WO2019109790A1 - 销量预测方法、装置和计算机可读存储介质 - Google Patents

销量预测方法、装置和计算机可读存储介质 Download PDF

Info

Publication number
WO2019109790A1
WO2019109790A1 PCT/CN2018/115652 CN2018115652W WO2019109790A1 WO 2019109790 A1 WO2019109790 A1 WO 2019109790A1 CN 2018115652 W CN2018115652 W CN 2018115652W WO 2019109790 A1 WO2019109790 A1 WO 2019109790A1
Authority
WO
WIPO (PCT)
Prior art keywords
sku
sales
tested
attribute
model
Prior art date
Application number
PCT/CN2018/115652
Other languages
English (en)
French (fr)
Inventor
高雨薇
宋磊
胡壁
马丽晨
Original Assignee
北京京东尚科信息技术有限公司
北京京东世纪贸易有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京京东尚科信息技术有限公司, 北京京东世纪贸易有限公司 filed Critical 北京京东尚科信息技术有限公司
Publication of WO2019109790A1 publication Critical patent/WO2019109790A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising

Definitions

  • the present disclosure relates to the field of data processing, and in particular, to a sales forecasting method, apparatus, and computer readable storage medium.
  • the sales plan is the execution plan of the sales target, and it undertakes the landing of the company's strategic plan, and guides the operation plan of suppliers, warehousing and logistics to play a core role.
  • the sales plan needs to be adjusted in a timely manner following factors such as market dynamics.
  • the formulation of sales plans is mainly based on experience and sales personnel.
  • One technical problem to be solved by the embodiments of the present disclosure is how to improve the accuracy of the sales forecast.
  • a method for predicting sales volume includes: determining a time attribute of a SKU to be tested according to a distribution of sales amount information of a unit of inventory SKU in different time units; obtaining a test to be tested a sales forecasting model corresponding to the feature vector of the SKU and the time attribute of the SKU to be tested, wherein the feature vector of the SKU to be tested includes historical data and predicted time information of the SKU to be tested; and the feature vector of the SKU to be tested is input into the predictive model, Get sales forecast results.
  • determining the time attribute corresponding to the SKU to be tested according to the distribution of the sales information of the SKU to be tested in different time units comprises: generating a trend vector for the plurality of SKUs including the SKU to be tested, wherein each trend vector The sales information of the corresponding SKU in each time unit is included; the trend vector corresponding to the plurality of SKUs is clustered, and the time attribute of the SKU to be tested is determined according to the category of the SKU to be tested in the clustering result.
  • the time attribute is a seasonal attribute.
  • the sales forecasting method further includes: in the plurality of commodity attributes of the SKU having the same time attribute, the proportion is exceeded
  • the commodity attribute of the preset value is determined as the commodity attribute corresponding to the time attribute; the commodity attribute of the SKU to be tested is matched with the commodity attribute corresponding to the different time attribute; and the time attribute of the SKU to be tested is determined according to the matching result.
  • the output of the model results in a sales forecast.
  • the sales forecast results are obtained based on the weighted sum of the output results of the plurality of sales forecast models, and the weight corresponding to the output of each of the sales forecast models is determined according to the prediction error of the sales forecast model.
  • the plurality of sales forecasting models include a first sales forecasting model and a second sales forecasting model; the sales forecasting result P is obtained using the following formula:
  • a is the prediction error of the first sales forecasting model
  • b is the prediction error of the second sales forecasting model
  • g1 is the output value of the first sales forecasting model
  • g2 is the output value of the second sales forecasting model.
  • the sales forecasting method further comprises: acquiring SKU feature vectors for training and corresponding tag values having the same time attribute, wherein the tag value is a sales volume of the corresponding SKU; a SKU feature vector to be used for training Input into the model to be trained; adjust the parameters of the model to be trained according to the difference between the output of the model to be trained and the mark value of the corresponding SKU feature vector, and obtain a sales forecast model.
  • the feature vector of the SKU to be tested includes one or more of the following characteristics: historical sales for each time period of the SKU, historical average sales for each time period of the SKU, sales of the SKU for each time unit in history, Sales of SKUs in the same period last year, promotional information related to SKUs, and product attributes of SKUs.
  • the sales forecasting method further includes: obtaining a predicted sales ratio of each SKU to be tested according to the sales forecast results of the plurality of SKUs to be tested; and according to the planned total sales volume and the predicted sales ratio of each SKU to be tested, Update the sales forecast results of each SKU to be tested.
  • a sales forecasting apparatus including: a time attribute determining module, configured to determine a SKU to be tested according to a distribution of sales amount information of a stock quantity unit SKU in different time units a time attribute; a feature vector and a model acquisition module, configured to acquire a feature vector of the SKU to be tested and a sales forecast model corresponding to the time attribute of the SKU to be tested, wherein the feature vector of the SKU to be tested includes historical data and prediction of the SKU to be tested Time information; a sales forecasting module, configured to input a feature vector of the SKU to be tested into the prediction model to obtain a sales forecast result.
  • the time attribute determination module is further configured to generate a trend vector for the plurality of SKUs including the SKU to be tested, wherein each trend vector includes sales information of the corresponding SKU in each time unit; corresponding to the plurality of SKUs The trend vector is clustered, and the time attribute of the SKU to be tested is determined according to the category of the SKU to be tested in the clustering result.
  • the time attribute is a seasonal attribute.
  • the sales forecasting device further includes: a new product time attribute determining module, configured to: in the case that the number of time units involved in the sales information of the SKU to be tested is less than a preset value, in the plurality of SKUs having the same time attribute In the product attribute, the item attribute exceeding the preset value is determined as the item attribute corresponding to the time attribute; the item attribute of the SKU to be tested is matched with the item attribute corresponding to the different time attribute; and the SKU of the to-be-tested is determined according to the matching result.
  • Time attribute configured to: in the case that the number of time units involved in the sales information of the SKU to be tested is less than a preset value, in the plurality of SKUs having the same time attribute In the product attribute, the item attribute exceeding the preset value is determined as the item attribute corresponding to the time attribute; the item attribute of the SKU to be tested is matched with the item attribute corresponding to the different time attribute; and the SKU of the to-be-tested is determined according to the matching result.
  • the feature vector and the model acquisition module are configured to acquire a feature vector of the SKU to be tested and a plurality of sales prediction models corresponding to the time attribute of the SKU to be tested; the sales prediction module is further configured to separately use the feature vector of the SKU to be tested Input into a plurality of sales forecasting models, and obtain sales forecast results based on output results of a plurality of sales forecasting models.
  • the sales forecasting module is further configured to obtain the sales forecasting result according to the weighted sum of the output results of the plurality of sales forecasting models, and the weight corresponding to the output result of each of the sales forecasting models is determined according to the forecasting error of the sales forecasting model.
  • the plurality of sales forecasting models include a first sales forecasting model and a second sales forecasting model; the sales forecasting module is further configured to obtain the sales forecasting result P using the following formula:
  • a is the prediction error of the first sales forecasting model
  • b is the prediction error of the second sales forecasting model
  • g1 is the output value of the first sales forecasting model
  • g2 is the output value of the second sales forecasting model.
  • the sales forecasting device further includes a sales forecasting model training module for acquiring SKU feature vectors for training and corresponding tag values having the same time attribute, wherein the tag value is the sales volume of the corresponding SKU;
  • the SKU feature vector for training is input into the model to be trained; the parameters of the model to be trained are adjusted according to the difference between the output result of the model to be trained and the tag value of the corresponding SKU feature vector, and a sales prediction model is obtained.
  • the feature vector of the SKU to be tested includes one or more of the following characteristics: historical sales for each time period of the SKU, historical average sales for each time period of the SKU, sales of the SKU for each time unit in history, Sales of SKUs in the same period last year, promotional information related to SKUs, and product attributes of SKUs.
  • the sales forecasting device further includes a planned sales disassembly module, configured to obtain a predicted sales ratio of each SKU to be tested according to the sales forecast results of the plurality of SKUs to be tested; The predicted sales ratio of the SKU is measured, and the sales forecast result of each SKU to be tested is updated.
  • a planned sales disassembly module configured to obtain a predicted sales ratio of each SKU to be tested according to the sales forecast results of the plurality of SKUs to be tested; The predicted sales ratio of the SKU is measured, and the sales forecast result of each SKU to be tested is updated.
  • a sales forecasting apparatus comprising: a memory; and a processor coupled to the memory, the processor being configured to be based on an instruction stored in the memory Execute any of the aforementioned sales forecasting methods.
  • a computer readable storage medium having stored thereon a computer program, wherein the program is executed by a processor to implement any of the foregoing sales forecasting methods.
  • Some of the above embodiments have the following advantages or advantages: by determining the sales forecasting model according to the time attribute of the SKU to be tested and using the determined sales forecasting model for automatic sales forecasting, the forecasting result can be made to match the SKU sales volume to be tested.
  • the distribution of time increases the accuracy of sales forecasting and reduces labor costs, which in turn improves the operational efficiency of the entire e-commerce platform.
  • FIG. 1 is a flow diagram of a method of forecasting sales volume, in accordance with some embodiments of the present disclosure.
  • FIG. 2 is a flow diagram of a method for determining a time attribute of a SKU, in accordance with some embodiments of the present disclosure.
  • FIG. 3 is a schematic flow chart of a method for determining a time attribute of a SKU according to further embodiments of the present disclosure.
  • FIG. 4 is a flow diagram of a method for forecasting sales using multiple models, in accordance with some embodiments of the present disclosure.
  • FIG. 5 is a flow diagram of a method of training a sales forecasting model in accordance with some embodiments of the present disclosure.
  • FIG. 6 is a schematic flow chart of a method for predicting sales volume according to further embodiments of the present disclosure.
  • FIG. 7 is a block diagram showing the structure of a sales forecasting device in accordance with some embodiments of the present disclosure.
