WO2017131696A1

WO2017131696A1 - Database server to predict sales

Info

Publication number: WO2017131696A1
Application number: PCT/US2016/015345
Authority: WO
Inventors: Choudur Lakshminarayan; John MARGAGLIONE; Tuan Anh PHAM
Original assignee: EntIT Software LLC
Current assignee: Micro Focus LLC
Priority date: 2016-01-28
Filing date: 2016-01-28
Publication date: 2017-08-03
Anticipated expiration: 2018-07-28

Abstract

In example implementations, a method executed by a processor is provided. The method collects historical sales data. The historical sales data is divided into a training data set and a testing data set. A classification model is calculated based on the training data set and validated with the testing data set. Information associated with a pending sale is received. The information comprises a plurality of factors. Based upon the classification model, a prediction is made whether the pending sale will successfully close. An action plan that includes changing at least one of the plurality of factors based on the predicting is generated.

Description

DATABASE SERVER TO PREDICT SALES

BACKGROUND

[0001] Sales forecasting is a tool used by sales organizations. Currently, sales forecasting is done in the aggregate (e.g., what will the total sales be for the entire organization for a quarter, or the year). Aggregate sales forecasting may help predict an overall sales, but may not provide information about a specific sale. As a result, aggregate sales forecasting may not provide any information that can be actionable for sales person or the organization.

BRIEF DESCRIPTION OF THE DRAWINGS

[0002] FIG. 1 is a block diagram of an example system of the present disclosure;

[0003] FIG. 2 is a block diagram of an example individual sales prediction database server;

[0004] FIG. 3 is a flow diagram of an example method for predicting sales; and

[0005] FIG. 4 is a block diagram of a non-transitory computer readable medium storing instructions executed by a processor.

DETAILED DESCRIPTION

[0006] The present disclosure discloses an example apparatus and method for predicting sales. As noted above, current sales forecasting methods forecast sales in the aggregate. However, predicting sales in the aggregate is too general to generate an action plan. For example, aggregate sales prediction methods do not provide any specific factors that may affect an outcome of a particular deal. As a result, sales forecasting methods that predict an aggregate sales amount do not allow a sales company to generate an action plan in response to a specific deal that may be at risk.

[0007] The present disclosure provides a more accurate and granular sales forecasting method that can provide a prediction for each pending sale. The correlations disclosed herein help to minimize type 2 error (e.g., error related to predicting a successful sale, when the sale was not successful) in a null hypothesis analysis. Moreover, the present disclosure can provide factors that affect the prediction of an outcome of each sale. The factors that affect the sale can be used to generate an action plan that provides action items for a sales manager. For example, if a negative factor is the sales representative that is assigned to the deal, the sales manager may immediately assign a new sales representative to the deal.

[0008] In addition, the present disclosure may use a database server that is dedicated for executing the prediction of the sales forecast. The data that is collected to train the prediction model may be stored in the database server and not moved from client to client as the prediction is executed. As a result, each client may connect to the database server to execute the prediction rather than downloading the data local on the client and moving the data from client to client each time the prediction is executed.

[0009] FIG. 1 illustrates an example system 100 of the present disclosure. In one example, the system 100 may include an individual sales prediction (ISP) database server 104 located in an Internet Protocol (IP) network 102. It should be noted that the IP network 102 has been simplified for ease of explanation. For example, the I P network 102 may include additional network elements that are not shown (e.g., a router, a gateway, a firewall, additional access networks, application servers, and the like).

[0010] In one example, a plurality of endpoints 1 10, 1 12 and 1 14 may be in communication with the ISP database server 104 via a wired or wireless connection. The endpoints 1 10, 1 12 and 1 14 may be any type of endpoint device such as a laptop computer, a desktop computer, a tablet computer, a smart phone, and the like. It should be noted that although three endpoint devices 1 10, 1 12 and 1 14 are illustrated in FIG. 1 that any number of endpoint devices may be deployed.

