鄭金峰,羅戎蕾
摘 要:服裝銷售預(yù)測(cè)是服裝企業(yè)商品企劃中必不可少的環(huán)節(jié)之一。為有效幫助服裝商品企劃人員及相關(guān)學(xué)者根據(jù)實(shí)際情況快速選擇合適的服裝銷售預(yù)測(cè)方法,對(duì)時(shí)間序列法、回歸分析法、灰色預(yù)測(cè)模型及人工神經(jīng)網(wǎng)絡(luò)4類定量銷售預(yù)測(cè)方法從優(yōu)缺點(diǎn)、優(yōu)化歷程及適用類型3個(gè)方面進(jìn)行梳理總結(jié),并對(duì)機(jī)器學(xué)習(xí)的部分組合算法進(jìn)行舉例與歸納。分析得出:時(shí)間序列法適用于歷史數(shù)據(jù)離散程度小且影響因素少的短中期服裝銷售預(yù)測(cè);回歸分析法中多元回歸法比一元回歸法在算理上更適合具有多因素影響的服裝銷售預(yù)測(cè);灰色預(yù)測(cè)模型適用于數(shù)據(jù)平滑且影響因素較少的服裝銷售預(yù)測(cè);人工神經(jīng)網(wǎng)絡(luò)則適合銷售數(shù)據(jù)離散程度大的時(shí)尚型服裝銷售預(yù)測(cè)。
關(guān)鍵詞:服裝銷售預(yù)測(cè);預(yù)測(cè)方法;時(shí)間序列法;回歸分析法;灰色模型;人工神經(jīng)網(wǎng)絡(luò)
中圖分類號(hào):TN941.26
文獻(xiàn)標(biāo)志碼:A
文章編號(hào):1009-265X(2022)02-0027-09
Research progress on quantitative forecast methods of clothing sales
ZHENG Jinfenga, LUO Rongleib,c
(a. School of Fashion Design & Engineering; b. School of International Education; c.Zhejiang Sickand Fashion Culture Research Center, Zhejiang Sci-Tech University, Hangzhou 310018, China)
Abstract: The forecast of clothing sales is one of the essential steps in the commodity planning of clothing enterprises. In order to effectively help garment planners and relevant scholars to choose appropriate forecast methods of clothing sales quickly as the case may be, this study summarizes the advantages and disadvantages, optimization process and application types of 4 kinds of quantitative sales forecast methods, including time series method, regression analysis method, grey prediction model and artificial neural network, illustrates and sums up some combined algorithms of machine learning. The results show that the time series method is suitable for short-and medium-term forecast of clothing sales with small discrete degree of historical data and few influence factors; in the regression analysis, multiple regression method is more suitable for the forecast of clothing sales with multiple influence factors than single regression method in computational theory; grey prediction model is suitable for the forecast of clothing sales with smooth data and few influence factors, while the artificial neural network is suitable for the forecast of sales of fashionable garments with highly discrete sales data.
Key words: forecast of clothing sales; forecast method; time series method; regression analysis method; grey model; artificial neural network
早在古代,人類就開(kāi)始根據(jù)自然現(xiàn)象預(yù)測(cè)天氣變化,由自然萬(wàn)物之間的聯(lián)系所建立的《周易》可以預(yù)測(cè)短期、中長(zhǎng)期及超長(zhǎng)期的天氣。直至20世紀(jì)60年代,預(yù)測(cè)才形成一門獨(dú)立的學(xué)科。預(yù)測(cè)學(xué)是一種運(yùn)用因果原理、慣性原理及相關(guān)原理研究關(guān)于預(yù)測(cè)的理論、方法、應(yīng)用及評(píng)價(jià)的綜合型數(shù)學(xué)學(xué)科,其核心是預(yù)測(cè)方法[1]。據(jù)有關(guān)資料統(tǒng)計(jì),被廣泛接受認(rèn)可的預(yù)測(cè)方法多達(dá)300多種[2]。因服裝具有趨勢(shì)性、季節(jié)性、周期性和隨機(jī)性,服裝銷售受天氣、地域及人為等因素影響,所以適合服裝銷售預(yù)測(cè)的方法約有幾十種。