  • FIG. 8 is a schematic structural diagram of a sales forecasting apparatus according to further embodiments of the present disclosure.
  • FIG. 9 is a schematic structural diagram of a sales forecasting apparatus according to still another embodiment of the present disclosure.
  • FIG. 1 is a flow diagram of a method of forecasting sales volume, in accordance with some embodiments of the present disclosure. As shown in FIG. 1, the sales forecasting method of this embodiment includes steps S102 to S106.
  • step S102 the time attribute of the SKU to be tested is determined according to the distribution of the sales information of the SKU (Stock Keeping Unit) in different time units.
  • the time unit can be set according to requirements, for example, each season can be a time unit, or each month, or every two months can be a time unit, and the like.
  • SKU sales trends can be reflected in time units. For example, SKUs for electric blankets are popular in winter; SKUs for mosquito nets and mosquito-repellent incense are selling well in summer; hairy crab SKUs are selling well in September and October; although wine products are purchased by a large number of users throughout the year, In January and February before and after the Spring Festival, sales of alcoholic products will peak during this period due to family gatherings and gatherings.
  • step S104 the feature vector of the SKU to be tested and the pin quantity prediction model corresponding to the time attribute of the SKU to be tested are acquired, wherein the feature vector of the SKU to be tested includes historical data and predicted time information of the SKU to be tested.
  • the sales forecasting model may be, for example, a Gradient Boosting Decision Tree (GBDT) model, a linear regression model, a Long Short-Term Memory (LSTM) model, or the like.
  • GBDT Gradient Boosting Decision Tree
  • LSTM Long Short-Term Memory
  • the historical data of the SKU to be tested reflects the sales situation of the SKU to be tested in the past period of time.
  • the historical data may be the data of the SKU itself or the data of the category of the SKU.
  • the feature vector of the SKU to be tested may include year-on-year growth information of the sales volume, such as the Gross Merchandise Volume (GMV) or the monthly average GMV of the SKU to be tested in the near future.
  • GMV Gross Merchandise Volume
  • the feature vector of the SKU to be tested may include the ring growth information of the sales volume, for example, the GMV of the category to which the SKU belongs is in the GMV of the same period of last year, or the GMV mean and variance of the period of the same period last year.
  • the predicted time information may be the time corresponding to the sales volume to be tested. For example, predicting the sales volume of a certain SKU in the winter, the winter may be the predicted time information; or predicting the sales volume of a certain SKU in December this year, the December may be the predicted time information.
  • the predicted time information can be represented in a variety of forms, such as dates, holiday indicia, and the like. If it is a date form, it can be either a solar date or a lunar date.
  • the feature vector of the SKU to be tested may also include promotional information related to the SKU. For example, the promotion category, the promotion duration, the number of SKUs covered by the promotion, the proportion of hot items in the promotion, and the like.
  • the promotional information of the SKU in the Friends platform can also be added to the feature vector.
  • the feature vector of the SKU to be tested may also include the commodity attributes of the SKU itself, such as specifications, sizes, colors, and the like.
  • step S106 the feature vector of the SKU to be tested is input into the prediction model to obtain a sales forecast result.
  • the sales forecasting model can be determined according to the time attribute of the SKU to be tested, and the determined sales forecasting model can be used for automatic sales forecasting, so that the forecasting result can be consistent with the distribution law of the SKU sales to be tested at different times.
  • the accuracy of sales forecasting reduces labor costs, which in turn improves the operational efficiency of the entire e-commerce platform.
  • the time attribute of the SKU can be determined by a method of clustering. An embodiment of a method of determining the temporal attribute of the disclosed SKU is described below with reference to FIG.
  • the time attribute determining method of the SKU of this embodiment includes steps S202 to S204.
  • step S202 a trend vector is generated for a plurality of SKUs, wherein each trend vector includes sales information of the corresponding SKU in each time unit.
  • the size of the time unit can be set as needed. For example, statistics can be made in units of months. If the sales volume of a SKU in January-December is 1, 0.8, 0.95, 0.98, 10.23, 15.6, 18.9, 19.7, 10.6, 2.5, 1, 1.2, and the unit is 10,000 yuan, the trend vector of the SKU is [ 1, 0.8, 0.95, 0.98, 10.23, 15.6, 18.9, 19.7, 10.6, 2.5, 1, 1.2].
  • step S204 the trend vectors corresponding to the plurality of SKUs are clustered, and the time attribute of the SKU to be tested is determined according to the category of the SKU to be tested in the clustering result.
  • the number of classifications may be 4, that is, each category corresponds to four seasons of spring, summer, autumn, and winter; the number of classifications may also be 5, that is, in addition to the four seasons, "the year-round heat is included.
  • Cell is a category that selects SKUs that do not change significantly over time.
  • the time attribute of each SKU can be automatically determined, thereby obtaining the time attribute of the SKU to be tested therein.
  • the method of the above embodiment can accurately and objectively identify the trend and regularity of the sales of the SKU over time, so that the time attribute of the SKU can be accurately identified.
  • the sales time of the SKU needs to reach a certain level before it can participate in the calculation, that is, the number of time units involved in the SKU needs to be greater than the preset value. For example, SKUs that have been sold for more than one year have data for each time period of the year; for some new categories or newly launched SKUs, the sales data is not enough to cover a whole year, so the time of the following embodiment can also be used to determine the time. Attributes. An embodiment of a method of determining the temporal attribute of the disclosed SKU is described below with reference to FIG.
  • FIG. 3 is a schematic flow chart of a method for determining a time attribute of a SKU according to further embodiments of the present disclosure. As shown in FIG. 3, the time attribute determining method of the SKU of this embodiment includes steps S302 to S306.
  • step S302 among the plurality of item attributes of the SKU having the same time attribute, the item attribute whose ratio exceeds the preset value is determined as the item attribute corresponding to the time attribute.
  • step S304 the item attributes of the SKU to be tested are matched with the item attributes corresponding to different time attributes.
  • step S306 the time attribute of the SKU to be tested is determined according to the matching result.
  • SKU1 having time attribute X has commodity attributes A, B, C
  • SKU2 has commodity attributes A, C, D
  • SKU3 has commodity attributes A, B, E
  • SKU2 has commodity attributes A, C, F, etc.
  • the commodity attributes A and C have the highest proportion.
  • the product attributes "thickening" and "plus velvet” have the highest proportion of product attributes, so they have the goods.
  • the SKUs of attributes A and C are most likely to be SKUs with time attribute X.
  • the new item SKU whose sales data amount coverage time is less than the preset value can also determine its time attribute, which further improves the applicable scenario of the present disclosure.
  • the sales duration of the new SKU reaches a preset value, it is also possible to re-verify whether the time attribute is reasonable by using the embodiment of FIG. 2 or the like.
  • the embodiment of the present disclosure may determine the sales forecast result based on the output result of one model when performing the prediction, or may be determined according to the output results of the plurality of models.
  • An embodiment in which the present disclosure employs a plurality of models for sales forecasting is described below with reference to FIG.
  • the sales prediction model training method of this embodiment includes steps S402 to S404.
  • step S402 a feature vector of the SKU to be tested and a plurality of sales prediction models corresponding to the time attribute of the SKU to be tested are acquired. For example, if the SKU to be tested is a winter hot item, then the GBDT winter model and the LSTM winter model can be obtained.
  • step S404 the feature vectors of the SKUs to be tested are respectively input into a plurality of sales prediction models, and the sales forecast results are obtained according to the output results of the plurality of sales prediction models.
  • the sales forecast result may be determined based on the weighted result of the output results of the plurality of sales forecast models.
  • the output results of multiple models can be integrated to improve the accuracy of the sales forecast.
  • the sales forecasting result may be obtained according to the weighted sum of the output results of the plurality of sales forecasting models, and the weight corresponding to the output result of each of the sales forecasting models is determined according to the forecasting error of the sales forecasting model.
  • An example of obtaining a prediction result is exemplarily provided below.
  • the two sales forecasting models are the first sales forecasting model and the second sales forecasting model respectively.
  • the sales forecast result P can be obtained by using formula (1):
  • a is the prediction error of the first sales forecasting model
  • b is the prediction error of the second sales forecasting model.
  • the prediction error can be determined according to the difference between the output value of the training stage and the mark value; g1 is the first sales forecast.
  • the output value of the model, g2 is the output value of the second sales forecast model. Therefore, the error magnitude of the model can be used as a weight to comprehensively consider the output of multiple models, further improving the accuracy of the sales forecast.
  • FIG. 5 is a flow diagram of a method of training a sales forecasting model in accordance with some embodiments of the present disclosure. As shown in FIG. 5, the sales prediction model training method of this embodiment includes steps S502 to S506.
  • step S502 the SKU feature vector for training and the corresponding tag value having the same time attribute are acquired, wherein the tag value is the sales volume of the corresponding SKU.
  • the feature vector used in the training phase is the same as the component vector used in the prediction phase, and will not be described here. Since the purpose of training is to obtain a proprietary model corresponding to a certain time attribute, the training data should also have the same time attribute.
  • step S504 the SKU feature vector for training is input into the model to be trained.
  • step S506 the parameters of the model to be trained are adjusted according to the difference between the output result of the model to be trained and the mark value of the corresponding SKU feature vector, and a sales prediction model is obtained.
  • GBDT is a gradient descent algorithm based on the Boosting integration method. The core idea is to build multiple decision trees in series. Each tree is used to correct the errors of the existing trees. GBDT is a machine learning algorithm that is widely used and has a good effect.
  • Linear regression is a regression analysis that models the relationship between one or more independent variables and dependent variables using a least squares function called a linear regression equation. This function is a linear combination of one or more model parameters called regression coefficients. The case of only one independent variable is called simple regression, and the case of more than one independent variable is called multiple regression.
  • LSTM is a special type of recurrent neural network that can learn long-term dependency information. LSTM avoids long-term dependencies through deliberate design.