[0011] In one example, the endpoints 1 10, 1 12 and 1 14 may be located remotely from the ISP database server 104. In another example, the endpoints 1 10, 1 12 and 1 14 and the ISP database server 104 may be part of a common sales organization or enterprise. The endpoints 1 10, 1 12 and 1 14 may be located remotely from one another and assigned to different sales managers of different regions. The sales managers may access the ISP database server 104 via a respective endpoint 1 10, 1 12 or 1 14 to have the ISP database server 104 predict an outcome of pending sales and provide an accurate sales forecast. Based on the predicted outcome of the pending sales, the ISP database server 104 may generate an action plan for each pending sale that can be used by the sales managers or the sales organization.

[0012] In one example, the system 100 may include a third party database (DB) 106 and a third party DB 108. Although two third party DBs 106 and 108 are illustrated in FIG. 1 , it should be noted that any number (e.g., more or less) of third party DBs may be deployed. The third party DBs 106 and 108 may provide historical sales data for the sales organization. The historical sales data may include a plurality of factors associated with each sale and an outcome of the previous sale (e.g., successful sale or unsuccessful sale). The historical sales data may be collected by the ISP database server 104 and used to generate a classification model. The classification model can then be used by the ISP DB 104 to predict an outcome of pending sales. An example of a third party DB 106 and 108 that can be used may be Salesforce.com®.

[0013] In one implementation, a plurality of ISP database servers 104 may be deployed and used to process large amounts of data collected from the third party DB 106 and/or the third party DB 108. For example, the data may be split into smaller subsets of data and each smaller subset of data may be processed by a respective one of the plurality of ISP database servers 104. As a result, the large amounts of data may be processed in parallel by the plurality of ISP database servers 104. [0014] FIG. 2 illustrates a block diagram of an example ISP database server 104. In one example, the ISP database server 104 may include a processor 202. The processor may be in communication with a data collection engine 204, a classification engine 206, a prediction engine 208 and an action plan generator 210. The processor may execute instructions associated with the functions performed by the data collection engine 204, the classification engine 206, the prediction engine 208 and the action plan generator 210.

[0015] In one implementation, the data collection engine 204 may collect historical sales data. For example, the processor 202 may establish a wired or wireless connection with a third party DB 106 or 108 and collect historical sales data. In one example, the data collection engine 204 may periodically collect historical sales data to periodically update the classification model that is generated.

[0016] In one example, the historical sales data may include various information associated with each historical sale. For example, the information may include whether the sale was successful or unsuccessful, a date the sale closed if successful, age (e.g. , a number of days until the win/loss), the sales representative, an age of the sales represented, a forecast of a sales manager, a region, whether a sales representative prepared a plan to tackle the potential sale, and the like.

[0017] In one example, the data can be organized into a table. Each row of the table may be associated with a potential or pending sale. Each column may be associated with one of the factors described above. A final column may represent a binary outcome of the potential or pending sale (e.g., successful sale or unsuccessful sale).

[0018] In one implementation, the classification engine 206 may calculate a classification model using the historical sales data collected by the data collection engine 204. For example, the classification engine 206 may divide the historical sales data into a training data set and a testing data set. The proportions of how the historical sales data are divided into the training data set and the testing data set may be pre-defined. In one example, 70% of the historical sales data may be used for the training data set and 30% of the historical sales data may be used for the testing data set.

[0019] In one example, the classification engine 206 may determine which factors are the most correlated to the outcome of a pending sale. In one example, the classification engine 206 may pre-process the data to remove outliers and derive new factors from the raw data collected from the third party DB 106 and/or the third party DB 108.

[0020] In one example, after the data is pre-processed, a modified stepwise- discriminant analysis (SWLDA) may be applied to the historical sales data to identify the most significant factors that correlate to the outcome of a pending sale. The modified SWLDA is an iterative method where at each iteration a factor with the highest correlation and statistical significance is selected. In one example, a significant factor may be identified as a factor that is greater than a threshold of 90% based on the modified SWLDA. In one example, the modified SWLDA may iterate until no factor meets the threshold.

[0021] The significant factors may then be correlated to the training data set. In one example, the classification model may be calculated based on the identified significant factors using any type of data analysis method, such as, a k-nearest neighbors (KNN) analysis, a non-parametric kernel density estimation analysis, a linear dirichlet allocation (LDA) analysis, a quadratic discriminant analysis, a qualitative data analysis (QDA), and the like.

