging to class y_i|y)。

The random forest is an integrated algorithm (Ensemble Learning), which belongs to Bagging type, and the final result is voted or averaged by combining a plurality of weak classifiers, so that the result of the whole model has higher accuracy and generalization performance. It can achieve good results, mainly attributed to "randomness" and "forest", one making it resistant to overfitting and the other making it more accurate. Random forests use CART decision trees as weak learners. When each tree is generated, the features selected by each tree are only a few features selected randomly, and the derivation of the total number m of the features is generally obtained by default, so that the feature randomness is ensured.

Although Support Vector Machines (SVMs) are short for more than twenty years, their own birth makes the field of Machine learning based on their good classification performance. If the ensemble learning algorithm is not considered, and a specific training data set is not considered, the performance SVM in the classification algorithm is excellent. The SVM is a binary classification algorithm, and both linear classification and nonlinear classification are supported. Through evolution, multivariate classification can be supported at present, and meanwhile, through expansion, the multivariate classification can also be applied to regression problems.

In the embodiment of the invention, when N is a positive integer greater than or equal to 2, the selected N algorithm models are respectively trained, wherein the training comprises the steps of adjusting each model to an optimal model by using a grid search cross-validation method, and then placing the N models in voting models for voting to obtain the final purchase probability.

That is, the step 36 of determining the purchasing probability prediction model of the financial product according to the trained N algorithm models includes:

step 361: acquiring the probability of purchasing the financial product at the fifth moment of the training customer predicted by the trained N algorithm models;

step 362: inputting the probability of purchasing the financial product at the fifth moment of the training client predicted by the trained N algorithm models into a Voting model (Voting Classifier), and obtaining the probability of purchasing the financial product at the fifth moment of the training client predicted by the purchasing probability prediction model of the financial product.

In the embodiment of the invention, N algorithm models are aggregated, N models are used simultaneously, and the prediction results of N models are organically combined, so that the prediction accuracy can be effectively improved.

In the embodiment of the present invention, preferably, the voting model is a majority voting model (hard voting), or a weighted voting model (soft voting).

The majority voting model is a model that uses the results of most of the algorithm models (e.g., 0 or 1) as the final prediction results, and the majority may be more than half of the algorithm models, for example.

The weighted voting model is that for the N algorithm models, each algorithm model is given a weight, the voting model multiplies the probability predicted by each algorithm model by the corresponding weight, and the multiplication results of all the algorithm models are summed to obtain the probability of the training client for purchasing the financial product.

In the embodiment of the invention, the weight of each algorithm model can be different, and the weight of the algorithm model is determined based on at least one of the prediction accuracy, the recall rate, the AUC score and the F1 score of the algorithm model. Generally, the higher the score of the index, the higher the weight.

In some embodiments of the invention, the weights of each of the algorithm models may be the same.

In the embodiment of the present invention, preferably, the voting model adopts a soft voting manner, that is, probabilities calculated by N algorithm models are subjected to weighted summation to obtain a weighted summation probability predicted to be 0 (not to be purchased) and a weighted summation probability predicted to be 1 (to be purchased), respectively, and the default weight is equal weight (each weight of N models is 1/N). The class with the high final selection probability is a predicted value, for example, 1 with the high probability is 1. The weights of multiple algorithm models may also be determined based on the accuracy with which each algorithm model predicts the results in combination with the weights. For example, the weight of each algorithm model is determined based on the accuracy rate, the higher the accuracy rate, the greater the weight. For example, the accuracy of the algorithm model 1 is 50%, the accuracy of the algorithm model 2 is 60%, the accuracy of the algorithm model 3 is 70%, the accuracy of the algorithm model 4 is 70%, and the accuracy of the algorithm model 5 is 80%, then the weight of the algorithm model 1 is 5/33, the weight of the algorithm model 2 is 6/33, the weight of the algorithm model 3 is 7/33, the weight of the algorithm model 4 is 7/33, and the weight of the algorithm model 5 is 8/33.