科學(xué)準(zhǔn)確的銷售預(yù)測(cè)可以幫助服裝企業(yè)制定科學(xué)的商品企劃方案,降低研發(fā)和生產(chǎn)成本,減少庫(kù)存,提高企業(yè)整體效益。每類銷售預(yù)測(cè)方法都有各自的算理、優(yōu)缺點(diǎn)及適用類型,如何針對(duì)不同類型的服裝、不同的銷售數(shù)據(jù)及預(yù)測(cè)時(shí)長(zhǎng)的需求選擇合適的銷售預(yù)測(cè)方法尤為重要[3]。而服裝商品企劃人員及相關(guān)學(xué)者實(shí)際做銷售預(yù)測(cè)時(shí),很難從大量相關(guān)文獻(xiàn)中找到合適的預(yù)測(cè)方法,且需花費(fèi)大量時(shí)間。
為解決上述問(wèn)題,本文對(duì)服裝銷售預(yù)測(cè)研究進(jìn)行了系統(tǒng)回顧,在常用服裝銷售定量預(yù)測(cè)方法中,主要對(duì)時(shí)間序列法(Time series prediction method, TSPM)、回歸分析法(Regression analysis method, RAM)、灰色預(yù)測(cè)模型(Grey model, GM)和人工神經(jīng)網(wǎng)絡(luò)(Artificial neural network, ANN)這4種方法從優(yōu)缺點(diǎn)、優(yōu)化情況及適用類型3個(gè)方面進(jìn)行歸納總結(jié),同時(shí)對(duì)機(jī)器學(xué)習(xí)的部分組合算法進(jìn)行了舉例與歸納。為服裝銷售預(yù)測(cè)研究人員及相關(guān)領(lǐng)域?qū)W者提供一個(gè)系統(tǒng)的參考,對(duì)服裝銷售預(yù)測(cè)領(lǐng)域的研究具有一定的積極推動(dòng)作用。
1 服裝銷售預(yù)測(cè)概述
1.1 服裝銷售預(yù)測(cè)概念
服裝銷售預(yù)測(cè)是指針對(duì)服裝這一產(chǎn)品屬性,通過(guò)分析天氣、流行趨勢(shì)、運(yùn)營(yíng)銷售、價(jià)格等影響服裝銷售的因素以及服裝銷售數(shù)據(jù)等信息,綜合個(gè)人經(jīng)驗(yàn)與事物發(fā)展規(guī)律,運(yùn)用統(tǒng)計(jì)學(xué)、數(shù)學(xué)以及邏輯學(xué)中的相關(guān)方法預(yù)測(cè)服裝在未來(lái)一段時(shí)間內(nèi)的銷量。目的是降低未來(lái)決策所產(chǎn)生的風(fēng)險(xiǎn),從而減少庫(kù)存。
1.2 銷售預(yù)測(cè)方法的分類
根據(jù)預(yù)測(cè)方法的性質(zhì),將服裝銷售預(yù)測(cè)方法分為定性預(yù)測(cè)與定量預(yù)測(cè),如圖1所示。
定性銷售預(yù)測(cè)是指擁有較好實(shí)踐能力和理論能力的相關(guān)專家,通過(guò)對(duì)歷史數(shù)據(jù)的分析與研究,結(jié)合個(gè)人經(jīng)驗(yàn)與實(shí)際情況,從性質(zhì)和程度兩個(gè)方面對(duì)事物未來(lái)的發(fā)展做出綜合的分析及判斷,形成綜合的評(píng)估方案[4]。在服裝銷售預(yù)測(cè)應(yīng)用研究中,常用的有德?tīng)柗品?、專家小組討論法及個(gè)人判斷法。
定量銷售預(yù)測(cè)是指將歷史銷售數(shù)據(jù)和影響因素作為輸入變量,根據(jù)輸入變量與未來(lái)銷量之間的相關(guān)關(guān)系用邏輯算法建立數(shù)學(xué)模型,通過(guò)數(shù)學(xué)運(yùn)算,計(jì)算出未來(lái)的銷量或需求[5]。優(yōu)點(diǎn)是客觀性強(qiáng),計(jì)算過(guò)程直觀,易操作,預(yù)測(cè)結(jié)果不會(huì)因人而異;缺點(diǎn)是需要一定的數(shù)據(jù),計(jì)算量大,運(yùn)算時(shí)間長(zhǎng)。在服裝銷售預(yù)測(cè)研究中,常用的有TSPM、RAM、GM及ANN等,如表1所示。
2 服裝銷售定量預(yù)測(cè)方法
2.1 時(shí)間序列法(TSPM)
19世紀(jì)30年代,Yule和Walker提出自回歸模型(AR)和移動(dòng)平均模型(MA),兩者結(jié)合形成自回歸移動(dòng)平均模型(ARMA)[47],AR、MA以及ARMA三者共同構(gòu)成了現(xiàn)代時(shí)間序列模型的基礎(chǔ)。TSPM的優(yōu)點(diǎn)是簡(jiǎn)單易掌握、運(yùn)算量小、計(jì)算速度快、精確度較高;缺點(diǎn)是不能準(zhǔn)確分析事物發(fā)展的內(nèi)在規(guī)律。TSPM應(yīng)用在服裝銷售預(yù)測(cè)中的主要有加權(quán)移動(dòng)平均模型(WMA)和差分自回歸移動(dòng)平均模型(ARIMA)。
Giri等[6]從Twitter、Instagram及Facebook上收集某品牌的服裝銷售數(shù)據(jù),用模糊綜合評(píng)判將收集到的數(shù)據(jù)信息模糊化,用時(shí)間序列模型預(yù)測(cè)服裝的銷量,證明了品牌服裝的推文與服裝銷量之間存在一定的相關(guān)性。李欣芮[7]用隨機(jī)森林算法優(yōu)化時(shí)間序列模型的殘差,提高了時(shí)間序列模型的預(yù)測(cè)精度。在一個(gè)數(shù)據(jù)序列中,越靠近預(yù)測(cè)時(shí)間的數(shù)據(jù)對(duì)銷售預(yù)測(cè)結(jié)果的影響越大,Xu等[8]按從過(guò)去到現(xiàn)在的時(shí)間順序依次增大給歷史數(shù)據(jù)賦權(quán),得到WMA模型。為解決WMA模型中存在的預(yù)測(cè)結(jié)果隨時(shí)間推移的現(xiàn)象,陳銀光等[9]在WMA模型的基礎(chǔ)上引入趨勢(shì)概念,將預(yù)測(cè)結(jié)果整體前移一個(gè)時(shí)間單位進(jìn)行預(yù)測(cè),有效提高了預(yù)測(cè)結(jié)果的準(zhǔn)確度。在缺少歷史銷售數(shù)據(jù)、影響因素多且一直在變化等多種不利條件影響的前提下,ARIMA模型與自適應(yīng)模糊神經(jīng)系統(tǒng)(ANFIS)相比,線性結(jié)構(gòu)簡(jiǎn)單的ARIMA模型的預(yù)測(cè)效果更好[10]。