[0022] In one example, the classification model may be graded based on a performance of a particular data analysis method. In one example, the performance may be measured based on a sensitivity, a specificity and a false positive rate (FPR). A classification model that has a high sensitivity, high specificity and low FPR is desired. In one example, the KNN analysis may provide the best data analysis for calculating the best classification model.

[0023] In one example, the classification engine 206 may validate the classification model by applying a null hypothesis analysis using the testing data set. Previous methods provided a less accurate forecasting model because the null hypothesis was applied to a data set that was assumed to have a Gaussian distribution. However, the testing data is used to calculate a distance between the testing data to one of the binary outcomes (e.g., successful sale or unsuccessful sale) determined by the classification model. The distance values comprise a chi squared (χ²) distribution as opposed to a Gaussian distribution.

[0024] In addition, the present disclosure introduces a new error constant term, a, that can be applied to each independent test of the classification model performed using the testing data set. Thus, if there are n independent tests performed by the testing data set, the potential tail error would be minimized to aⁿ. In one example, a may be set to a value of 0.05.

[0025] In one example, the null hypothesis analysis may be applied using the testing data set on the classification model to minimize a type 2 error. In one example, since the classification model has a binary outcome, there may be two types of error. Type 1 error may be error introduced by classifying a pending sale as unsuccessful when the sale was actually a successful sale. Type 2 error may be introduced by classifying a pending sale as successful when the sale was actually an unsuccessful sale.

[0026] By applying the aⁿ error constant to the n independent tests performed by the testing data set, the probability that the classification model is invalidated by a new data point is exponentially reduced. In other words, the validation that is performed by the testing data by applying the aⁿ error constant ensures that the classification model is accurate at predicting a sales outcome of a pending sale.

[0027] In one implementation, the prediction engine 208 may predict whether a pending sale will successfully close based upon the classification model. For example, the prediction engine 208 may receive a pending sale from a communication session established with one of the endpoints 1 10, 1 12 or 1 14. For example, a sales manager may want to forecast sales and provide a pending sale or a plurality of pending sales to the ISP database server 104. In one example, the pending sale may include information associated with the pending sale (e.g., a plurality of factors described above). The prediction engine 208 may use the classification method to calculate a distance between a predicted outcome of the pending sale to one of the binary outcomes of the classification method (e.g., a successful sale or an unsuccessful sale). The prediction engine 208 may classify the pending sale based on which outcome the pending sale is closer to.

[0028] In one implementation, the action plan generator 210 may generate an action plan based upon the predicted outcome of the pending sale. In one example, if the pending sale is predicted to be successful, then the action plan may be to take no additional action. In other words, no factors are changed and the sales manager may continue to implement the action plan currently in place to close the pending sale.

[0029] However, if the pending sale is predicted to be unsuccessful, the action plan generator 210 may generate an action plan that changes at least one factor associated with the pending sale. In one example, a factor that has a highest significant factor for determining the outcome may be changed. For example, the sales representative may be the highest significant factor as described above using a SWLDA analysis. Thus, the sales representative that is currently assigned may have a low success rate for closing deals and the sales manager may reassign a new sales representative to the pending sale.

[0030] In another example, an age of the pending sale may be the highest significant factor. As a result, the action plan may be to implement a more accelerated action plan to work towards closing the pending sale sooner otherwise the pending sale could become stale at which point the pending sale would close unsuccessfully. In some examples the generated action plan may including changing more than one factor.

[0031] In one example, the classification model engine 206 may calculate the classification model locally and the prediction engine 208 may predict an outcome of the pending sale locally on the ISP database server 104. In other words, the historical sales data, the classification model and the functions of the prediction engine 208 are not transferred or downloaded locally to an endpoint device 1 10, 1 12 or 1 14 when a pending sales prediction is requested. As a result, the ISP database server 104 ensures that the same classification model is applied for each prediction requested by the remotely located endpoint devices 1 10, 1 12 and 1 14. In addition, the ISP database server 104 ensures that the classification model that is applied for each prediction request is calculated based on the same historical sales data. As a result, different sales managers in different regions may receive an accurate sales prediction for each pending sale that is based off of a consistent and accurate classification model.

[0032] FIG. 3 illustrates a flow diagram of an example method 300 for predicting sales. In one example, the blocks of the method 300 may be performed by the ISP database server 104.