  • the model in addition to different models according to time attributes, the model can be further subdivided.
  • the training data with promotional resources can be separately trained.
  • promotional resources for example, a SKU for a homepage push, a SKU with an activity column, a SKU with a large promotion force, and the like; in addition, a new category can also be used.
  • SKU training data is trained separately and so on.
  • FIG. 6 is a schematic flow chart of a method for predicting sales volume according to further embodiments of the present disclosure. As shown in FIG. 6, the sales forecasting method of this embodiment includes steps S602 to S620.
  • step S602 the time attribute of the SKU to be tested is determined according to the distribution of the sales amount information of the stock quantity unit SKU to be tested in different time units.
  • step S604 a feature vector of the SKU to be tested is acquired.
  • step S606 it is determined whether the SKU to be tested has a promotion resource. If yes, step S608 is performed; if not, step S610 is performed.
  • step S608 the sales model corresponding to the time attribute of the SKU to be tested is used to predict the sales volume of the SKU to be tested.
  • the promotion model is a sales prediction model obtained by training based on training data with promotional resources.
  • step S610 it is determined whether the SKU to be tested is a new product. If yes, step S612 is performed; if not, step S614 is performed.
  • step S612 the sales volume of the SKU to be tested is predicted by using the new product model corresponding to the time attribute of the SKU to be tested.
  • the new product model is a sales forecast model obtained by training based on the training data corresponding to the new product.
  • step S614 the sales volume of the SKU to be tested is predicted by using a common model corresponding to the time attribute of the SKU to be tested.
  • the general model is a sales forecast model obtained by training based on training data corresponding to non-new products and non-promotional items.
  • step S616 the prediction result is loaded into the cache so that the user can perform fast reading as needed.
  • a sales plan report is generated according to the predicted result. For example, an overall plan based on the estimated sales volume of the SKU to be tested may be generated in units of salespersons, categories, brands, and departments.
  • the predicted sales volume ratio of each SKU to be tested may be obtained according to the sales forecast results of the plurality of SKUs to be tested; and then each is updated according to the planned total sales volume and the predicted sales ratio of each SKU to be tested.
  • the sales forecast of the SKU to be tested may be obtained.
  • the plan can be disassembled according to the predicted result of the SKU, so that the predicted result can better meet the actual demand.
  • step S620 the data is persistently stored, for example, by HDFS (Hadoop Distributed File System).
  • HDFS Hadoop Distributed File System
  • the sales forecasting apparatus 70 of the embodiment includes: a time attribute determining module 710, configured to determine a time attribute of the SKU to be tested according to a distribution of sales amount information of the in-stock quantity unit SKU in different time units;
  • the feature vector and model acquisition module 720 is configured to acquire a feature vector of the SKU to be tested and a sales forecast model corresponding to the time attribute of the SKU to be tested, where the feature vector of the SKU to be tested includes historical data and predicted time information of the SKU to be tested;
  • the sales prediction module 730 is configured to input the feature vector of the SKU to be tested into the prediction model to obtain a sales prediction result.
  • the time attribute determination module 710 can be further configured to generate a trend vector for the plurality of SKUs including the SKU to be tested, wherein each trend vector includes sales information of the corresponding SKU in each time unit; The trend vector corresponding to the SKU is clustered, and the time attribute of the SKU to be tested is determined according to the category of the SKU to be tested in the clustering result.
  • the time attribute is a seasonal attribute.
  • the sales forecasting device 70 may further include: a new product time attribute determining module 740, configured to: in the case of the SKU having the same time attribute, if the number of time units involved in the sales information of the SKU to be tested is less than a preset value Among the plurality of commodity attributes, the commodity attribute having a ratio exceeding the preset value is determined as the commodity attribute corresponding to the time attribute; the commodity attribute of the SKU to be tested is matched with the commodity attribute corresponding to the different time attribute; Measure the time attribute of the SKU.
  • a new product time attribute determining module 740 configured to: in the case of the SKU having the same time attribute, if the number of time units involved in the sales information of the SKU to be tested is less than a preset value Among the plurality of commodity attributes, the commodity attribute having a ratio exceeding the preset value is determined as the commodity attribute corresponding to the time attribute; the commodity attribute of the SKU to be tested is matched with the commodity attribute corresponding to the different time attribute; Measure the time attribute of the
  • the feature vector and model acquisition module 720 can be configured to acquire a feature vector of the SKU to be tested and a plurality of sales prediction models corresponding to time attributes of the SKU to be tested; the sales prediction module 730 can be further configured to use the SKU to be tested.
  • the feature vectors are respectively input into a plurality of sales prediction models, and the sales forecast results are obtained based on the output results of the plurality of sales prediction models.
  • the sales forecasting module 730 is further configured to obtain the sales forecasting result according to the weighted sum of the output results of the plurality of sales forecasting models, and the weight corresponding to the output result of each of the sales forecasting models is determined according to the forecasting error of the sales forecasting model. .
  • the plurality of sales forecasting models may include a first sales forecasting model and a second sales forecasting model; the sales forecasting module 730 may be further configured to obtain the sales forecasting result P using the following formula:
  • a is the prediction error of the first sales forecasting model
  • b is the prediction error of the second sales forecasting model
  • g1 is the output value of the first sales forecasting model
  • g2 is the output value of the second sales forecasting model.
  • the sales forecasting device 70 may further include a sales forecasting model training module 750 for acquiring SKU feature vectors for training and corresponding tag values having the same time attribute, wherein the tag value is a corresponding SKU Sales volume; the SKU feature vector used for training is input into the model to be trained; the parameters of the model to be trained are adjusted according to the difference between the output result of the model to be trained and the tag value of the corresponding SKU feature vector, and a sales prediction model is obtained.
  • a sales forecasting model training module 750 for acquiring SKU feature vectors for training and corresponding tag values having the same time attribute, wherein the tag value is a corresponding SKU Sales volume; the SKU feature vector used for training is input into the model to be trained; the parameters of the model to be trained are adjusted according to the difference between the output result of the model to be trained and the tag value of the corresponding SKU feature vector, and a sales prediction model is obtained.
  • the feature vector of the SKU to be tested may include one or more of the following characteristics: historical sales volume for each time period of the SKU, historical average sales volume for each time period of the SKU, and sales volume of the SKU for each time unit in the history Sales of SKUs in the same period last year, promotion information related to SKUs, and product attributes of SKUs.
  • the sales forecasting device 70 may further include a planned sales disassembly module 760, configured to obtain a predicted sales ratio of each SKU to be tested according to the sales forecast results of the plurality of SKUs to be tested; The predicted sales volume of each SKU to be tested, and the sales forecast result of each SKU to be tested is updated.
  • a planned sales disassembly module 760 configured to obtain a predicted sales ratio of each SKU to be tested according to the sales forecast results of the plurality of SKUs to be tested; The predicted sales volume of each SKU to be tested, and the sales forecast result of each SKU to be tested is updated.
  • FIG. 8 is a schematic structural diagram of a sales forecasting apparatus according to further embodiments of the present disclosure.
  • the sales forecasting apparatus 800 of this embodiment includes a memory 810 and a processor 820 coupled to the memory 810, the processor 820 being configured to perform any of the foregoing implementations based on instructions stored in the memory 810.
  • the sales forecast method in the example is a schematic structural diagram of a sales forecasting apparatus according to further embodiments of the present disclosure.
  • the sales forecasting apparatus 800 of this embodiment includes a memory 810 and a processor 820 coupled to the memory 810, the processor 820 being configured to perform any of the foregoing implementations based on instructions stored in the memory 810.
  • the sales forecast method in the example is a schematic structural diagram of a sales forecasting apparatus according to further embodiments of the present disclosure.
  • the sales forecasting apparatus 800 of this embodiment includes a memory 810 and a processor 820 coupled to the memory 810, the processor 820 being configured to perform any of the foregoing implementations based on instructions stored in
  • the memory 810 may include, for example, a system memory, a fixed non-volatile storage medium, or the like.
  • the system memory stores, for example, an operating system, an application, a boot loader, and other programs.
  • FIG. 9 is a schematic structural diagram of a sales forecasting apparatus according to still another embodiment of the present disclosure.
  • the sales forecasting apparatus 900 of this embodiment includes a memory 910 and a processor 920, and may further include an input/output interface 930, a network interface 940, a storage interface 950, and the like. These interfaces 930, 940, 950 and between the memory 910 and the processor 920 can be connected, for example, via a bus 960.
  • the input/output interface 930 provides a connection interface for input and output devices such as a display, a mouse, a keyboard, and a touch screen.
  • Network interface 940 provides a connection interface for various networked devices.
  • the storage interface 950 provides a connection interface for an external storage device such as an SD card or a USB flash drive.
  • Embodiments of the present disclosure also provide a computer readable storage medium having stored thereon a computer program, wherein the program is executed by a processor to implement any of the foregoing sales forecasting methods.
  • embodiments of the present disclosure can be provided as a method, system, or computer program product. Accordingly, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or a combination of software and hardware aspects. Moreover, the present disclosure may take the form of a computer program product embodied on one or more computer-usable non-transitory storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. .
  • the computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device.
  • the apparatus implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on a computer or other programmable device to produce computer-implemented processing for execution on a computer or other programmable device.
  • the instructions provide steps for implementing the functions specified in one or more of the flow or in a block or blocks of a flow diagram.