TSPM因只考慮時(shí)間因素對(duì)銷售結(jié)果的影響[11],忽略了其他影響因素與銷量之間的因果關(guān)系,且歷史數(shù)據(jù)的時(shí)間與預(yù)測(cè)未來(lái)銷量的時(shí)間相隔越遠(yuǎn),歷史數(shù)據(jù)的影響越小[12-13],所以TSPM適用于歷史數(shù)據(jù)離散程度小且影響因素小的短中期基礎(chǔ)型服裝銷售預(yù)測(cè)。
2.2 回歸分析法(RAM)
1855年,Galton在研究人類遺傳問(wèn)題時(shí)提出回歸分析法(RAM)[48],RAM是一種研究自變量與因變量相關(guān)關(guān)系的預(yù)測(cè)性建模方法。優(yōu)點(diǎn)是簡(jiǎn)單易計(jì)算,可計(jì)量各因素之間的相關(guān)程度與擬合程度,多元回歸適用于多變量預(yù)測(cè);缺點(diǎn)是相對(duì)簡(jiǎn)單低級(jí),易出現(xiàn)過(guò)擬合。在服裝銷售預(yù)測(cè)領(lǐng)域常用的有一元回歸(SRA)及多元回歸(MRA)等。
Nivasanon等[14]用指數(shù)平滑法、MA與RAM分別做服裝銷售預(yù)測(cè),結(jié)果顯示3種方法中RAM的預(yù)測(cè)誤差最小。Demiriz等[15]對(duì)服裝產(chǎn)品按屬性的相似度分類,用MRA預(yù)測(cè),分類后可提高服裝銷售預(yù)測(cè)的精度。李鋒等[16]用MRA模擬預(yù)測(cè)紡織品銷量,用逐步回歸剔除影響紡織品銷售較小的因素,解決了多元回歸影響因子較多時(shí)自由度高、共線性低的缺點(diǎn)。池可等[17]對(duì)多種服裝銷售定量預(yù)測(cè)算法比較,發(fā)現(xiàn)MRA適用于季節(jié)型服裝銷售預(yù)測(cè),對(duì)時(shí)尚型服裝銷量預(yù)測(cè)時(shí)SRA的預(yù)測(cè)精度較高,主要原因是在時(shí)尚型服裝的所有影響因素中,流行趨勢(shì)影響最大,忽略了其他影響相對(duì)較小的因素[18]。
RAM根據(jù)歷史銷售數(shù)據(jù)及影響因素等自變量預(yù)測(cè)銷量(因變量),但影響因素是隨著時(shí)間的變化在不斷變化的,且時(shí)間越長(zhǎng),影響因素變化的概率越大,所以RAM不適合長(zhǎng)期預(yù)測(cè)。在算理上,擁有多個(gè)自變量的MRA比只有一個(gè)自變量的SRA更適合具有多因素影響的服裝銷售預(yù)測(cè),其中SRA適用于時(shí)尚型服裝銷售預(yù)測(cè),MRA更適用于季節(jié)型服裝銷售預(yù)測(cè)。
2.3 灰色預(yù)測(cè)模型(GM)
1982年,鄧聚龍教授[49]發(fā)表了《灰色控制系統(tǒng)》,此后灰色系統(tǒng)理論經(jīng)過(guò)近40年的蓬勃發(fā)展,形成了一門包含分析、控制、決策、優(yōu)化及預(yù)測(cè)等多功能的學(xué)科結(jié)構(gòu)體系[50]。GM的優(yōu)點(diǎn)[51]是計(jì)算量小,少量數(shù)據(jù)就可以預(yù)測(cè),適用于短、中、長(zhǎng)期預(yù)測(cè);缺點(diǎn)是預(yù)測(cè)對(duì)象的原始數(shù)據(jù)需符合殘差檢驗(yàn)或經(jīng)變換處理后符合殘差檢驗(yàn)[19]。在服裝銷售預(yù)測(cè)領(lǐng)域常用的有單維灰色模型(GM(1,1))和多維灰色模型(GM(1,N))。
江玉杰[20]將GM(1,1)預(yù)測(cè)后得到預(yù)測(cè)值置入到其原始數(shù)據(jù)序列x(0)(k+1)之后,去掉原始數(shù)據(jù)序列中距離預(yù)測(cè)時(shí)間最遠(yuǎn)的一個(gè)數(shù)據(jù),形成新序列X(0)={x(0),x(0)(3)…,x(0)(k),x(0)(k+1)},發(fā)展系數(shù)減小,預(yù)測(cè)精度提高,實(shí)驗(yàn)證明優(yōu)化后的新生成GM(1,1)模型預(yù)測(cè)結(jié)果的平均相對(duì)誤差及均方差均比原始GM(1,1)模型小。由于GM(1,1)模型只受歷史銷售數(shù)據(jù)和發(fā)展系數(shù)的影響,Xia等[21]將季節(jié)性因素引入到GM(1,1)模型中,韓曙光等[22]在GM(1,1)模型的基礎(chǔ)上引入已銷售產(chǎn)品的影響因子和銷售過(guò)程的干擾因子,一定程度上彌補(bǔ)了GM(1,1)模型只受歷史銷售數(shù)據(jù)和發(fā)展系數(shù)影響的缺點(diǎn)。
王昕彤等[23-24]在GM(1,N)的微分方程中加入線性修正量和灰色作用量來(lái)優(yōu)化GM(1,N)的建模機(jī)理、參數(shù)使用及模型結(jié)構(gòu)上的缺陷,形成OGM(1,N)模型,實(shí)驗(yàn)證明,優(yōu)化后的OGM(1,N)模型預(yù)測(cè)誤差小于GM(1,N),并通過(guò)統(tǒng)一參數(shù)估計(jì)和時(shí)間響應(yīng)式構(gòu)建DGM(1,1)模型,在DGM(1,1)的基礎(chǔ)上引入平滑性算子得到ROGM(1,1),該模型適合銷售數(shù)據(jù)離散程度大的中短期預(yù)測(cè)。
為解決歷史數(shù)據(jù)不符合殘差檢驗(yàn)問(wèn)題,黃鴻云等[25]引入臨時(shí)常數(shù)使歷史數(shù)據(jù)曲線平滑,將GM(1,1)模型中的控制灰數(shù)引入到GM(1,N)中,得到優(yōu)化算法IGM(1,N),再將IGM(1,N)預(yù)測(cè)的結(jié)果用BP神經(jīng)網(wǎng)絡(luò)進(jìn)行兩次殘差預(yù)測(cè),得到新組合算法,該組合算法與ARIMA、GM(1,1)和GM(1,N)相比預(yù)測(cè)效果最好。徐倩如[26]將GM模型引入馬爾可夫鏈算法來(lái)優(yōu)化服裝銷售受外界因素影響大的問(wèn)題。GM-馬爾科夫鏈組合模型既降低了單項(xiàng)模型易出錯(cuò)的風(fēng)險(xiǎn),又解決了受外界因素影響較大時(shí)預(yù)測(cè)精度降低的問(wèn)題。
