【作者】 刘晓宇；

【导师】 许建军；

【作者基本信息】 东北财经大学 ， 应用统计， 2023， 硕士

【摘要】 据2022年国家统计局公布数据,我国消费品零售总额比上一年有所下降,商品零售总额的下降幅度为0.1%,是其他行业零售总额下降幅度的几十分之一,这也表示了我国的商品零售有着较强的韧性。线上购物平台的飞速发展对线下店铺零售造成了一定的冲击,加之疫情的短期效应,部分线下店铺面临着入不敷出的风险。我国消费市场有着强大的消费基础,且消费规模扩大、结构升级的态势仍将延续。国家提出一系列促进消费政策成效初步显现,商品和服务消费需求将不断释放,消费市场有望继续稳步恢复持续恢复。目前国内外学者的对于零售数据的研究多基于一手获取的信息,但是零售数据除零售额及产品价格等直观因素外,难以获得更准确的其他信息,往往忽略了消费者自身以及外部环境因素。因此对于分析影响因素,利用现有信息进行数据整合及重构,构建多因素预测模型的研究具有一定的理论价值与实际意义。本文以P企业下一家百货商场的热销产品——YSL品牌口红的历史销售量作为研究对象,进行产品需求预测。探索分析企业销售现状,构建了特征变量的原始数据集,同时采用一定的方法利用已有数据转换为可度量的特征加入数据集。基于构造的数据集,利用机器学习模型进行了特征的筛选、单一模型预测及多模型融合的优化。主要工作内容如下:首先,本文整理分析了国内外对于产品需求,销售,特别是消费品需求预测的研究内容,对其中的预测方法进行总结归纳,基于已有的研究内容,寻求特征变量的构建,从微观和宏观角度进行特征变量的选择,从企业和消费者视角理解产品购买的影响因素。本文依托零售企业会员消费历史信息数据开展实证研究,使用SQL对该商场的零售数据相关表格进行处理,进行多表连接和筛选,整理展示单一产品的销售情况,分析影响其需求的多维因素。构建初始特征变量集,再运用机器学习算法进行特征的重要性评分,按照重要性进行一轮特征变量的降维。其次,对经过两次处理的特征变量集进行数据的清洗,包括重复值、空值、异常值的处理。最终特征变量10个,有效数据数量为631条。利用单一模型筛选特征变量,使用六种单一机器学习模型进行了探索性分析。在建模过程中运用网格搜索法及交叉验证法进行参数的调优,综合对比单模型拟合效果及评价指标,选择优良的模型作为初级学习器进行stacking模型融合。同时在模型融合的过程中,创新的对生成的输入变量进行干预,不直接进行stacking,将原始数据集中部分特征以及输入变量的均值添加进入,使得次级学习器学习到数据的原始结构,构造出基于三种初级学习器的销量预测模型,经改良过后的模型误差要小于初始stacking模型,且使用测试数据证明该模型的预测准确率较高。最后,利用该模型对实际数据进行分析,得到了准确性较高的预测模型,并且利用在变量集时间戳之后的子集数据即会员历史销售量进行了模型的推广验证,证实该模型在数据结构一致时可用于其他子集数据。结合研究过程发现目前存在的问题,给出提高销量的合理意见。帮助企业在获取消费者数据的且抗敏感性的数据指标范围,探索销量提升的关键因素,优化公司经营策略,以达到提高营销准确性的目的,帮助企业增强行业内竞争力、提升品牌价值及顾客满意度进而提高企业效益。更多还原

【Abstract】 According to the data released by the National Bureau of Statistics in 2022,China’s total retail sales of consumer goods decreased by 0.1%compared to the previous year,which is more than one tenth of the decline of the total retail sales of other industries,which also shows that China’s retail sales have strong resilience.The rapid development of online shopping platforms has caused a certain impact on the retail of offline stores.Coupled with the short-term effect of the epidemic,some offline stores are facing the risk of being unable to make ends meet.Chinese consumer market has a strong consumption base,and the trend of expanding consumption scale and upgrading structure will continue.The government has put forward a series of policies to promote consumption,and the consumer demand for goods and services will continue to be released,and the consumer market is expected to continue to recover steadily.At present,domestic and foreign scholars’ research on retail data is mostly based on first-hand information,but in addition to retail sales and product prices and other intuitive factors,it is difficult to obtain more accurate information in retail data,often ignoring consumers themselves and external environmental factors.Therefore,it is of certain theoretical value and practical significance to analyze the influencing factors,use the existing information to integrate and reconstruct the data,and construct the multi-factor prediction model.This paper takes the sales volume of YSL lipstick,a popular product in the department store next to P company,as the research object,analyzes the sales status of the company,constructs the original data set of characteristic variables,and uses certain methods to convert the existing data into measurable features and add them into the data set.Based on the constructed data set,feature screening,single model prediction and multi-model fusion optimization are carried out by using machine learning model.The main work contents are as follows:Firstly,This paper sorted out and analyzed the domestic and foreign research contents on product demand,sales,especially consumer demand forecast,and summarized the forecasting methods.Based on the existing research contents,it sought the construction of characteristic variables,selected characteristic variables from micro and macro perspectives,and understood the influencing factors of product purchase from the perspectives of enterprises and consumers.Based on the consumption history information data of retail enterprise members,this paper carries out empirical research.First,SQL is used to process the sales flow table,member consumption detail table and commodity information table,organize and display the sales situation of a single product,and analyze the multi-dimensional factors affecting its demand.The initial feature variable set is constructed,and then the machine learning algorithm is used to score the importance of the features,and a round of dimension reduction of the feature variables is carried out according to the importance.Secondly,the data of the characteristic variable set processed twice is cleaned,including the processing of repeated values,null values and outliers.There are 10 final characteristic variables and 631 effective data.LASSO regression was used to score the feature importance,and the feature variables were selected by coefficient and significance.Subsequently,six single models were used:LASSO regression,random forest regression,K-nearest neighbor regression,support vector machine regression,time series regression and ’short and long time memory network’ to model the data and conduct exploratory analysis.In the modeling process,mesh search method and cross validation method were used to optimize the parameters,and the fitting effect of single model and the evaluation index:MSE,RMSE,decision coefficient were compared comprehensively.Select good models as primary learning tools to integrate stacking models.At the same time,in the process of model fusion,it innovatively interferes with the generated input variables and does not directly stacking.Partial features of the original data set and the mean value of input variables are added to enable secondary learners to learn the original structure of the data and construct a sales prediction model based on three primary learners.The error of the improved model is less than the initial stacking model,and the prediction accuracy of the model is proved by test data.Finally,through empirical analysis,a prediction model with high accuracy was obtained,and the subset data after the time stamp of the variable set,namely the historical sales volume of members,was used to promote the model,confirming that the model can be used for other subset data when the data structure is consistent.Combined with the research process to find the existing problems,give reasonable suggestions to improve sales.Help enterprises to build a system of influencing factors,a reasonable range of consumer data acquisition,and explore new marketing strategies,so as to achieve the purpose of improving marketing accuracy,and strive to enhance the competitiveness of retail enterprises,enhance the brand,bring new profit growth points for retail enterprises,so as to enhance the competitiveness of enterprises.更多还原