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种销量预测方法、装置和计算机可读存储介质，涉及数据处理领域。销量预测方法包括：根据待测库存量单位SKU在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性（S102）；获取待测SKU的特征向量以及待测SKU的时间属性对应的销量预测模型，其中，待测SKU的特征向量包括待测SKU的历史数据和预测时间信息（S104）；将待测SKU的特征向量输入到预测模型中，获得销量预测结果（S106）。从而，可以使预测结果符合待测SKU销量在不同时间的分布规律，提高了销量预测的准确率，降低了人工成本，进而提升了整个电子商务平台的运行效率。

Description

销量预测方法、装置和计算机可读存储介质
本申请是以CN申请号为201711291842.7，申请日为2017年12月8日的申请为基础，并主张其优先权，该CN申请的公开内容在此作为整体引入本申请中。
技术领域
本公开涉及数据处理领域，特别涉及一种销量预测方法、装置和计算机可读存储介质。
背景技术
在电子商务技术中，销售计划作为销售目标的执行方案，向上承接公司战略计划的落地，向下指导供应商、仓储、物流的运营计划，起到了核心枢纽的作用。销售计划需要跟随市场动态等因素做出及时的调整。目前，销售计划的制定主要通过采销人员凭经验制定。
发明内容
发明人认识到，通过采销人员凭经验制定销售计划的方式无法对销售计划进行精细化管理，造成了销量预测的准确性低、销售计划与实际的销售情况不匹配，进而影响了电子商务***的整体运行效率。
本公开实施例所要解决的一个技术问题是：如何提升销量预测的准确性。
根据本公开一些实施例的第一个方面，提供一种销量预测方法，包括：根据待测库存量单位SKU在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性；获取待测SKU的特征向量以及待测SKU的时间属性对应的销量预测模型，其中，待测SKU的特征向量包括待测SKU的历史数据和预测时间信息；将待测SKU的特征向量输入到预测模型中，获得销量预测结果。
在一些实施例中，根据待测SKU在不同时间单位的销量信息的分布情况，确定待测SKU对应的时间属性包括：为包括待测SKU的多个SKU生成趋势向量，其中，每个趋势向量包括相应的SKU在每个时间单位的销量信息；对多个SKU对应的趋势向量进行聚类，根据聚类结果中待测SKU所属类别确定待测SKU的时间属性。
在一些实施例中，时间属性为季节属性。
在一些实施例中，在待测SKU的销量信息涉及的时间单位数量小于预设值的情况下，销量预测方法还包括：在具有同一时间属性的SKU的多个商品属性中，将占比超过预设值的商品属性确定为时间属性对应的商品属性；将待测SKU的商品属性与不同的时间属性对应的商品属性进行匹配；根据匹配结果确定待测SKU的时间属性。
在一些实施例中，获取待测SKU的特征向量以及待测SKU的时间属性对应的多个销量预测模型；将待测SKU的特征向量分别输入到多个销量预测模型中，根据多个销量预测模型的输出结果获得销量预测结果。
在一些实施例中，根据多个销量预测模型的输出结果的加权和获得销量预测结果，每个销量预测模型的输出结果所对应的权重根据销量预测模型的预测误差确定。
在一些实施例中，多个销量预测模型包括第一销量预测模型和第二销量预测模型；采用以下公式获得销量预测结果P：
Figure PCTCN2018115652-appb-000001
其中，a为第一销量预测模型的预测误差，b为第二销量预测模型的预测误差，g1为第一销量预测模型的输出值，g2为第二销量预测模型的输出值。
在一些实施例中，销量预测方法还包括：获取具有相同时间属性的、用于训练的SKU特征向量以及相应的标记值，其中，标记值为相应SKU的销量；将用于训练的SKU特征向量输入到待训练模型中；根据待训练模型的输出结果与相应SKU特征向量的标记值的差距调整待训练模型的参数，获得销量预测模型。
在一些实施例中，待测SKU的特征向量包括以下一种或多种特征：SKU的每个时段的历史销量、SKU的每个时段的历史平均销量、SKU在历史每个时间单位的销量、SKU在去年同期的销量、SKU涉及的促销信息、SKU的商品属性。
在一些实施例中，销量预测方法还包括：根据多个待测SKU的销量预测结果，获得每个待测SKU的预测销量比例；根据计划销量总量和每个待测SKU的预测销量比例，更新每个待测SKU的销量预测结果。
根据本公开一些实施例的第二个方面，提供一种销量预测装置，包括：时间属性确定模块，用于根据待测库存量单位SKU在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性；特征向量和模型获取模块，用于获取待测SKU的特征向量以及待测SKU的时间属性对应的销量预测模型，其中，待测SKU的特征向量包括 待测SKU的历史数据和预测时间信息；销量预测模块，用于将待测SKU的特征向量输入到预测模型中，获得销量预测结果。
在一些实施例中，时间属性确定模块进一步用于为包括待测SKU的多个SKU生成趋势向量，其中，每个趋势向量包括相应的SKU在每个时间单位的销量信息；对多个SKU对应的趋势向量进行聚类，根据聚类结果中待测SKU所属类别确定待测SKU的时间属性。
在一些实施例中，时间属性为季节属性。
在一些实施例中，销量预测装置还包括：新品时间属性确定模块，用于在待测SKU的销量信息涉及的时间单位数量小于预设值的情况下，在具有同一时间属性的SKU的多个商品属性中，将占比超过预设值的商品属性确定为时间属性对应的商品属性；将待测SKU的商品属性与不同的时间属性对应的商品属性进行匹配；根据匹配结果确定待测SKU的时间属性。
在一些实施例中，特征向量和模型获取模块用于获取待测SKU的特征向量以及待测SKU的时间属性对应的多个销量预测模型；销量预测模块进一步用于将待测SKU的特征向量分别输入到多个销量预测模型中，根据多个销量预测模型的输出结果获得销量预测结果。
在一些实施例中，销量预测模块进一步用于根据多个销量预测模型的输出结果的加权和获得销量预测结果，每个销量预测模型的输出结果所对应的权重根据销量预测模型的预测误差确定。
在一些实施例中，多个销量预测模型包括第一销量预测模型和第二销量预测模型；销量预测模块进一步用于采用以下公式获得销量预测结果P：
Figure PCTCN2018115652-appb-000002
其中，a为第一销量预测模型的预测误差，b为第二销量预测模型的预测误差，g1为第一销量预测模型的输出值，g2为第二销量预测模型的输出值。
在一些实施例中，销量预测装置还包括销量预测模型训练模块，用于获取具有相同时间属性的、用于训练的SKU特征向量以及相应的标记值，其中，标记值为相应SKU的销量；将用于训练的SKU特征向量输入到待训练模型中；根据待训练模型的输出结果与相应SKU特征向量的标记值的差距调整待训练模型的参数，获得销量预测模型。
在一些实施例中，待测SKU的特征向量包括以下一种或多种特征：SKU的每个时段的历史销量、SKU的每个时段的历史平均销量、SKU在历史每个时间单位的销量、SKU在去年同期的销量、SKU涉及的促销信息、SKU的商品属性。
在一些实施例中，销量预测装置还包括计划销量拆解模块，用于根据多个待测SKU的销量预测结果，获得每个待测SKU的预测销量比例；根据计划销量总量和每个待测SKU的预测销量比例，更新每个待测SKU的销量预测结果。
根据本公开一些实施例的第三个方面，提供一种销量预测装置，包括：存储器；以及耦接至所述存储器的处理器，所述处理器被配置为基于存储在所述存储器中的指令，执行前述任意一种销量预测方法。
根据本公开一些实施例的第四个方面，提供一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现前述任意一种销量预测方法。
上述公开中的一些实施例具有如下优点或有益效果：通过根据待测SKU的时间属性确定销量预测模型、并采用确定的销量预测模型进行自动销量预测，可以使预测结果符合待测SKU销量在不同时间的分布规律，提高了销量预测的准确率，降低了人工成本，进而提升了整个电子商务平台的运行效率。
通过以下参照附图对本公开的示例性实施例的详细描述，本公开的其它特征及其优点将会变得清楚。
附图说明
为了更清楚地说明本公开实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本公开的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为根据本公开一些实施例的销量预测方法的流程示意图。
图2为根据本公开一些实施例的SKU的时间属性确定方法的流程示意图。
图3为根据本公开另一些实施例的SKU的时间属性确定方法的流程示意图。
图4为根据本公开一些实施例的采用多个模型进行销量预测的方法的流程示意图。
图5为根据本公开一些实施例的销量预测模型训练方法的流程示意图。
图6为根据本公开另一些实施例的销量预测方法的流程示意图。
图7为根据本公开一些实施例的销量预测装置的结构示意图。
图8为根据本公开另一些实施例的销量预测装置的结构示意图。
图9为根据本公开又一些实施例的销量预测装置的结构示意图。