由于GM模型需要原始數(shù)據(jù)符合殘差檢驗(yàn),符合殘差檢驗(yàn)的數(shù)據(jù)相對(duì)平滑[27];同時(shí)影響因素(對(duì)應(yīng)GM中的發(fā)展系數(shù))越小,預(yù)測(cè)精度越高[28-29],因此GM模型適合歷史數(shù)據(jù)平滑且影響因素小的中短期銷售預(yù)測(cè)。在算理上,考慮多種影響因素的GM(1,N)比GM(1,1)更適合具有多因素影響的服裝銷售預(yù)測(cè)。
2.4 人工神經(jīng)網(wǎng)絡(luò)法(ANN)
1943年,Mcculloch和Pitts通過(guò)對(duì)人腦細(xì)胞神經(jīng)元結(jié)構(gòu)的分析和研究,首次提出神經(jīng)元數(shù)學(xué)模型,開(kāi)啟了神經(jīng)學(xué)理論研究時(shí)代[52]。ANN的優(yōu)點(diǎn)是可以最大限度逼近任意非線性關(guān)系,具有自學(xué)習(xí)及尋找最優(yōu)解的能力;缺點(diǎn)是可視性差、訓(xùn)練時(shí)間長(zhǎng)、需要大量數(shù)據(jù)。在服裝銷售預(yù)測(cè)中常用的主要有BP神經(jīng)網(wǎng)絡(luò)、極限學(xué)習(xí)機(jī)(ELM)及徑向基神經(jīng)網(wǎng)絡(luò)(RBF)等。
2.4.1 BP神經(jīng)網(wǎng)絡(luò)
BP神經(jīng)網(wǎng)絡(luò)是一種將誤差逆向傳播以獲得規(guī)律的前饋型神經(jīng)網(wǎng)絡(luò),BP神經(jīng)網(wǎng)絡(luò)的權(quán)值和閾值被遺傳算法優(yōu)化后,預(yù)測(cè)結(jié)果的精度提高[30-31]。喻寶祿等[32]利用主成分分析法對(duì)影響預(yù)測(cè)的因素分類,將分類后的主要因素輸入到BP神經(jīng)網(wǎng)絡(luò),用遺傳算法迭代優(yōu)化連接權(quán)值和閾值,雖然預(yù)測(cè)精度有所增加,但也增加了計(jì)算時(shí)間。汪蕓芳等[33]用BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)服裝的庫(kù)存,用GM(1,1)預(yù)測(cè)節(jié)假日購(gòu)物時(shí)期的庫(kù)存作為BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)的補(bǔ)充,兩種方法相結(jié)合的預(yù)測(cè)效果更好。
2.4.2 極限學(xué)習(xí)機(jī)(ELM)
ELM是一種前饋式神經(jīng)網(wǎng)絡(luò)算法,泛化能力強(qiáng),只產(chǎn)生唯一最優(yōu)解,解決了ANN易陷入局部最優(yōu)解的缺點(diǎn)[34]。在銷售預(yù)測(cè)應(yīng)用上,ELM比BP神經(jīng)網(wǎng)絡(luò)的預(yù)測(cè)誤差小,穩(wěn)定性高,計(jì)算時(shí)間短。ELM的權(quán)值和閾值在訓(xùn)練過(guò)程中會(huì)隨機(jī)自動(dòng)生成,縮短了模型的訓(xùn)練學(xué)習(xí)時(shí)間,運(yùn)算效率高,但也會(huì)造成模型每次輸出的結(jié)果都不同,于淼[35]在研究服裝銷售預(yù)測(cè)時(shí)用集成理論優(yōu)化了ELM的這一問(wèn)題。Sun等[36]用結(jié)構(gòu)方程模型對(duì)銷售數(shù)據(jù)歸一化處理,選用尺寸、顏色及價(jià)格等作為輸入變量,用回歸積分法對(duì)ELM的輸出結(jié)果優(yōu)化得到組合預(yù)測(cè)模型ELME,實(shí)驗(yàn)證明ELME預(yù)測(cè)結(jié)果的標(biāo)準(zhǔn)差和均方差比ELM小20%,但ELME的預(yù)測(cè)總時(shí)間是ELM的近100倍。變異系數(shù)(CVS)表示銷售數(shù)據(jù)的波動(dòng)程度,CVS越大,ELM的預(yù)測(cè)誤差越低,反之預(yù)測(cè)誤差越大,這是因?yàn)镋LM具有記憶功能,當(dāng)訓(xùn)練CVS較小的數(shù)據(jù)時(shí),ELM很難記憶和區(qū)分?jǐn)?shù)據(jù)的規(guī)律,導(dǎo)致學(xué)習(xí)效率低,誤差高。
2.4.3 徑向基神經(jīng)網(wǎng)絡(luò)(RBF)
RBF是一種前饋型神經(jīng)網(wǎng)絡(luò),具有全局最優(yōu)逼近性能,不存在局部極小問(wèn)題,具有較好的映射能力,泛化能力和分類能力好,學(xué)習(xí)速度和收斂速度快。RBF可以對(duì)影響服裝銷售因素的指標(biāo)數(shù)據(jù)分類,識(shí)別輸入數(shù)據(jù)與基函數(shù)中心點(diǎn)的距離,將距離帶入徑向基函數(shù)作為輸入層與隱含層的連接,最后把所有徑向基函數(shù)的輸出值加權(quán)合并,得出預(yù)測(cè)值。RBF在選擇不同的預(yù)測(cè)精度與徑向基函數(shù)寬度組合時(shí)會(huì)產(chǎn)生不同的收斂效果與逼近能力,影響網(wǎng)絡(luò)的擬合與泛化。池可[37]采用網(wǎng)格法優(yōu)化了RBF的這一問(wèn)題,證明了RBF在服裝銷售預(yù)測(cè)中應(yīng)用的可行性。
由于ANN具有自學(xué)習(xí)能力以及可以逼近任意非線性關(guān)系的優(yōu)點(diǎn),所以對(duì)離散程度越大即規(guī)律性越強(qiáng)的數(shù)據(jù)學(xué)習(xí)效果越好,而離散程度小的數(shù)據(jù)規(guī)律性小,學(xué)習(xí)效果差;銷售的影響因素越小,ANN中相應(yīng)的權(quán)值越小,學(xué)習(xí)效果則越差,因此ANN適合歷史數(shù)據(jù)離散程度大且影響因素大的中短期服裝銷售預(yù)測(cè)。
2.5 機(jī)器學(xué)習(xí)組合算法