In the embodiment of the invention, the prediction is completed by adopting a voting model. Where the voting model combines N algorithm models (machine learning classifiers) that are conceptually different and uses a majority voting model (hard voting) or a weighted voting model (soft voting) to predict the class label (0 or 1). Such a classifier can be used for a set of well-behaved models in order to balance the weaknesses of its individuals.

In the above embodiment of the present invention, the fitness of the algorithm model may be determined by using a preset evaluation index, where the preset evaluation index includes at least one of prediction accuracy, recall rate, AUC score and F1 score.

(2) Data pre-processing

In the embodiment of the present invention, before performing model training, data preprocessing may also be performed on a training sample set and/or a test sample set to be processed, so as to obtain the training sample set and/or the test sample set after data preprocessing.

The data preprocessing process includes data cleansing and/or data normalization. The purpose of data cleansing and data normalization is to facilitate comparison and evaluation. Specifically, preferably, the data preprocessing includes at least one of missing value calculation, outlier exclusion, data transformation, non-dimensionalization, and normalization. The dimensionless method includes normalization, interval scaling, etc., the missing value calculation includes missing value filling, etc., and the data transformation includes polynomial data conversion, etc. For example, customer data is deleted or filled in, where the filling may be to complement a default value such as zero or an average value.

(3) Customer feature selection

In the embodiment of the invention, before model training, the characteristic selection model can be used for processing the client characteristics in the training sample set to be processed, the selected client characteristics are determined, and the selected client characteristics are used for screening the training sample set.

The customer characteristics are fields in the customer data that have a correlation with the likelihood of purchase by the customer. For example, the hit rate of searching potential customers from customers with the total assets above 5 ten thousand yuan is 35 times higher than that of searching potential customers from customers with the total assets below 5 ten thousand yuan, and the customers with the total assets above 5 ten thousand yuan is one of the main indexes of the potential customers. The amount of the customer's last transaction and the time of the last transaction have a considerable effect on their purchase of regular products. The positive change of the monthly-daily average change rate of the deposit of the client and the higher year and day of the deposit express the stable growth of the client assets and have stable fund sources. Accumulating the amount of the periodic purchases, the number of products on periodic deposits held, and the status of the assets of the customer over the previous quarter all have a large impact on the customer's purchase of the periodic products.

In an embodiment of the present invention, the customer data may include at least one of the following data: transaction behavior feature data, RFM (Recency, Frequency, currency) behavior pattern data, customer attribute feature data, asset liability feature data, credit attribute feature data, holding behavior feature data, and relationship delineation information.

Wherein the transaction behavior characteristics include at least one of: the accumulated transaction times of the month/quarter, the bank loan account-entering times, the transfer account-transferring times, the payroll income times, the cash deposit times, the loan repayment times, the transfer account-transferring times, the cash payment times, the consumption payment times, the life payment times and the last transaction amount of the client;

the RFM behavioral patterns include at least one of: time of last time of purchasing financing products, frequency of purchasing financing products in the last March, amount of money purchased in the last March, time of purchasing fund products in the last March, frequency of purchasing fund products in the last March, amount of money purchased fund products in the last March, number of days this day after card swiping in the last consumption, frequency of consumption in the last March, amount of consumption in the last March, the first name of the type of the last March consumption merchant (top1), the last account moving date of all the accounts in the name, and the number of days this day after the last account moving;

the customer attribute characteristics include at least one of: gender, age, home address, industry, job title, academic calendar, marital status, general family population, cell phone number, affiliation institution, customer manager, age of relationship with me (length of time to become a bank customer);

the asset liability characteristics include at least one of: maximum asset concentration, deposit time point balance, financing time point balance, fund time point balance, national debt time point balance, whether national debt is signed, whether trust is signed, whether precious metal is signed, whether third party inventory is signed, whether insurance is signed, deposit month-day is uniform, financing month-day is uniform, national debt month-day is uniform, deposit quarter-day is uniform, financing season-day is uniform, fund quarter-day is uniform, national debt season-day is uniform, deposit year-day is uniform, financing year-day is uniform, fund year-day is uniform, national debt year-day is uniform;

