台灣咖啡銷售量之分析與預測

Abstract

中文摘要

為了瞭解目前國內咖啡市場銷售量的變化與趨勢，本研究使用經濟部統計處的「國內咖啡銷售量」的資料，其時間從2000年1月至2019年12月，共240筆月資料，並將此資料分成樣本內期間以及樣本外期間，以分析樣本內期間國內咖啡月銷售量的狀況，並保留樣本外期間的最後十二筆作為預測的對照依據。此份報告使用四種時間序列分析的模型/方法，依序為ARIMA ( Autoregressive Integrated Moving Average )模型、ARIMA之介入分析、時間序列迴歸法及指數平滑法，對樣本內期間配適預測模式，檢驗預測是否準確，最後採用三種評估準則，分別為RMSE ( Root Mean Square Error )，MAE ( Mean Absolute Error )以及MAPE ( Mean Absolute Percentage Error )，來評估何者為最佳之模型/方法。結果發現，相乘之指數平滑法的預測表現最良好。此外，由於位居連鎖咖啡排行冠軍的星巴克在1999年開始至2005年進行多元化的經營模式，更是在2002年11月，開始正式推出方便攜帶又免用現金的星巴克隨身卡，促進咖啡市場擴張。故本研究考慮到會受到外在因素的影響，因此也將介入函數放入模型中。結果發現，樣本內期間選擇時間序列迴歸法來配適模型的預測表現為最佳，樣本外期間選擇相乘之指數平滑法來配適的預測表現為最佳。

Abstract

In order to understand the consumption of coffee in Taiwan, we analyze the data of "domestic coffee sales volume" obtained from the Statistics Department of the Ministry of Economic Affairs. We collect monthly coffee sales volume from January 2000 to December 2019 (a total of 240 observations) and devide our data into two periods: in-sample period and out-of-sample period. We use four time series forecasting models/methods including ARIMA model, ARIMA model with intervention analysis, time series regression, and exponential smoothing method to fit the data during the in-sample period. We then test the accuracy of the out-of-sample prediction using three evaluation criteria, namely root mean square error, mean absolute error, and mean absolute percentage error. The results show that the prediction performance of the exponential smoothing with multiplicative holt-winters method outperforms other alternative models/methods. We notice that Starbucks, the champion coffee chain, launched a portable and cash-free Starbucks card in November 2002. Thus, we consider the influence of external factors on target time series by adding the intervention function into the examined models. The results indicate that selecting the time series regression to fit the in-sample period is the most preferable, and choosing the exponential smoothing with multiplicative holt-winters method brings the best prediction for the out-of-sample period.