機(jī)器學(xué)習(xí)是研究計(jì)算機(jī)重復(fù)同樣或類似工作時(shí),如何模擬人類學(xué)習(xí)行為以獲取新的知識(shí)或技能,重組已有的知識(shí)結(jié)構(gòu)并不斷改善自身的性能[53],研究目標(biāo)是從大量異構(gòu)數(shù)據(jù)集中提取隱藏規(guī)律。在預(yù)測(cè)領(lǐng)域常用的有ANN、樸素貝葉斯、決策樹(shù)、支持向量機(jī)(SVM)以及隨機(jī)森林等[54]。機(jī)器學(xué)習(xí)組合算法是指將機(jī)器學(xué)習(xí)所包含的算法與其他或自身算法相組合,以提高計(jì)算精度。
機(jī)器學(xué)習(xí)中的SVM、隨機(jī)森林、線性回歸、梯度提升算法、LASSO回歸、前向選擇回歸等預(yù)測(cè)方法中,非線性模型(SVM、梯度提升模型和隨機(jī)森林)的銷售預(yù)測(cè)性能與準(zhǔn)確率要優(yōu)于線性模型[38](線性回歸、前向選擇回歸和Lasso回歸模型)。Teucke等[39]用決策樹(shù)法對(duì)季節(jié)型服裝的歷史銷售數(shù)據(jù)分類,用SVM模型對(duì)分類后的數(shù)據(jù)訓(xùn)練,但由于沒(méi)有足夠的數(shù)據(jù)并未驗(yàn)證預(yù)測(cè)的效果。Singh等[40]用以梯度提升算法為核心的XGBoost庫(kù)和長(zhǎng)短記憶神經(jīng)網(wǎng)絡(luò)(LSTM)訓(xùn)練服裝歷史銷售數(shù)據(jù),結(jié)果顯示XGBoost庫(kù)的預(yù)測(cè)效果較好。為探索產(chǎn)品需求和定價(jià)對(duì)銷量的影響,F(xiàn)erreira等[41]利用RAM和決策樹(shù)研發(fā)了LP Bound Algorithm算法解決商品定價(jià)的問(wèn)題,并將該算法應(yīng)用到美國(guó)電商Rue La La的日常定價(jià)決策系統(tǒng)中,建立需求預(yù)測(cè)模型,發(fā)現(xiàn)提升中高端產(chǎn)品的價(jià)格并不會(huì)導(dǎo)致銷售額下降。
由于服裝銷售的影響因素非常復(fù)雜,將影響因素模糊化引入到ANN,得到模糊神經(jīng)網(wǎng)絡(luò)(FNN)。趙學(xué)斌等[42]將MRA與FNN結(jié)合,并將MRA、FNN及其組合算法分別應(yīng)用到服裝銷售預(yù)測(cè)中,結(jié)果顯示其組合算法的預(yù)測(cè)精度最高。張小娟[43]在ANN的基礎(chǔ)上引入模糊推理系統(tǒng)(FIS),形成自適應(yīng)模糊神經(jīng)系統(tǒng)(ANFIS),ANFIS在算理上更適用于影響因素不確定、數(shù)據(jù)非線性或缺失的服裝銷售預(yù)測(cè)。Huang等[44]將銷量預(yù)測(cè)轉(zhuǎn)化為消費(fèi)者需求預(yù)測(cè),用相似屬性服裝的銷量和影響因素代替預(yù)售服裝的歷史銷量及影響因素,分別用ANFIS與ANN預(yù)測(cè)新服裝的需求量,結(jié)果顯示ANFIS的預(yù)測(cè)精度比ANN高30%以上,但將ANFIS應(yīng)用在中長(zhǎng)期服裝銷售預(yù)測(cè)時(shí)[45],其預(yù)測(cè)效果并不優(yōu)于ANN。Au等[46]在研究貝葉斯信息準(zhǔn)則(BIC)優(yōu)化進(jìn)化神經(jīng)網(wǎng)絡(luò)(ENN)時(shí),提出一種尋找最大隱藏神經(jīng)元個(gè)數(shù)的預(yù)搜索機(jī)制。優(yōu)化后的ENN計(jì)算速度更快。將其與SARIMA對(duì)比,ENN對(duì)CVS值小的銷售數(shù)據(jù)預(yù)測(cè)效果較好,SARIMA對(duì)CVS值大的銷售數(shù)據(jù)預(yù)測(cè)效果較好。
3 結(jié)論與展望
服裝銷售預(yù)測(cè)作為服裝商品企劃的核心技術(shù),在大型服裝企業(yè)得到廣泛的研究和應(yīng)用。本文基于銷售預(yù)測(cè)方法的算法原理,對(duì)4類銷售預(yù)測(cè)方法的優(yōu)缺點(diǎn)、優(yōu)化歷程及適用類型進(jìn)行梳理分析,總結(jié)如下:
a) TSPM適用于歷史數(shù)據(jù)離散程度小且影響因素少的短中期服裝銷售預(yù)測(cè)。
b) RAM中擁有多個(gè)自變量的MRA比只有一個(gè)自變量的SRA更適合具有多因素影響的服裝銷售預(yù)測(cè),其中MRA較適用于季節(jié)型服裝銷售預(yù)測(cè),SRA較適用于時(shí)尚型服裝銷售預(yù)測(cè)。
c) GM模型適合歷史數(shù)據(jù)平滑且影響因素少的中短期銷售預(yù)測(cè),在算理上,GM(1,N)比GM(1,1)更適合具有多因素影響的服裝銷售預(yù)測(cè),ROGM(1,1)適合銷售數(shù)據(jù)離散程度大的短中期服裝銷售預(yù)測(cè)。
d) ANN適合歷史數(shù)據(jù)離散程度大且影響因素大的中短期服裝銷售預(yù)測(cè)。
e) 機(jī)器學(xué)習(xí)組合算法中ANFIS適用于數(shù)據(jù)非線性或缺失的短期銷售預(yù)測(cè);ENN對(duì)歷史數(shù)據(jù)CVS值小的預(yù)測(cè)效果較好,SARIMA對(duì)歷史數(shù)據(jù)CVS值大的預(yù)測(cè)效果較好。機(jī)器學(xué)習(xí)中的非線性模型(SVM、梯度提升模型和隨機(jī)森林)要優(yōu)于線性模型的銷售預(yù)測(cè)性能與準(zhǔn)確率。
服裝商品企劃人員及相關(guān)學(xué)者可根據(jù)以上研究結(jié)論快速找到合適的預(yù)測(cè)方法,但本文僅對(duì)4類常用銷售預(yù)測(cè)方法進(jìn)行相關(guān)分析與歸納,相關(guān)學(xué)者可以在此基礎(chǔ)之上對(duì)其他算法總結(jié),最后將所有常用服裝銷售預(yù)測(cè)方法寫入軟件或系統(tǒng),以便服裝企業(yè)人員使用。
參考文獻(xiàn):
[1]司守奎,孫璽菁,張德存.數(shù)學(xué)建模算法與應(yīng)用習(xí)題解答[M].2版.北京:國(guó)防工業(yè)出版社,2015:397-425.