the credit attribute features include at least one of: the method comprises the following steps of (1) collecting a core client number, a client current grade, a client credit grade, a current loan five-grade classification, a comprehensive credit line, a client last grade, grade change time, a client last credit grade, credit grade change time, a last loan five-grade classification, a loan classification change date, a social insurance balance and a public fund monthly payment;

the holding behavior characteristics include at least one of: the method comprises the following steps of (1) total amount of assets, deposit balance, held product number, deposited product number, loan product number, current debit card type number, current credit card type number, current debit card number, current credit card number, financing product number, fund product number, signed service type product number, signed channel type product number, accumulated purchase periodic time number, accumulated purchase periodic sum, held periodic deposit product number, accumulated loan time number, accumulated loan application time number, accumulated purchase financing time number, accumulated fund purchase time number, total assets of a customer in the previous quarter, whether a periodic product is held, maximum asset balance and maximum asset category;

the relationship delineation includes at least one of: transfer application, whether line is crossed, the number of strokes transferred in this month, the amount of money transferred in this month, and the bank of the opposite party;

the above-mentioned accumulation is performed based on this month, this quarter, and the like.

Preferably, in an embodiment of the present invention, the step of processing the client features in the training sample set to be processed by using the feature selection model, and determining the selected client features includes:

screening the client characteristics in the training sample set to be processed by adopting at least one screening method of mutual information, chi-square test and F verification; the step is a coarse screening step;

when at least two screening methods are adopted, carrying out aggregation processing on the client characteristics in the training sample set to be processed screened by adopting different screening methods to obtain the client characteristics after the aggregation processing; the aggregation processing is, for example, to take an intersection, a union, or an optimum set.

Further, at least one of a recursive feature elimination method and a feature model elimination method is adopted to screen the customer features to obtain the selected customer features, and the step is a fine screening step. The recursive feature elimination method and the feature model elimination method are based on different processing mechanisms, and specifically, feature screening is performed based on algorithm models such as extreme trees, random forests, bayes and the like.

In the embodiment of the present invention, further, the method may further include:

and performing dimension reduction processing on the client characteristics by adopting at least one of characteristic orthogonalization, characteristic principal component analysis and matrix decomposition. Feature dimension reduction refers to reducing the dimension of a feature, and feature transformation may be included in feature dimension reduction, that is, a feature is fundamentally changed and an original feature disappears (although some properties of the original feature are also maintained by a new feature).

In an embodiment of the present invention, after the step of determining the purchasing probability prediction model of the financial product, the method further includes:

calculating importance information of customer features through a purchase probability prediction model of the financial product;

In the embodiment of the invention, before rough screening, customer characteristics (fields with correlation to the prediction result) can be added according to business knowledge (based on business requirements), because the characteristics before rough screening are conventional characteristics, and customer characteristics (customer characteristics which are designed by users based on business understanding and business requirements and have strong interpretability on prediction targets) customized by users need to be added.

For example, the following additional fields are added for a financial deposit product.

Newly adding fields: whether target clients are regularly target clients in one year, whether target clients are large in one year, whether target clients are regularly newly added in one year, whether target clients are frequently added in one year, whether target clients are cancelled in one year, whether target clients are regularly lost in one year, two product interest rates are fixed, a deposit product interest rate is notified in one day, a deposit product interest rate is notified in seven days, a month product interest rate, a three month product interest rate, a six month product interest rate, a nine month product interest rate, a one year product interest rate, a two year product interest rate, a three year product interest rate, and a five year product interest rate.

For example, for a financial savings product, the features selected for the final feature include: the total assets of the client in the previous quarter, whether the client holds regular products, the last transaction amount of the client, the monthly and daily average change rate of the deposit, the annual and daily average of the deposit, the accumulated purchase regular amount, the number of products holding the regular deposit, the maximum asset balance, the maximum asset class and the like.