具体实施方式
下面将结合本公开实施例中的附图，对本公开实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本公开一部分实施例，而不是全部的实施例。以下对至少一个示例性实施例的描述实际上仅仅是说明性的，决不作为对本公开及其应用或使用的任何限制。基于本公开中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本公开保护的范围。
除非另外具体说明，否则在这些实施例中阐述的部件和步骤的相对布置、数字表达式和数值不限制本公开的范围。
同时，应当明白，为了便于描述，附图中所示出的各个部分的尺寸并不是按照实际的比例关系绘制的。
对于相关领域普通技术人员已知的技术、方法和设备可能不作详细讨论，但在适当情况下，所述技术、方法和设备应当被视为授权说明书的一部分。
在这里示出和讨论的所有示例中，任何具体值应被解释为仅仅是示例性的，而不是作为限制。因此，示例性实施例的其它示例可以具有不同的值。
应注意到：相似的标号和字母在下面的附图中表示类似项，因此，一旦某一项在一个附图中被定义，则在随后的附图中不需要对其进行进一步讨论。
图1为根据本公开一些实施例的销量预测方法的流程示意图。如图1所示，该实施例的销量预测方法包括步骤S102～S106。
在步骤S102中，根据待测SKU(Stock Keeping Unit，库存量单位)在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性。
时间单位可以根据需求设定，例如，可以令每个季节为一个时间单位，也可以令每个月、或每两个月为一个时间单位等等。可以通过时间单位体现SKU的销售趋势。例如，电热毯类的SKU在冬季热销；蚊帐、蚊香类的SKU在夏季热销；大闸蟹类SKU在九月、十月左右热销；酒类产品虽然在全年都有大量用户购买，但是在春节前后的一月、二月时，由于家庭聚餐、聚会较多，因此酒类产品的销量会在这段时间达到峰值。
在步骤S104中，获取待测SKU的特征向量以及待测SKU的时间属性对应的销 量预测模型，其中，待测SKU的特征向量包括待测SKU的历史数据和预测时间信息。
本公开在进行预测时，并不是将所有待测SKU都通过相同的模型进行预测，而是按照时间属性预先训练不同的销量预测模型。即，在训练时，采用具有同一时间属性的训练数据对模型进行训练。从而在预测时，可以获得更准确的预测结果。
销量预测模型例如可以为梯度提升决策树(Gradient Boosting Decision Tree，简称：GBDT)模型、线性回归模型、长短记忆网络(Long Short-Term Memory，简称：LSTM)模型等等。
在待测SKU的特征向量中，待测SKU的历史数据反映了待测SKU在过去一段时间的销售情况，这些历史数据可以是SKU本身的数据、也可以是该SKU所属品类的数据。在一些实施例中，待测SKU的特征向量可以包括销量的同比增长信息，例如待测SKU在近一段时间内每个月的实际成交总额(Gross Merchandise Volume，简称：GMV)、或者月均GMV；在一些实施例中，待测SKU的特征向量可以包括销量的环比增长信息，例如待测SKU所属的品类在去年同期的GMV，或者去年同期的一段时间内的GMV均值、方差等统计信息。
预测时间信息可以为待测销量所对应的时间，例如预测今年冬季某SKU的销量，则冬季可以为预测时间信息；或者预测今年12月某SKU的销量，则12月可以为预测时间信息。除了上述两种示例性的方式以外，预测时间信息可以采用多种形式表示，例如日期、节假日标记等等。如果是日期形式，既可以是阳历日期、也可以是阴历日期。
待测SKU的特征向量还可以包括SKU涉及的促销信息。例如该SKU参与的促销类别、促销时长、促销覆盖的SKU数量、促销中的热销品所占比例等等。而SKU在友商平台中的促销信息也可以添加到特征向量中。
待测SKU的特征向量还可以包括SKU自身的商品属性，例如规格、尺寸、颜色等等。
在步骤S106中，将待测SKU的特征向量输入到预测模型中，获得销量预测结果。
通过上述实施例的方法，可以根据待测SKU的时间属性确定销量预测模型、并采用确定的销量预测模型进行自动销量预测，从而可以使预测结果符合待测SKU销量在不同时间的分布规律，提高了销量预测的准确率，降低了人工成本，进而提升了整个电子商务平台的运行效率。
在一些实施例中，可以通过聚类的方法确定SKU的时间属性。下面参考图2描 述本公开SKU的时间属性的确定方法的实施例。
图2为根据本公开一些实施例的SKU的时间属性确定方法的流程示意图。如图2所示，该实施例的SKU的时间属性确定方法包括步骤S202～S204。
在步骤S202中，为多个SKU生成趋势向量，其中，每个趋势向量包括相应的SKU在每个时间单位的销量信息。
时间单位的大小可以根据需要来设定。例如，可以以月份为单位进行统计。如果某个SKU在1～12月份的销量分别为1、0.8、0.95、0.98、10.23、15.6、18.9、19.7、10.6、2.5、1、1.2，单位为万元，则该SKU的趋势向量为[1,0.8,0.95,0.98,10.23,15.6,18.9,19.7,10.6,2.5,1,1.2]。
在步骤S204中，对多个SKU对应的趋势向量进行聚类，根据聚类结果中待测SKU所属类别确定待测SKU的时间属性。
例如，如果时间属性为季节属性，那么分类数量可以为4，即每个类别分别对应春、夏、秋、冬四个季节；分类数量也可以为5，即除了四季以外还包括“全年热销”这样一个类别，以将销量不随时间的变化而有显著改变的SKU挑选出来。
通过上述实施例的方法，可以自动确定各个SKU的时间属性，从而获得了其中的待测SKU的时间属性。上述实施例的方法能够准确、客观地识别SKU的销量随时间变化的趋势和规律，从而可以准确地识别出SKU的时间属性。
在进行聚类时，SKU的销售时间需要达到一定程度才可以参与计算，即SKU涉及的时间单位数量需要大于预设值。例如，已销售一年以上的SKU具有一年中各个时段的数据；而对于一些新品类或新上线的SKU，其销售数据不足以覆盖一整年，因此还可以采用以下实施例的方法确定时间属性。下面参考图3描述本公开SKU的时间属性的确定方法的实施例。
图3为根据本公开另一些实施例的SKU的时间属性确定方法的流程示意图。如图3所示，该实施例的SKU的时间属性确定方法包括步骤S302～S306。
在步骤S302中，在具有同一时间属性的SKU的多个商品属性中，将占比超过预设值的商品属性确定为时间属性对应的商品属性。
在步骤S304中，将待测SKU的商品属性与不同的时间属性对应的商品属性进行匹配。
在步骤S306中，根据匹配结果确定待测SKU的时间属性。
例如，已确定具有时间属性X的SKU1具有商品属性A、B、C，SKU2具有商品 属性A、C、D，SKU3具有商品属性A、B、E，SKU2具有商品属性A、C、F等等。经过统计，具有时间属性X的SKU中，商品属性A和C的占比最高，例如具有冬季属性的SKU中，商品属性“加厚”、“加绒”的商品属性占比最高，因此具有商品属性A和C的SKU极有可能为具有时间属性X的SKU。
从而，对于销售数据量覆盖时间长度小于预设值的新品SKU也可以确定其时间属性，进一步提高了本公开的适用场景。当该新品SKU的销售时长达到预设值后，还可以采用图2实施例或类似的方法重新验证其时间属性是否合理。
本公开的实施例在进行预测时可以基于一个模型的输出结果确定销量预测结果，也可以根据多个模型的输出结果来确定。下面参考图4描述本公开采用多个模型进行销量预测的实施例。
图4为根据本公开一些实施例的采用多个模型进行销量预测的方法的流程示意图。如图4所示，该实施例的销量预测模型训练方法包括步骤S402～S404。
在步骤S402中，获取待测SKU的特征向量以及待测SKU的时间属性对应的多个销量预测模型。例如，待测SKU为冬季热销品，那么可以获取GBDT冬季模型以及LSTM冬季模型。
在步骤S404中，将待测SKU的特征向量分别输入到多个销量预测模型中，根据多个销量预测模型的输出结果获得销量预测结果。
例如，可以根据多个销量预测模型的输出结果的加权结果确定销量预测结果。通过上述实施例的方法，可以综合多个模型的输出结果，提高销量预测的准确性。
在一些实施例中，可以根据多个销量预测模型的输出结果的加权和获得销量预测结果，每个销量预测模型的输出结果所对应的权重根据销量预测模型的预测误差确定。下面示例性地提供一种预测结果的获得方法。
设基于两个销量预测模型进行预测，这两个销量预测模型分别为第一销量预测模型和第二销量预测模型。可以采用公式(1)获得销量预测结果P：
Figure PCTCN2018115652-appb-000003
在公式(1)中，a为第一销量预测模型的预测误差，b为第二销量预测模型的预测误差，预测误差可以根据训练阶段输出结果与标记值的差距确定；g1为第一销量预测模型的输出值，g2为第二销量预测模型的输出值。从而，可以将模型的误差大小作为权值以综合考虑多个模型的输出结果，进一步提升销量预测的准确性。
下面参考图5描述本公开销量预测模型训练方法的实施例。
图5为根据本公开一些实施例的销量预测模型训练方法的流程示意图。如图5所示，该实施例的销量预测模型训练方法包括步骤S502～S506。