SI Shoukui, SUN Xijing, ZHANG Decun. Mathematical Modeling Algorithm and Application[M]. 2nd Edition. Beijing: National Defense Industry Press, 2015: 397-425.
[2]鞠雨宏,張煥偉.經(jīng)濟(jì)預(yù)測(cè)方法綜述與選擇[C]∥后疫情時(shí)代中國(guó)經(jīng)濟(jì)發(fā)展論壇論文集.長(zhǎng)春:吉林省財(cái)政科學(xué)研究所,2020:10.
JU Yuhong, ZHANG Huanwei. Summary and selection of economic forecasting methods[C]∥ Proceedings of the Forum on China's Economic Development in the Post-Epidemic Era. Changchun: Jilin Institute of financial Sciences, 2020: 10.
[3]周小溪,徐行,孟劍飛,等.服裝銷售預(yù)測(cè)方法研究進(jìn)展[J].針織工業(yè),2020(3):68-72.
ZHOU Xiaoxi, XU Hang, MENG Jianfei, et al. Research progress of garment sales forecasting methods[J]. Knitting Industries. 2020(3): 68-72.
[4]李笑然,姜毅.銷售預(yù)測(cè)中定量與定性分析方法的比較與探究[J].商業(yè)會(huì)計(jì),2017(6):65-67.
LI Xiaoran, JIANG Yi. Comparison and exploration of quantitative and qualitative analysis methods in sales forecast[J]. Commercial Accounting, 2017(6): 65-67.
[5]梁炯聰.基于時(shí)間序列的服裝銷售預(yù)測(cè)系統(tǒng)研究與應(yīng)用[D].廣州:廣東工業(yè)大學(xué),2015.
LIANG Jiongcong. Research and Application of Clothing Sales Forecasting System Based on Time Series[D]. Guangzhou: Guangdong University of Technology, 2015.
[6]GIRI C, THOMASSEY S, ZENG X. Exploitation of social network data for forecasting garment sales[J]. International Journal of Computational Intelligence Systems, 2019, 12(2): 1423-1435.
[7]李欣芮.基于組合模型的銷售量預(yù)測(cè)及優(yōu)化[D].北京:華北電力大學(xué),2019.
LI Xinrui. Sales Forecasting Based on Combination Model[D]. Beijing: North China Electric Power University, 2019.
[8]XU W X, LIAO F J, LI S, et al. Design and application for the method of dynamic weighted moving average forecasting[C]// Sixth International Conference on Intelligent Systems Design and Engineering Applications. IEEE Computer Society, Guiyang, China. IEEE, 2015: 278-280.
[9]陳銀光,于守健.基于改進(jìn)加權(quán)移動(dòng)平均法的服裝銷售預(yù)測(cè)[J].智能計(jì)算機(jī)與應(yīng)用,2018,8(6):54-56,60.
CHEN Yinguang,YU Shoujian. Clothing sales forecast based on improved weighted moving average method[J]. Intelligent Computer and Application, 2018, 8(6): 54-56, 60.
[10]THOMASSEY S. Sales forecasts in clothing industry: The key success factor of the supply chain management[J]. International Journal of Production Economics, 2010, 128(2): 470-483.
[11]ISSAO CHIBA G, MARIA MENDES LUNA M. Time series models: An analysis of demand forecast accuracy for a product line in a clothing company[J]. Revista Gesto da Produo Operaes e Sistemas, 2020, 15(4): 219-251.
[12]宋憲捷,周建軍,鹿士凱,等.基于客流量的貝葉斯回歸服裝銷售預(yù)測(cè)方法[J].絲綢,2018,55(4):44-48.
SONG Xianjie, ZHOU Jianjun, LU Shikai, et al. Clothing sales forecast based on passenger flow and Bayesian regression model[J]. Journal of Silk, 2018, 55(4): 44-48.
[13]DOSZYN M. Intermittent demand forecasting in the enterprise: Empirical verification[J]. Journal of Forecasting, 2019, 38(5): 459-469.
[14]NIVASANON C, RUEKKASAEM L, AUNGKULA-NON P. Demand forecasting for online market stock[C]∥ IC4E'19: Proceedings of the 10th International Conference on E-Education, E-Business, E-Management and E-Learning. Tokyo, Japan, New York: ACM Press, 2019: 292-297.
[15]DEMIRIZ A. Demand Forecasting based on pairwise item associations[J]. Procedia Computer Science, 2014, 36: 261-268.
[16]李鋒,袁瑣云.紡織印染品銷售趨勢(shì)分析系統(tǒng)的設(shè)計(jì)與實(shí)現(xiàn)[J].計(jì)算機(jī)應(yīng)用,2019,39(S2):271-274.
LI Feng, YUAN Suoyun. Design and implementation of textile printing and dyeing product sales trend analysis system[J]. Journal of Computer Applications,2019, 39(S2): 271-274.
[17]池可,陳雁.基于對(duì)象信息的服裝銷售預(yù)測(cè)方法選擇[J].紡織學(xué)報(bào),2009,30(10):139-142.
CHI Ke, CHEN Yan. Forecast method for garment sales based on object information[J]. Journal of Textile Research, 2009, 30(10): 139-142.