(4) Financial product recommendation

In an embodiment of the present invention, after the step of predicting the probability that the customer to be predicted purchases the financial product at the second time, the method further includes:

1) aiming at a financial product, sorting the predicted probability of purchasing the financial product of a plurality of clients to be predicted from high to low, selecting the first Y clients to be predicted (for example 200 clients), recommending the financial product to the first Y clients to be predicted, wherein Y is a positive integer greater than or equal to 1, and the numerical value of Y can be adjusted according to the manpower of a website.

2) And aiming at one customer to be predicted, sorting the probability of purchasing a plurality of financial products at the second moment of the customer to be predicted according to the sequence from high to low, selecting the first Z financial products, recommending the first Z financial products to the customer to be predicted, wherein Z is a positive integer greater than or equal to 1, and the numerical value of Z can be adjusted according to the manpower of the website.

For example:

customer 1: it is suitable for recommending large inventory (probability: 0.62), and then recommending one year and more for settlement (probability 0.35).

And (3) a client 2: it is suitable for recommending large inventory (probability: 0.5), and then recommending one year and more (probability 0.38).

And 3, a client: it is suitable to recommend a reserve of one year and more (probability: 0.54), and then recommend a large deposit list (probability 0.31).

And a client 4: it is suitable for recommending one year and more for storage (probability 0.53).

In the specific application, in order to evaluate the effectiveness of marketing, two similar websites (for example, two banking websites determined according to the number of tellers and the number of self-service devices) are selected, one is an intervention group and is given to a prediction list for marketing, and the other is a comparison group and is not given to the prediction list, and a traditional method is adopted. Marketing measures include special activities, emotions, physical objects and the like. After two months, the marketing results of the two websites are compared. The results of the marketing include marketing success and pull-up (i.e., expanded extravehicular funds). After the method is used, the marketing success rate and the refresh rate of the intervention group are obviously improved compared with those of the control group.

If the marketing effect of the intervention group is not improved or the effect is not improved as compared with the control group, the previous steps can be adjusted, such as adjusting the customer characteristics, and the customer characteristics can be specifically increased, such as defining some customer characteristics outside the range of the customer characteristics considered before.

In the embodiment of the invention, the purchase probability prediction model of the financial product is obtained by training the real historical customer data of a large number of customers and depends on a big data technology, so that the probability prediction accuracy of the purchase probability prediction model obtained by training on the financial product purchased by the customer at a certain moment is higher, the target customer meeting the business marketing work requirement can be calculated, the accurate screening and personalized hierarchical marketing of the customer are realized, the customer experience can be improved, and the income of a financial institution can be improved at the same time.

Based on the same inventive concept, referring to fig. 5, the present invention further provides a financial product purchase prediction system 50, comprising:

a first obtaining module 51, configured to obtain customer data of a customer to be forecasted who does not hold a financial product at a first time;

a forecasting module 52, configured to input customer data of each customer to be forecasted into a purchasing probability forecasting model of the financial product, and forecast a probability that each customer to be forecasted purchases the financial product at a second time, where the second time is later than the first time; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers.

Preferably, when there are 12 said purchasing probability prediction models corresponding to each said financial product, said prediction module is configured to determine the month in which said second time is located; and selecting a purchase probability prediction model corresponding to the month in which the second moment is positioned, and predicting the probability of purchasing the financial product by each customer to be predicted at the second moment.

Preferably, the financial product purchase prediction system further comprises:

a training module, configured to set parameters of each of the algorithm models, input the training sample set to each of the algorithm models, and predict a probability that the customer for training purchases the financial product at a fifth past time, where the fifth past time is later than the fourth past time;

the adjusting module is used for adjusting parameters corresponding to the algorithm models according to the predicted probability that the training customer purchases the financial product at the fifth moment to obtain N trained algorithm models;