[0033] At block 302, the method 300 begins. At block 304, the method 300 collects historical sales data. For example, the historical sales data may be collected from a third party database such as Salesforce.com.

[0034] At block 306, the method 300 divides the historical sales data into a training data set and a testing data set. For example, a portion of the historical sales data may be used as a training data set to calculate a classification model and the remaining portion of the historical sales data may be used to validate the classification model. In one example, 70 percent of the historical sales data may be separated as the training data set and the remaining 30 percent of the historical sales data may be separated out as the testing data set.

[0035] At block 308, the method 300 calculates a classification model based on the training data set. The classification model may be validated with the testing data set that has a chi squared distribution by applying a null hypothesis and an error constant to each one of the testing data set that is compared to the classification model. For example, null hypothesis may be applied to the chi squared distribution (as opposed to a Gaussian distribution that is typically used) that comprises a distribution of distances that are calculated between the testing data set and the classification model. The null hypothesis may validate the classification model by minimizing a type 2 error (e.g., an error associated with classifying a pending sale as a successful sale when the sale is an unsuccessful sale) by applying the error constant to each one of the testing data set that is compared to the classification model.

[0036] As described above, the error constant may be increased

exponentially by a value associated with the number of independent tests performed on the classification method. For example, if there are 30

independent tests run on the classification method using 30 data points in the testing data set, then the error probability of the classification method may be equal to the error constant raised to a power of 30 (e.g., a³⁰). In other words, the testing data is used to validate the classification method to ensure that the classification method provides an accurate prediction.

[0037] The classification model may be calculated using any one of a variety of different statistical analysis methods. In one example, the classification model may be calculated using at least one of a KNN analysis, a non-parametric kernel density estimation analysis, an LDA analysis, a quadratic discriminant analysis, or a QDA analysis. In one example, KNN analysis may provide the most accurate classification model.

[0038] In one implementation, the classification method may also be based on a subset of factors extracted from the historical sales data. As described above, the subset of factors may be extracted using a SWLDA analysis to identify significant factors that determine an outcome of a pending sale.

[0039] In one example, the method 300 may periodically repeat blocks 304, 306 and 308. For example, to continually provide the most accurate

classification model, the method 300 may periodically recalculate the

classification model using the most recent historical sales data.

[0040] At block 310, the method 300 receives information associated with a pending sale, wherein the information comprises a plurality of factors. For example, a sales manager may want to forecast sales for a time period. The sales manager may send, via an endpoint, pending sales and information associated with the pending sales to the ISP database server to receive a prediction of whether each one of the pending sales will be successful or unsuccessful.

[0041] At block 312, the method 300 predicts whether the pending sale will successfully close based upon the classification model. For example, the method 300 may use the classification method to calculate a distance between a predicted outcome of the pending sale to one of the binary outcomes of the classification method (e.g., a successful sale or an unsuccessful sale). The method 300 may classify the pending sale based on which outcome the pending sale is closer to.

[0042] At block 314, the method 300 may generate an action plan that includes changing at least one of the plurality of factors based on the predicting. In one example, if the pending sale is predicted to be successful, then the action plan may be to take no additional action. In other words, no factors are changed and the sales manager may continue to implement the action plan currently in place to close the pending sale.

[0043] However, if the pending sale is predicted to be unsuccessful, the action plan may change at least one factor associated with the pending sale. In one example, a factor that has a highest significant factor for determining the outcome may be changed. In another example, a plurality of significant factors may be changed. At block 316, the method 300 ends.

[0044] FIG. 4 illustrates an example of an apparatus 400. In one example, the apparatus may be the ISP database server 104. In one example, the apparatus 400 may include a processor 402 and a non-transitory computer readable storage medium 404. The non-transitory computer readable storage medium 404 may include instructions 406, 408, 410 and 412 that when executed by the processor 402, cause the processor 402 to perform various functions.

[0045] In one example, the instructions 406 may include instructions to divide collected historical sales data into a training data set and a testing data set. The instructions 408 may include instructions to calculate a classification model based. The classification model may be based on the training data set, wherein the classification model is validated with the testing data set that has a chi squared distribution by applying a null hypothesis and an error constant to each one of the testing data set that is compared to the classification model. The instructions 410 may include instructions to predict whether a pending sale will successfully close. The prediction may be based on a comparison between a plurality of factors associated with the pending sale and the classification model. The instructions 412 may include instructions to generate an action plan. The action plan may include changing at least one of the plurality of factors based on the instructions to predict.