[18]TEHRANI A F, AHRENS D. Enhanced predictive models for purchasing in the fashion field by using kernel machine regression equipped with ordinal logistic regression[J]. Journal of Retailing and Consumer Services, 2016, 32: 131-138.
[19]孫利輝.基于灰色理論的服裝銷售預(yù)測(cè)及庫(kù)存優(yōu)化研究[D].杭州:浙江理工大學(xué),2016.
SUN Lihui. Study on the Sales Forecasting of Clothing andInventory Problem Based on the Grey Theory[D]. Hangzhou: Zhejiang Sci-Tech University, 2016.
[20]江玉杰.基于新陳代謝GM(1,1)模型的服裝產(chǎn)品銷售預(yù)測(cè)[J].浙江紡織服裝職業(yè)技術(shù)學(xué)院學(xué)報(bào),2017,16(4):63-67.
JIANG Yujie. Sales forecast of clothing products based on metabolic GM(1, 1) model[J]. Journal of Zhejiang Textile & Fashion College, 2017, 16(4): 63-67.
[21]XIA M, WONG WK. A seasonal discrete grey forecasting model for fashion retailing[J]. Knowledge-Based Systems, 2014, 57: 119-126.
[22]韓曙光,孫利輝.基于改進(jìn)GM(1,1)模型的服裝配貨決策研究[J].浙江理工大學(xué)學(xué)報(bào)(社會(huì)科學(xué)版),2015(6):508-511.
HAN Shuguang, SUN Lihui. Study on clothing distribution decisions based on improved GM(1, 1 model[J]. Journal of Zhejiang Sci-Tech University(Social Sciences Edition), 2015(6): 508-511.
[23]王昕彤,王秀敏,郭瑞良,等.基于灰色理論的服裝企業(yè)銷售預(yù)測(cè)模型[J].絲綢,2020,57(2):55-60.
WANG Xintong, WANG Xiumin, GUO Ruiliang, et al. Sales prediction model of clothing enterprises based on grey theory[J].Journal of Silk, 2020, 57(2): 55-60.
[24]王昕彤.基于灰色理論的服裝企業(yè)多渠道銷售預(yù)測(cè)模型的研究[D].北京:北京服裝學(xué)院,2019.
WANG Xintong. Research on the Prediction Modelof Clothing Enterprises Multichannel Sales Based on Grey Theory[D]. Beijing: Beijing Institute of Fashion Technology, 2019.
[25]黃鴻云,劉衛(wèi)校,丁佐華.基于多維灰色模型及神經(jīng)網(wǎng)絡(luò)的銷售預(yù)測(cè)[J].軟件學(xué)報(bào),2019,30(4):1031-1044.
HUANG Hongyun, LIU Weixiao, DING Zuohua. Sales forecasting based on multi-dimensional grey model and neural network[J]. Journal of Software, 2019, 30(4): 1031-1044.
[26]徐倩茹.基于組合預(yù)測(cè)模型的服裝銷量預(yù)測(cè)研究[D].西安:長(zhǎng)安大學(xué),2018.
XU Qianru. Research of the Forecast of Apparel Sales Based on the Combined Forecasting Model. [D].Xi′an: Chang′an University, 2018.
[27]劉思峰,曾波,劉解放,等.GM(1,1)模型的幾種基本形式及其適用范圍研究[J].系統(tǒng)工程與電子技術(shù),2014,36(3):501-508.
LIU Sifeng, ZENG Bo, LIU Jiefang, et al. Several basic models of GM(1, 1) and their applicable bound[J]. Systems Engineering and Electronics, 2014, 36(3): 501-508.
[28]劉衛(wèi)校.基于多維灰色模型與神經(jīng)網(wǎng)絡(luò)的銷售預(yù)測(cè)模型研究[D].杭州:浙江理工大學(xué),2017.
LIU Weixiao. The Research on Sales Forecasting Model Based on Multidimensional Grey Model and Neural Network[D]. Hangzhou: Zhejiang Sci-Tech University, 2017.
[29]劉衛(wèi)校.基于離散灰色預(yù)測(cè)模型與人工神經(jīng)網(wǎng)絡(luò)混合智能模型的時(shí)尚銷售預(yù)測(cè)[J].計(jì)算機(jī)應(yīng)用,2016,36(12):3378-3384.
LIU Weixiao. Hybrid intelligent model for fashion sales forecasting based ondiscrete grey forecasting model and artificial neural network[J]. Journal of Computer Applications, 2016, 36(12): 3378-3384.
[30]羅戎蕾,劉紹華,蘇晨.基于遺傳算法的BP神經(jīng)網(wǎng)絡(luò)服裝銷售預(yù)測(cè)方法[J].北京郵電大學(xué)學(xué)報(bào),2014,37(4):39-43.
LUO Ronglei, LIU Shaohua, SU Chen. Garment sales forecast method based on geneticalgorithm and BP neural network[J]. Journal of Beijing University of Posts and Telecommunications, 2014, 37(4): 39-43.
[31]圣文順,趙翰馳,孫艷文.基于改進(jìn)遺傳算法優(yōu)化BP神經(jīng)網(wǎng)絡(luò)的銷售預(yù)測(cè)模型[J].計(jì)算機(jī)系統(tǒng)應(yīng)用,2019,28(12):200-204.
SHENG Wenshun, ZHAO Hanchi, SUN Yanwen. Sales forecasting model based on BP neural network optimized by improved genetic algorithms[J]. Computer Systems & Applications, 2019, 28(12): 200-204.
[32]喻寶祿,段迅,吳云.BP神經(jīng)網(wǎng)絡(luò)數(shù)據(jù)預(yù)測(cè)模型的建立及應(yīng)用[J].計(jì)算機(jī)與數(shù)字工程,2016,44(3):482-486,545.
YU Baolu, DUAN Xun, WU Yun. Establishment and application of data prediction model based on BP neural network[J].Computer & Digital Engineering, 2016, 44(3): 482-486, 545.
[33]汪蕓芳,史意,陳麗華.基于BP神經(jīng)網(wǎng)絡(luò)及灰色GM(1,1)模型的服裝供應(yīng)鏈第三方庫(kù)存預(yù)測(cè)應(yīng)用研究[J].數(shù)學(xué)的實(shí)踐與認(rèn)識(shí),2020,50(3):277-285.