Preferably, the adjusting module is configured to adjust parameters of corresponding algorithm models according to the predicted probability that the training customer purchases the financial product at the fifth moment and a preset evaluation index, so as to obtain N adjusted algorithm models; inputting the training sample set into the adjusted algorithm model, and predicting the probability that the training customers in the training sample set purchase the financial products at the fifth moment; determining a first evaluation index value of each algorithm model according to the predicted probability of the training customer in the training sample set purchasing the financial product at the fifth moment and a preset evaluation index; obtaining a test sample set, wherein the test sample set is a set of historical customer data of a plurality of training customers before a fourth time, the test sample set is different from the training sample set, the test sample set can be obtained by randomly splitting customer data of the training customers into the training sample set and the test sample set, and further the data splitting is performed before training an algorithm model; inputting the test sample set into the correspondingly adjusted algorithm model, and predicting the probability that the training customer in the test sample set purchases the financial product at the fifth moment; determining a second evaluation index value of each algorithm model according to the predicted probability of the training customer in the test sample set purchasing the financial product at the fifth moment and a preset evaluation index; comparing the first evaluation index value with the second evaluation index value to obtain the fitting degree of each algorithm model; and adjusting parameters of the corresponding algorithm model according to the fitting degree of the algorithm model to obtain the trained algorithm model.

Preferably, each financial product corresponds to a purchase probability prediction model, and the historical customer data of the second customer includes customer data of any one month in the data extraction period Te before the fourth time.

Preferably, each of the financial products corresponds to 12 purchase probability prediction models, and for a purchase probability prediction model for a specified month, the historical customer data of the first customer and the second customer includes customer data for a month before the specified month in M consecutive years before the third time, where M is a positive integer greater than or equal to 2.

Preferably, the weight of each of the algorithm models is the same.

Preferably, the second obtaining module is configured to perform dimension reduction processing on the selected customer feature by using at least one of feature orthogonalization, principal component analysis of the feature, and matrix decomposition.

Preferably, the financial product purchase prediction system further comprises:

the first recommending module is used for sorting the predicted probability of purchasing the financial products of the multiple to-be-predicted customers from high to low according to a financial product, selecting the front Y to-be-predicted customers and recommending the financial product to the front Y to-be-predicted customers, wherein Y is a positive integer greater than or equal to 1; and/or

Referring to FIG. 6, the present invention further provides a financial product purchase prediction system 60, which includes a memory 61, a processor 62, and a computer program stored in the memory 61 and executable on the processor; the computer program when executed by the processor 62 implements the steps of:

The processor 62 is responsible for managing the bus architecture and general processing, and the memory 61 may store data used by the processor 62 in performing operations.

Preferably, when 12 of the purchasing probability prediction models correspond to each of the financial products, the computer program when executed by the processor 62 further implements the steps of:

determining the month in which the second moment is located;

Preferably, the computer program when executed by the processor 62 further implements the steps of:

setting parameters of each algorithm model, inputting the training sample set into each algorithm model, and predicting the probability that the customer for training purchases the financial product at a fifth past moment, wherein the fifth past moment is later than the fourth moment;

Preferably, each financial product corresponds to 12 purchase probability prediction models, and for a purchase probability prediction model for a specified month, the historical customer data of the first customer and the second customer is customer data of a month before the specified month in M consecutive years before the year at the fourth time, where M is a positive integer greater than or equal to 2.

Preferably, N is a positive integer greater than or equal to 2, and the computer program when executed by the processor 62 further implements the steps of:

Preferably, the weight of each of the algorithm models is the same.

Preferably, the computer program when executed by the processor 62 further implements the steps of: and carrying out data preprocessing on the training sample set to be processed to obtain the training sample set after data preprocessing.

Preferably, the computer program when executed by the processor 62 further implements the steps of: and processing the customer characteristics in the training sample set to be processed by adopting a characteristic selection model, determining the selected customer characteristics, and screening the training sample set by adopting the selected customer characteristics.

