炎炎夏日茶的飲品為現今飲料業銷售主力之一,茶飲料似乎成了每天的必 需品,這樣的消費習性不知不覺地融入我們的生活當中,由於市場上消費者購買 行為的改變便利商店通路更是茶飲的主要銷售管道,由於物品在現代通路的流通 速度變快加上通路商瓜分了銷售商品的利潤,製造商為了提高本身的收益越來越 重視銷售成本的降低。因而生產計畫就成為製造商要求的重點,生產計畫做的好 除了可降低茶飲的生產成本外也因生產機台與人員協調佳避免了企業資源的浪 費。然而生產計畫要做的好則客戶對茶飲的訂單需求預測就要做的準確,但是客 戶訂單需求又是完全取決於消費者的行為。然而在商品的銷售與消費者或與其他 產品之關係是非常錯綜複雜,面對茶飲商品銷售的預測問題,以傳統資料分析方 式是不容易窺探出未來市場的需求,製造商莫不期望能有更好的決策支援系統來 輔助。 本研究是以連鎖便利商店銷售點(Point of sale,POS)的茶飲銷售資料為基礎, 採用資料探勘的方法論建構銷售預測模式,在預測模式的建構時間序列部分採用 移動平均法(MA)及自我迴歸整合移動平均法(ARIMA);類神經網路部分則以倒傳 遞法為主,搭配Granger 因果檢定與自我迴歸(AR) 等法建構了4種預測模型。最 後以實際的銷售資料進行預測模型的評估,得出結果發現以Granger 因果檢定加 上自我迴歸(AR)與類神經網路結合的倒傳遞類神經網路模型的預測績效最好。
In hot summer, tea is one of the main sales in beverage, and drinking tea is becoming one of daily necessities today. Such a habit of drinking tea is quietly integrated into our lives. Because of changes in behavior of consumers in the market, convenience stores have become the main channel for selling tea. Today the speed of modern sales channel and the dividing profits by distributors have changed the manufacturers’ attitude to put effort on reducing the cost of sales in order to enhance their revenue. Therefore developing a production plan has become the important issue for the manufacturers. A well- developed production plan not only can reduce the production cost of tea, but also can enhance coordination between machines and employees, resulting in little waste of production resources. A good production plan relies on the accurate prediction of tea sales that is mainly dependent on consumer behaviors. However, the relationship among sale of products, consumers and other products is very complex. To address the issue about the prediction of tea sales, the traditional data analysis is not easy to predict the sales in the future market. Therefore manufacturers are looking forward to a better decision support systems for prediction. This study used tea sales data from point of sale (POS) in convenience store chain. We used data mining methodology to construct sales prediction model. The prediction method in the model is based on statistical time series moving average (MA) and autoregression integrated moving average (ARIMA). We also used neural network based on back-propagation network with changes in causal parameters for prediction models: One includes Granger causality test and the neural network ; another one includes Granger causality test, the neural network and more, the autoregressive (AR). We used actual sales data to assess these prediction models. Our results have shown that using Granger causality test plus autoregressive within back-propagation neural networks has the best prediction for these data.