WANG Yunfang, SHI Yi, CHEN Lihua. Third-Party inventory forecasting application research of apparel supply chain based on BP neural network and grey GM(1, 1) model[J].Mathematics in Practice and Theory,2020,50(3):277-285.
[34]WONG W K, GUO Z X. A hybrid intelligent model for medium-term sales forecasting in fashion retail supply chains using extreme learning machine and harmony search algorithm[J]. International Journal of Production Economics, 2010, 128(2): 614-624.
[35]于淼.基于機(jī)器學(xué)習(xí)理論的服裝中期銷售預(yù)測(cè)研究[D].杭州:浙江理工大學(xué),2015.
YU Miao. Study on the Mid-term Sales Forecasting of Clothing Based on the Theory of Extreme Learning Machine[D]. Hangzhou: Zhejiang Sci-Tech Univer-sity, 2015.
[36]SUN Z L, CHOI T, AUK, et al. Sales forecasting using extreme learning machine with applications in fashion retailing[J]. Decision Support Systems, 2008, 46(1): 411-419.
[37]池可.服裝銷售預(yù)測(cè)方法及RBF神經(jīng)網(wǎng)絡(luò)模型研究[D].蘇州:蘇州大學(xué),2009.
CHI Ke. Research on the Methods and RBF Neural Network Model for Clothing Sale Prediction[D]. Suzhou: Soochow University, 2009.
[38]CUI R, GALLINO S, MORENO A, et al. The operational value of social media information[J]. Production and Operations Management, 2018, 27(10): 1749-1769.
[39]TEUCKE M, AIT-ALLA A, EL-BERISHY N, et al. Forecasting of seasonal apparel products[C]// 4th International Conference on Dynamics in Logistics. Cham: Springer International Publishing. 2015: 633-642.
[40]SINGH PK, GUPTA Y, JHA N, et al. Fashion retail: Forecasting demand for new items[C]//KDD 2019 Workshop. Alaska-USA: Anchorage, 2019.
[41]FERREIRAK J, LEE B H A, SIMCHI-LEVI D. Analytics for an online retailer: Demand forecasting and price optimization[J]. Manufacturing & Service Operations Management, 2016, 18(1): 69-88.
[42]趙學(xué)斌,李大學(xué),謝名亮.短生命周期產(chǎn)品的銷量預(yù)測(cè)模型研究[J].計(jì)算機(jī)工程與設(shè)計(jì),2010,31(11):2527-2530.
ZHAO Xuebin, LI Daxue, XIE Mingliang. Study on sales forecasting models of short lifecycle products[J]. Computer Engineering and Design, 2010, 31(11): 2527-2530.
[43]張小娟.自適應(yīng)神經(jīng)模糊推理系統(tǒng)(ANFIS)及其仿真[J].電子設(shè)計(jì)工程,2012,20(5):11-13.
ZHANG Xiaojuan. Study on the adaptive network-based fuzzy inference system and simulation[J]. Electronic Design Engineering, 2012, 20(5): 11-13.
[44]HUANG H, LIU Q. Intelligent retail forecasting system for new clothing products considering stock-out[J]. Fibres and Textiles in Eastern Europe, 2017, 25(1): 10-16.
[45]AKSOY A, OZTURK N, SUCKY E. A decision support system for demand forecasting in the clothing industry[J]. International Journal of Clothing Science and Technology, 2012, 24(4): 221-236.
[46]AUK, CHOI T, YU Y. Fashion retail forecasting by evolutionary neural networks[J]. International Journal of Production Economics, 2008, 114(2): 615-630.
[47]FABOZZI F J, FOCARDI S M, RACHEV S T, et al. Autoregressive Moving Average Models[M]. New Jersey: John Wiley & Sons Inc, 2014:171-190.
[48]李清茹,吳俊,葉冬青.相關(guān)和回歸分析的創(chuàng)始人:弗朗西斯·高爾頓[J].中華疾病控制雜志,2018,22(9):981-983.
LI Qingru, WU Jun, YE Dongqing. The pioneer of correlation and regression analysis: Francis Galton[J]. Chinese Journal of Disease Control & Prevention, 2018, 22(9): 981-983.
[49]鄧聚龍.灰色控制系統(tǒng)[J].華中工學(xué)院學(xué)報(bào),1982,10(3):9-18.
DENG Julong. The grey control system[J].Journal of Huazhong University of Science and Technology, 1982, 10(3): 9-18.
[50]周文浩,曾波.灰色關(guān)聯(lián)度模型研究綜述[J].統(tǒng)計(jì)與決策,2020,36(15):29-34.
ZHOU Wenhao, ZENG Bo. A research review of grey relational degree model[J]. Statistics and Decision,2020,36(15):29-34.
[51]張晶晶,李澤霞.灰色預(yù)測(cè)法在情報(bào)研究中的應(yīng)用研究綜述[J].圖書(shū)情報(bào)工作,2020,64(24):106-114.
ZHANG Jingjing, LI Zexia. Review on the application of grey prediction method in information research[J]. Library and Information Service, 2020, 64(24): 106-114.
[52]張馳,郭媛,黎明.人工神經(jīng)網(wǎng)絡(luò)模型發(fā)展及應(yīng)用綜述[J].計(jì)算機(jī)工程與應(yīng)用,2021,57(11):57-69.
ZHANG Chi, GUO Yuan, LI Ming. Review of development and application of artificial neural network models[J]. Computer Engineering and Applications, 2021,57(11):57-69.
[53]楊劍鋒,喬佩蕊,李永梅,等.機(jī)器學(xué)習(xí)分類問(wèn)題及算法研究綜述[J].統(tǒng)計(jì)與決策,2019,35(6):36-40.
YANG Jianfeng, QIAO Peirui, LI Yongmei,et al. A review of machine-learning classification and algorithms[J]. Statistics and Decision, 2019,35(6):36-40.
[54]DIFRANCO G, SANTURRO M. Machine learning, artificial neural networks and social research[J]. Quality & Quantity: International Journal of Methodology, 2021, 55(3): 1007-1025.