Preferably, the computer program when executed by the processor 62 further implements the steps of: screening the client characteristics in the training sample set to be processed by adopting at least two screening methods of mutual information, chi-square test and F verification;

aggregating the client characteristics in the set of training samples to be processed screened by adopting different screening methods to obtain the aggregated client characteristics;

Preferably, the computer program when executed by the processor 62 further implements the steps of: and performing dimension reduction processing on the client characteristics by adopting at least one of characteristic orthogonalization, characteristic principal component analysis and matrix decomposition.

Preferably, the computer program when executed by the processor 62 further implements the steps of: calculating importance information of the selected customer characteristics through a purchase probability prediction model of the financial product;

Preferably, the computer program when executed by the processor 62 further implements the steps of: aiming at a financial product, sorting the predicted probability of purchasing the financial product of a plurality of clients to be predicted from high to low, selecting front Y clients to be predicted, and recommending the financial product to the front Y clients to be predicted, wherein Y is a positive integer greater than or equal to 1; and/or

The embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the computer program implements each process of the foregoing financial product purchase prediction method embodiment, and can achieve the same technical effect, and in order to avoid repetition, details are not repeated here. The computer-readable storage medium may be a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk.

While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A financial product purchase forecasting method, comprising:

inputting customer data of each customer to be predicted into a purchasing probability prediction model of the financial product, and predicting the probability of purchasing the financial product by each customer to be predicted at a second moment, wherein the second moment is later than the first moment; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers;

before the step of predicting the probability that each customer to be predicted purchases the financial product at the second moment, the method further comprises:

obtaining N algorithm models;

determining a purchase probability prediction model of the financial product according to the trained N algorithm models;

wherein N is a positive integer greater than or equal to 2, and the step of determining the purchase probability prediction model of the financial product according to the trained N algorithm models comprises:

2. The financial product purchase prediction method of claim 1, wherein the customer data of the customer to be predicted is customer data between the first time and a third time, the third time being earlier than the first time.

3. The financial product purchase prediction method of claim 1,

each financial product corresponds to a purchasing probability prediction model; or

4. The financial product purchase prediction method of claim 3, wherein when there are 12 purchase probability prediction models corresponding to each financial product, the step of inputting customer data of each customer to be predicted into the purchase probability prediction models of the financial products, and the step of predicting the probability that each customer to be predicted purchases the financial product at the second time comprises:

determining the month in which the second moment is located;

5. The financial product purchase prediction method of claim 1 wherein the step of adjusting parameters of corresponding algorithm models based on the predicted probability of the training customer purchasing the financial product at the fifth time to obtain N trained algorithm models comprises:

6. The financial product purchase prediction method of claim 5, wherein the preset evaluation index includes at least one of a prediction accuracy, a recall rate, an area under the curve AUC score and an F1 score.

7. The financial product purchase forecasting method of claim 1, wherein the plurality of training customers includes: a first customer who already holds the financial product at the fourth time, and a second customer who does not hold the financial product at the fourth time.

8. The financial product purchase prediction method of claim 7, wherein each of the financial products corresponds to a purchase probability prediction model, and the historical customer data of the first customer includes customer data between a sixth time and a seventh time, the sixth time being a time point when the first customer purchases the financial product, and the seventh time being earlier than the sixth time.

9. The financial product purchase prediction method of claim 8, wherein a time period between the sixth time and the seventh time is equal to a time period between the fifth time and the fourth time.

10. The financial product purchase prediction method of claim 7 or 8, wherein each of the financial products corresponds to a purchase probability prediction model, and the historical customer data of the second customer includes customer data of any one month in a data extraction period before the fourth time.

11. The financial product purchase prediction method of claim 7, wherein each of the financial products corresponds to 12 purchase probability prediction models, and for a purchase probability prediction model for a given month, the historical customer data for the first customer and the second customer comprises customer data for a month prior to the given month in M consecutive years prior to the year at the fourth time, where M is a positive integer greater than or equal to 2.

12. The financial product purchase prediction method of claim 1 wherein the N algorithmic models are selected from the following algorithmic models: the system comprises a gradient lifting algorithm model, a neural network algorithm model, a polynomial Bayesian algorithm model, a random forest algorithm model and a support vector machine algorithm model.