在步骤S502中，获取具有相同时间属性的、用于训练的SKU特征向量以及相应的标记值，其中，标记值为相应SKU的销量。
训练阶段所采用的特征向量与预测阶段所采用的特征向量的组成元素相同，这里不再赘述。由于训练的目的是获得某种时间属性对应的专有模型，因此训练数据也应当具有相同的时间属性。
在步骤S504中，将用于训练的SKU特征向量输入到待训练模型中。
在步骤S506中，根据待训练模型的输出结果与相应SKU特征向量的标记值的差距调整待训练模型的参数，获得销量预测模型。
模型的具体训练方法可以参考现有技术中的训练方法，这里不再赘述。本领域技术人员可以根据需要选择待训练的模型。下面示例性地对几种模型进行简要的介绍。
GBDT是基于提升(Boosting)集成方法的梯度下降算法，核心是思想是串行建立多棵决策树，每一棵树都是用来纠正前面已有树的错误。GBDT是目前应用比较广泛、效果较好的一种机器学习算法。
线性回归是利用称为线性回归方程的最小二乘函数对一个或多个自变量和因变量之间关系进行建模的一种回归分析。这种函数是一个或多个称为回归系数的模型参数的线性组合。只有一个自变量的情况称为简单回归，大于一个自变量情况的叫做多元回归。
LSTM是一种递归神经网络的特殊类型，可以学习长期依赖信息。LSTM通过刻意的设计来避免长期依赖问题。
在训练阶段，除了按照时间属性划分不同的模型以外，还可以对模型进行进一步细分。例如，可以对有促销资源的训练数据单独进行训练，有促销资源的情况例如可以为首页推送的SKU、具有活动专栏的SKU、促销力度较大的SKU等等；此外，也可以对新品类的SKU的训练数据进行单独训练等等。下面参考图6描述本公开销量预测方法的实施例。
图6为根据本公开另一些实施例的销量预测方法的流程示意图。如图6所示，该实施例的销量预测方法包括步骤S602～S620。
在步骤S602中，根据待测库存量单位SKU在不同时间单位的销量信息的分布情 况，确定待测SKU的时间属性。
在步骤S604中，获取待测SKU的特征向量。
在步骤S606中，判断待测SKU是否有促销资源。如果是，执行步骤S608；如果不是，执行步骤S610。
在步骤S608中，采用待测SKU的时间属性对应的促销模型预测待测SKU的销量。促销模型为根据具有促销资源的训练数据进行训练而得到的销量预测模型。
在步骤S610中，判断待测SKU是否为新品。如果是，执行步骤S612；如果不是，执行步骤S614。
在步骤S612中，采用待测SKU的时间属性对应的新品模型预测待测SKU的销量。新品模型为根据新品对应的训练数据进行训练而得到的销量预测模型。
在步骤S614中，采用待测SKU的时间属性对应的普通模型预测待测SKU的销量。普通模型为根据非新品、非促销品对应的训练数据进行训练而得到的销量预测模型。
在步骤S616中，将预测结果加载到缓存中，以便于用户根据需要进行快速读取。
在步骤S618中，根据预测结果生成销售计划报表，例如可以根据待测SKU的预计销量生成以采销员、品类、品牌、部门为单位的总体计划。
在一些实施例中，可以根据多个待测SKU的销量预测结果，获得每个待测SKU的预测销量比例；然后再根据计划销量总量和每个待测SKU的预测销量比例，更新每个待测SKU的销量预测结果。从而，当采销员或者部分获得了生产任务之后，可以根据SKU的预测结果来对计划进行拆解，从而使预测结果更能够符合实际需求。
在步骤S620中，对数据进行持久化存储，例如可以通过HDFS(Hadoop Distributed File System，分布式文件存储***)保存。
通过上述实施例的方法，能够自动化地针对待测SKU的具体类型进行精准的销量预测，并且能够将数据进行缓存以及持久化的存储，为后续的供货、仓储、物流、运营等环节提供了极大的便利。
下面参考图7描述本公开销量预测装置的实施例。
图7为根据本公开一些实施例的销量预测装置的结构示意图。如图7所示，该实施例的销量预测装置70包括：时间属性确定模块710，用于根据待测库存量单位SKU在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性；特征向量和模型获取模块720，用于获取待测SKU的特征向量以及待测SKU的时间属性对应的销 量预测模型，其中，待测SKU的特征向量包括待测SKU的历史数据和预测时间信息；销量预测模块730，用于将待测SKU的特征向量输入到预测模型中，获得销量预测结果。
在一些实施例中，时间属性确定模块710可以进一步用于为包括待测SKU的多个SKU生成趋势向量，其中，每个趋势向量包括相应的SKU在每个时间单位的销量信息；对多个SKU对应的趋势向量进行聚类，根据聚类结果中待测SKU所属类别确定待测SKU的时间属性。
在一些实施例中，时间属性为季节属性。
在一些实施例中，销量预测装置70还可以包括：新品时间属性确定模块740，用于在待测SKU的销量信息涉及的时间单位数量小于预设值的情况下，在具有同一时间属性的SKU的多个商品属性中，将占比超过预设值的商品属性确定为时间属性对应的商品属性；将待测SKU的商品属性与不同的时间属性对应的商品属性进行匹配；根据匹配结果确定待测SKU的时间属性。
在一些实施例中，特征向量和模型获取模块720可以用于获取待测SKU的特征向量以及待测SKU的时间属性对应的多个销量预测模型；销量预测模块730可以进一步用于将待测SKU的特征向量分别输入到多个销量预测模型中，根据多个销量预测模型的输出结果获得销量预测结果。
在一些实施例中，销量预测模块730进一步用于根据多个销量预测模型的输出结果的加权和获得销量预测结果，每个销量预测模型的输出结果所对应的权重根据销量预测模型的预测误差确定。
在一些实施例中，多个销量预测模型可以包括第一销量预测模型和第二销量预测模型；销量预测模块730可以进一步用于采用以下公式获得销量预测结果P：
Figure PCTCN2018115652-appb-000004
其中，a为第一销量预测模型的预测误差，b为第二销量预测模型的预测误差，g1为第一销量预测模型的输出值，g2为第二销量预测模型的输出值。
在一些实施例中，销量预测装置70还可以包括销量预测模型训练模块750，用于获取具有相同时间属性的、用于训练的SKU特征向量以及相应的标记值，其中，标记值为相应SKU的销量；将用于训练的SKU特征向量输入到待训练模型中；根据待训练模型的输出结果与相应SKU特征向量的标记值的差距调整待训练模型的参数， 获得销量预测模型。
在一些实施例中，待测SKU的特征向量可以包括以下一种或多种特征：SKU的每个时段的历史销量、SKU的每个时段的历史平均销量、SKU在历史每个时间单位的销量、SKU在去年同期的销量、SKU涉及的促销信息、SKU的商品属性。
在一些实施例中，销量预测装置70还可以包括计划销量拆解模块760，用于根据多个待测SKU的销量预测结果，获得每个待测SKU的预测销量比例；根据计划销量总量和每个待测SKU的预测销量比例，更新每个待测SKU的销量预测结果。
图8为根据本公开另一些实施例的销量预测装置的结构示意图。如图8所示，该实施例的销量预测装置800包括：存储器810以及耦接至该存储器810的处理器820，处理器820被配置为基于存储在存储器810中的指令，执行前述任意一个实施例中的销量预测方法。
其中，存储器810例如可以包括***存储器、固定非易失性存储介质等。***存储器例如存储有操作***、应用程序、引导装载程序(Boot Loader)以及其他程序等。
图9为根据本公开又一些实施例的销量预测装置的结构示意图。如图9所示，该实施例的销量预测装置900包括：存储器910以及处理器920，还可以包括输入输出接口930、网络接口940、存储接口950等。这些接口930，940，950以及存储器910和处理器920之间例如可以通过总线960连接。其中，输入输出接口930为显示器、鼠标、键盘、触摸屏等输入输出设备提供连接接口。网络接口940为各种联网设备提供连接接口。存储接口950为SD卡、U盘等外置存储设备提供连接接口。
本公开的实施例还提供一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现前述任意一种销量预测方法。
本领域内的技术人员应当明白，本公开的实施例可提供为方法、***、或计算机程序产品。因此，本公开可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本公开可采用在一个或多个其中包含有计算机可用程序代码的计算机可用非瞬时性存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本公开是参照根据本公开实施例的方法、设备(***)、和计算机程序产品的流程图和/或方框图来描述的。应理解为可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数 据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
以上所述仅为本公开的较佳实施例，并不用以限制本公开，凡在本公开的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本公开的保护范围之内。