[0046] It will be appreciated that variants of the above-disclosed and other features and functions, or alternatives thereof, may be combined into many other different systems or applications. Various presently unforeseen or unanticipated alternatives, modifications, variations, or improvements therein may be subsequently made by those skilled in the art which are also intended to be encompassed by the following claims.

Claims

1 . A method, comprising:

collecting, by a processor, historical sales data;

dividing, by the processor, the historical sales data into a training data set and a testing data set;

calculating, by the processor, a classification model based on the training data set, wherein the classification model is validated with the testing data set that has a chi squared distribution by applying a null hypothesis and an error constant to each one of the testing data set that is compared to the

classification model;

receiving, by the processor, information associated with a pending sale, wherein the information comprises a plurality of factors;

predicting, by the processor, whether the pending sale will successfully close based upon the classification model; and

generating, by the processor, an action plan that includes changing at least one of the plurality of factors based on the predicting.

2. The method of claim 1 , wherein the chi squared distribution comprises a distribution of distances that are calculated between the testing data set and the classification model.

3. The method of claim 1 , wherein the null hypothesis validates the classification model by minimizing a type 2 error by applying the error constant to the each one of the testing data set that is compared to the classification model.

4. The method of claim 3, wherein the type 2 error is an error associated with classifying the pending sale as a successful sale when the sale is an unsuccessful sale.

5. The method of claim 1 , wherein the historical sales data is collected from a third party database.

6. The method of claim 1 , wherein the classification model is based on a subset of factors extracted from the historical sales data using a modified stepwise-discriminant analysis.

7. The method of claim 6, wherein at least one of: a k-nearest neighbors analysis, a non-parametric kernel density estimation analysis, a linear dirichlet allocation analysis, a quadratic discriminant analysis, or a qualitative data analysis, is applied to the subset of factors and the historical sales data.

8. The method of claim 1 , wherein the action plan comprises changing at least one of the plurality of factors that will change a prediction of the pending sale to a successful sale when the pending sale is predicted to be an unsuccessful sale.

9. An apparatus, comprising:

a processor;

a data collection engine that collects historical sales data;

a classification engine that calculates a classification model based on a training data set divided out of the historical sales data, wherein the

classification model is validated with a testing data set by applying a null hypothesis and an error constant to each one of the testing data set that is compared to the classification model, wherein the testing data set divided out of the historical sales data that has a chi squared distribution;

a prediction engine that predicts whether a pending sale will successfully close based upon the classification model; and

an action plan generator that generates an action plan based on the predicting.

10. The apparatus of claim 9, wherein the chi squared distribution comprises a distribution of distances that are calculated between the testing data set and the classification model.

1 1 . The apparatus of claim 9, wherein the null hypothesis validates the classification model by minimizing a type 2 error by applying the error constant to the each one of the testing data set that is compared to the classification model.

12. The apparatus of claim 9, wherein the classification model is based on a subset of factors extracted from the historical sales data using a modified stepwise-discriminant analysis.

13. The apparatus of claim 12, wherein at least one of: a k-nearest neighbors analysis, a non-parametric kernel density estimation analysis, a linear dirichlet allocation analysis, a quadratic discriminant analysis, or a qualitative data analysis, is applied to the subset of factors and the historical sales data.

14. A non-transitory computer-readable storage medium encoded with instructions executable by a processor, the computer-readable storage medium comprising:

instructions to divide collected historical sales data into a training data set and a testing data set;

instructions to calculate a classification model based on the training data set, wherein the classification model is validated with the testing data set that has a chi squared distribution by applying a null hypothesis and an error constant to each one of the testing data set that is compared to the

classification model;

instructions to predict whether a pending sale will successfully close based on a comparison between a plurality of factors associated with the pending sale and the classification model; and

instructions to generate an action plan that includes changing at least one of the plurality of factors based on the instructions to predict.

15. The non-transitory computer readable storage medium of claim 14, wherein the null hypothesis validates the classification model by minimizing a type 2 error by applying the error constant to the each one of the testing data set that is compared to the classification model.