13. The financial product purchase prediction method of claim 1, wherein the voting model is a majority voting model or a weighted voting model.

14. The financial product purchase prediction method of claim 13, wherein the weight of the algorithm model is determined based on at least one of a prediction accuracy, a recall rate, an AUC score and an F1 score of the algorithm model.

15. The financial product purchase prediction method of claim 13, wherein the weight of each of the algorithm models is the same.

16. The financial product purchase prediction method of claim 1 wherein the step of obtaining a training sample set comprises:

17. The financial product purchase prediction method of claim 16, wherein the data preprocessing comprises at least one of missing value calculation, outlier rejection, data transformation, dimensionless, and normalization.

18. The financial product purchase prediction method of claim 1 wherein the step of obtaining a training sample set comprises:

19. The financial product purchase prediction method of claim 18 wherein the step of processing customer characteristics in the training sample set to be processed using the characteristic selection model to determine selected customer characteristics comprises:

20. The financial product purchase forecasting method of claim 19, wherein the step of determining the selected customer characteristic further comprises:

21. The financial product purchase prediction method of claim 18, wherein the step of determining a purchase probability prediction model for the financial product is followed by further comprising:

22. The financial product purchase prediction method of claim 1, wherein the step of predicting the probability that the customer to be predicted will purchase the financial product at the second time is followed by further comprising:

23. A financial product purchase forecasting system, comprising:

the forecasting module is used for inputting the customer data of each customer to be forecasted into the purchasing probability forecasting model of the financial product and forecasting the probability that each customer to be forecasted purchases the financial product at a second moment, wherein the second moment is later than the first moment; the purchase probability prediction model of the financial product is obtained by training historical customer data of a plurality of training customers;

the third acquisition module is used for acquiring N algorithm models;

the determining module is used for determining a purchasing probability prediction model of the financial product according to the trained N algorithm models;

wherein, N is a positive integer greater than or equal to 2, the determining module is configured to obtain a probability that the trained customer who is predicted by the trained N algorithm models purchases the financial product at the fifth time; inputting the probability of purchasing the financial product at the fifth moment of the training client predicted by the trained N algorithm models into a voting model to obtain the probability of purchasing the financial product at the fifth moment of the training client predicted by the purchasing probability prediction model of the financial product.

24. The financial product purchase prediction system of claim 23,

the customer data of the customer to be predicted is customer data between the first time and a third time, and the third time is earlier than the first time.

25. The financial product purchase prediction system of claim 23,

26. The financial product purchase prediction system of claim 25, wherein the prediction module, when there are 12 purchase probability prediction models for each of the financial products, is configured to determine the month in which the second time is located; and selecting a purchase probability prediction model corresponding to the month in which the second moment is positioned, and predicting the probability of purchasing the financial product by each customer to be predicted at the second moment.

27. The financial product purchase prediction system of claim 23,

the adjusting module is used for adjusting parameters of corresponding algorithm models according to the predicted probability that the training customer purchases the financial product at the fifth moment and a preset evaluation index to obtain N adjusted algorithm models; inputting the training sample set into the adjusted algorithm model, and predicting the probability that the training customers in the training sample set purchase the financial products at the fifth moment; determining a first evaluation index value of each algorithm model according to the predicted probability of the training customer in the training sample set purchasing the financial product at the fifth moment and a preset evaluation index; obtaining a test sample set, wherein the test sample set is a set of historical customer data of a plurality of training customers before a fourth time, and the test sample set is different from the training sample set; inputting the test sample set into the correspondingly adjusted algorithm model, and predicting the probability that the training customer in the test sample set purchases the financial product at the fifth moment; determining a second evaluation index value of each algorithm model according to the predicted probability of the training customer in the test sample set purchasing the financial product at the fifth moment and a preset evaluation index; comparing the first evaluation index value with the second evaluation index value to obtain the fitting degree of each algorithm model; and adjusting parameters of the corresponding algorithm models according to the fitting degree of each algorithm model to obtain the trained algorithm models.