Claims (20)

  1. 一种销量预测方法，包括：
    根据待测库存量单位SKU在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性；
    获取待测SKU的特征向量以及所述待测SKU的时间属性对应的销量预测模型，其中，所述待测SKU的特征向量包括待测SKU的历史数据和预测时间信息；
    将所述待测SKU的特征向量输入到所述预测模型中，获得销量预测结果。
  2. 根据权利要求1所述的销量预测方法，其中，所述根据待测SKU在不同时间单位的销量信息的分布情况，确定待测SKU对应的时间属性包括：
    为包括待测SKU的多个SKU生成趋势向量，其中，每个趋势向量包括相应的SKU在每个时间单位的销量信息；
    对所述多个SKU对应的趋势向量进行聚类，根据聚类结果中待测SKU所属类别确定待测SKU的时间属性。
  3. 根据权利要求1或2所述的销量预测方法，其中，所述时间属性为季节属性。
  4. 根据权利要求1所述的销量预测方法，其中，在待测SKU的销量信息涉及的时间单位数量小于预设值的情况下，所述销量预测方法还包括：
    在具有同一时间属性的SKU的多个商品属性中，将占比超过预设值的商品属性确定为所述时间属性对应的商品属性；
    将待测SKU的商品属性与不同的时间属性对应的商品属性进行匹配；
    根据匹配结果确定待测SKU的时间属性。
  5. 根据权利要求1所述的销量预测方法，其中，
    获取待测SKU的特征向量以及所述待测SKU的时间属性对应的多个销量预测模型；
    将所述待测SKU的特征向量分别输入到多个销量预测模型中，根据多个销量预测模型的输出结果获得销量预测结果。
  6. 根据权利要求5所述的销量预测方法，其中，根据多个销量预测模型的输出结果的加权和获得销量预测结果，每个销量预测模型的输出结果所对应的权重根据销量预测模型的预测误差确定。
  7. 根据权利要求1所述的销量预测方法，还包括：
    获取具有相同时间属性的、用于训练的SKU特征向量以及相应的标记值，其中，所述标记值为相应SKU的销量；
    将所述用于训练的SKU特征向量输入到待训练模型中；
    根据所述待训练模型的输出结果与相应SKU特征向量的标记值的差距调整所述待训练模型的参数，获得销量预测模型。
  8. 根据权利要求1所述的销量预测方法，其中，所述待测SKU的特征向量包括以下一种或多种特征：
    SKU的每个时段的历史销量、SKU的每个时段的历史平均销量、SKU在历史每个时间单位的销量、SKU在去年同期的销量、SKU涉及的促销信息、SKU的商品属性。
  9. 根据权利要求1所述的销量预测方法，还包括：
    根据多个待测SKU的销量预测结果，获得每个待测SKU的预测销量比例；
    根据计划销量总量和每个待测SKU的预测销量比例，更新每个待测SKU的销量预测结果。
  10. 一种销量预测装置，包括：
    时间属性确定模块，用于根据待测库存量单位SKU在不同时间单位的销量信息的分布情况，确定待测SKU的时间属性；
    特征向量和模型获取模块，用于获取待测SKU的特征向量以及所述待测SKU的时间属性对应的销量预测模型，其中，所述待测SKU的特征向量包括待测SKU的历史数据和预测时间信息；
    销量预测模块，用于将所述待测SKU的特征向量输入到所述预测模型中，获得销量预测结果。
  11. 根据权利要求10所述的销量预测装置，其中，所述时间属性确定模块进一步用于为包括待测SKU的多个SKU生成趋势向量，其中，每个趋势向量包括相应的SKU在每个时间单位的销量信息；对所述多个SKU对应的趋势向量进行聚类，根据聚类结果中待测SKU所属类别确定待测SKU的时间属性。
  12. 根据权利要求10或11所述的销量预测装置，其中，所述时间属性为季节属性。
  13. 根据权利要求10所述的销量预测装置，还包括：
    新品时间属性确定模块，用于在待测SKU的销量信息涉及的时间单位数量小于 预设值的情况下，在具有同一时间属性的SKU的多个商品属性中，将占比超过预设值的商品属性确定为所述时间属性对应的商品属性；将待测SKU的商品属性与不同的时间属性对应的商品属性进行匹配；根据匹配结果确定待测SKU的时间属性。
  14. 根据权利要求10所述的销量预测装置，其中，
    所述特征向量和模型获取模块用于获取待测SKU的特征向量以及所述待测SKU的时间属性对应的多个销量预测模型；
    所述销量预测模块进一步用于将所述待测SKU的特征向量分别输入到多个销量预测模型中，根据多个销量预测模型的输出结果获得销量预测结果。
  15. 根据权利要求14所述的销量预测装置，其中，所述销量预测模块进一步用于其中，根据多个销量预测模型的输出结果的加权和获得销量预测结果，每个销量预测模型的输出结果所对应的权重根据销量预测模型的预测误差确定
  16. 根据权利要求10所述的销量预测装置，还包括销量预测模型训练模块，用于获取具有相同时间属性的、用于训练的SKU特征向量以及相应的标记值，其中，所述标记值为相应SKU的销量；将所述用于训练的SKU特征向量输入到待训练模型中；根据所述待训练模型的输出结果与相应SKU特征向量的标记值的差距调整所述待训练模型的参数，获得销量预测模型。
  17. 根据权利要求10所述的销量预测装置，其中，所述待测SKU的特征向量包括以下一种或多种特征：
    SKU的每个时段的历史销量、SKU的每个时段的历史平均销量、SKU在历史每个时间单位的销量、SKU在去年同期的销量、SKU涉及的促销信息、SKU的商品属性。
  18. 根据权利要求10所述的销量预测装置，还包括计划销量拆解模块，用于根据多个待测SKU的销量预测结果，获得每个待测SKU的预测销量比例；根据计划销量总量和每个待测SKU的预测销量比例，更新每个待测SKU的销量预测结果。
  19. 一种销量预测装置，其中：
    存储器；以及
    耦接至所述存储器的处理器，所述处理器被配置为基于存储在所述存储器中的指令，执行如权利要求1～9中任一项所述的销量预测方法。
  20. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现权利要求1～9中任一项所述的销量预测方法。
PCT/CN2018/115652 2017-12-08 2018-11-15 销量预测方法、装置和计算机可读存储介质 WO2019109790A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711291842.7A CN109903064A (zh) 2017-12-08 2017-12-08 销量预测方法、装置和计算机可读存储介质
CN201711291842.7 2017-12-08