28. The financial product purchase prediction system of claim 27, wherein the preset assessment indicators include at least one of prediction accuracy, prediction precision, recall, area under the curve AUC score and F1 score.

29. The financial product purchase prediction system of claim 23, wherein the plurality of training customers comprises: a first customer who already holds the financial product at the fourth time, and a second customer who does not hold the financial product at the fourth time.

30. The financial product purchase prediction system of claim 29, wherein each of the financial products corresponds to a purchase probability prediction model, and the historical customer data of the first customer includes customer data between a sixth time and a seventh time, the sixth time being a time point at which the first customer purchased the financial product, the seventh time being earlier than the sixth time.

31. The financial product purchase prediction system of claim 30, wherein a time period between the sixth time and the seventh time is equal to a time period between the fifth time and the fourth time.

32. The financial product purchase prediction system of claim 29 or 30, wherein each of the financial products corresponds to a purchase probability prediction model, and the historical customer data of the second customer comprises customer data of any one month in a data extraction cycle prior to the fourth time.

33. The financial product purchase prediction system of claim 29, wherein each of the financial products corresponds to 12 purchase probability prediction models, and wherein for a purchase probability prediction model for a given month, the historical customer data for the first customer and the second customer comprises customer data for a month prior to the given month in M consecutive years prior to the year at the fourth time, wherein M is a positive integer greater than or equal to 2.

34. The financial product purchase prediction system of claim 23 wherein the N algorithmic models are selected from the following algorithmic models: the system comprises a gradient lifting algorithm model, a neural network algorithm model, a polynomial Bayesian algorithm model, a random forest algorithm model and a support vector machine algorithm model.

35. The financial product purchase prediction system of claim 23, wherein the voting model is a majority voting model or a weighted voting model.

36. The financial product purchase prediction system of claim 35, wherein the weight of the algorithm model is determined based on at least one of a prediction accuracy, a prediction precision, a recall, an AUC score and an F1 score of the algorithm model.

37. The financial product purchase prediction system of claim 35 wherein the weight of each algorithm model is the same.

38. The financial product purchase prediction system of claim 23, wherein the second obtaining module is configured to perform data preprocessing on the training sample set to be processed to obtain a training sample set after data preprocessing.

39. The financial product purchase prediction system of claim 38, wherein the data preprocessing comprises at least one of missing value calculation, outlier rejection, data transformation, dimensionless, and normalization.

40. The financial product purchase prediction system of claim 23, wherein the second acquisition module is configured to process customer characteristics in the training sample set to be processed using a characteristic selection model, determine selected customer characteristics, and filter the training sample set using the selected customer characteristics.

41. The financial product purchase prediction system of claim 40, wherein the second obtaining module is configured to filter customer characteristics in the training sample set to be processed using at least two filtering methods selected from mutual information, chi-square test, and F-test; aggregating the client characteristics in the training sample set to be processed screened by adopting different screening methods to obtain the aggregated client characteristics; and screening the client characteristics after the polymerization treatment by adopting at least one of a recursive characteristic elimination method and a characteristic model elimination method to obtain the selected client characteristics.

42. The financial product purchase prediction system of claim 40, wherein the second capture module is configured to perform dimension reduction on the customer characteristics using at least one of feature orthogonalization, principal component analysis of features, and matrix decomposition.

43. The financial product purchase prediction system of claim 40, wherein the second obtaining module is configured to calculate importance information of the selected customer characteristics through a purchase probability prediction model of the financial product; and adjusting the adopted client characteristics according to the importance information of the client characteristics.

44. The financial product purchase prediction system of claim 23, further comprising:

45. A financial product purchase prediction system comprising a memory, a processor and a computer program stored on the memory and executable on the processor; wherein the processor, when executing the computer program, implements the method of predicting a purchase of a financial product as recited in any one of claims 1-22.

46. A computer-readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the steps in the method for predicting the purchase of a financial product according to any one of claims 1-22.