Publications (1)

Publication Number Publication Date
WO2019109790A1 true WO2019109790A1 (zh) 2019-06-13

Family

ID=66751301

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/115652 WO2019109790A1 (zh) 2017-12-08 2018-11-15 销量预测方法、装置和计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN109903064A (zh)
WO (1) WO2019109790A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113283936A (zh) * 2021-05-28 2021-08-20 深圳千岸科技股份有限公司 一种销量预测方法、销量预测装置及电子设备
CN113379125A (zh) * 2021-06-11 2021-09-10 武汉大学 一种基于TCN和LightGBM组合模型的物流仓储销售量预测方法

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112906925B (zh) * 2019-11-19 2024-06-18 北京沃东天骏信息技术有限公司 货物补充方法、装置和计算机可读存储介质
CN111160968A (zh) * 2019-12-27 2020-05-15 清华大学 Sku级别商品销量预测方法及装置
CN111652655A (zh) * 2020-06-10 2020-09-11 创新奇智(上海)科技有限公司 商品销量预测方法、装置、电子设备及可读存储介质
CN113971471A (zh) * 2020-07-22 2022-01-25 上海顺如丰来技术有限公司 产品信息预测模型构建方法与产品信息预测方法
CN111882039A (zh) * 2020-07-28 2020-11-03 平安科技（深圳）有限公司 物理机销售数据预测方法、装置、计算机设备及存储介质
CN112258224A (zh) * 2020-10-19 2021-01-22 北京沃东天骏信息技术有限公司 信息生成方法、装置、终端、***和存储介质
CN113487359B (zh) * 2021-07-12 2024-03-22 华润数字科技有限公司 基于多模态特征的商品销量预测方法、装置及相关设备
CN114155024A (zh) * 2021-11-30 2022-03-08 北京京东振世信息技术有限公司 确定目标物品的方法、装置、设备及介质
CN114677174A (zh) * 2022-03-25 2022-06-28 北京京东尚科信息技术有限公司 一种计算未上架物品销量的方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1423215A (zh) * 2001-11-27 2003-06-11 株式会社世界 销售预测装置及销售预测方法
CN104700152A (zh) * 2014-10-22 2015-06-10 浙江中烟工业有限责任公司 一种融合季节销售信息与搜索行为信息的烟草销量预测方法
US20160117702A1 (en) * 2014-10-24 2016-04-28 Vedavyas Chigurupati Trend-based clusters of time-dependent data
CN106570573A (zh) * 2015-10-13 2017-04-19 阿里巴巴集团控股有限公司 预测包裹属性信息的方法及装置
CN106951514A (zh) * 2017-03-17 2017-07-14 合肥工业大学 一种考虑品牌情感的汽车销量预测方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102968670B (zh) * 2012-10-23 2016-08-17 北京京东世纪贸易有限公司 预测数据的方法和装置
US20140200992A1 (en) * 2013-01-14 2014-07-17 Oracle International Corporation Retail product lagged promotional effect prediction system
CN106408341A (zh) * 2016-09-21 2017-02-15 北京小米移动软件有限公司 货品销量的预测方法及装置、电子设备
CN106971249A (zh) * 2017-05-05 2017-07-21 北京挖玖电子商务有限公司 一种销量预测及补货方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1423215A (zh) * 2001-11-27 2003-06-11 株式会社世界 销售预测装置及销售预测方法
CN104700152A (zh) * 2014-10-22 2015-06-10 浙江中烟工业有限责任公司 一种融合季节销售信息与搜索行为信息的烟草销量预测方法
US20160117702A1 (en) * 2014-10-24 2016-04-28 Vedavyas Chigurupati Trend-based clusters of time-dependent data
CN106570573A (zh) * 2015-10-13 2017-04-19 阿里巴巴集团控股有限公司 预测包裹属性信息的方法及装置
CN106951514A (zh) * 2017-03-17 2017-07-14 合肥工业大学 一种考虑品牌情感的汽车销量预测方法

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113283936A (zh) * 2021-05-28 2021-08-20 深圳千岸科技股份有限公司 一种销量预测方法、销量预测装置及电子设备
CN113379125A (zh) * 2021-06-11 2021-09-10 武汉大学 一种基于TCN和LightGBM组合模型的物流仓储销售量预测方法
CN113379125B (zh) * 2021-06-11 2022-05-13 武汉大学 一种基于TCN和LightGBM组合模型的物流仓储销售量预测方法

Also Published As

Publication number Publication date
CN109903064A (zh) 2019-06-18

Similar Documents

Publication Publication Date Title
WO2019109790A1 (zh) 销量预测方法、装置和计算机可读存储介质
US11195193B2 (en) System and method for price testing and optimization
Wu et al. House price index construction in the nascent housing market: The case of China
US8738421B1 (en) Driver moderator method for retail sales prediction
CN110009400A (zh) 商品定价方法、终端及计算机可读存储介质
US11727420B2 (en) Time series clustering analysis for forecasting demand
Salehi et al. An EOQ model with random disruption and partial backordering
CN110555578B (zh) 一种销量预测方法和装置
CN107465741A (zh) 信息推送方法和装置
US10896218B2 (en) Computerized geo-referencing for images
Demirag et al. A note on inventory policies for products with residual-life-dependent demand
US20190251609A1 (en) Commodity demand prediction system, commodity demand prediction method, and commodity demand prediction program
US20160162830A1 (en) Inventory management based on automatically generating recommendations
US20190385178A1 (en) Prediction system and prediction method
Mascle et al. Inventory management maximization based on sales forecast: case study
CN110275897A (zh) 用于基于位置的数据检索的启用信标的***
US10672053B1 (en) Systems, manufactures, and methods for comparative bid analysis and purchase order preparation
JP2020057164A (ja) 解析装置
CN115860800A (zh) 一种节假日商品销量预测方法、装置及计算机存储介质
Wen et al. Dynamic capacity management with uncertain demand and dynamic price
CN106682947A (zh) 用于共享环境的时间价值评估方法和***
JP6193817B2 (ja) 需要予測装置およびプログラム
Pal et al. A marketing-oriented inventory model with three-component demand rate dependent on displayed stock level (DSL)
US20230052352A1 (en) Systems and methods for assigning attribution weights to nodes
CN113435541B (zh) 品类规划方法、装置、存储介质及计算机设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18887047

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 18/09/2020)

122 Ep: pct application non-entry in european phase

Ref document number: 18887047

Country of ref document: EP

Kind code of ref document: A1