以自我回歸移動平均模型與標題熱門程度分析預測雜誌銷售量

Abstract

台灣的便利商店擁有相當大的市場，許多廠商紛紛與便利商店合作，企圖藉此拓展自己產品的市占率。但便利商店店面大小有限，因此需要精確的銷售量預測系統，以減少發生庫存成本增加以及產品短缺的機會。學者發現，顧客在購買產品前會透過網路找尋產品評價等行為。因此許多研究者透過Google的網路搜尋量預測消費者需求。但過去的研究其搜尋指數是由人事先定義，非以產品內容為依據。 本研究提出自我回歸移動平均模式中加入雜誌內容熱門指數的方法進行銷售預測：產品的內容是預測顧客需求的重要因素，透過Google搜尋可得到內容的熱門搜尋指數，熱門指數越高者表示其內容越能吸引顧客購買慾望；本研究並且提出比較不同時期雜誌文章中的標題字熱門搜尋指數之方法，以解決Google-Insights在搜尋時字數上的限制。 實驗結果顯示本研究所提出加入內容熱門指數的方法比傳統自我回歸移動平均模型方法能更能有效改善預測的準確性。

Convenience stores have a huge market in Taiwan and many providers want to cooperate with them. However, stores’ space is limited, and a high quality prediction is required to reduce resource waste. An accurate sales forecasting can reduce inventory costs and shortage, and is helpful to plan profit levels and capital needs. Before customers make decision, they search product via Internet. Many studies investigate how to apply the search indexes derived from the Google insights or Google trend data to improve predictions in different domains such as housing prices/sales or travelers. However, most studies use some search terms, which are related to the predicting products and are manually decided, rather than the contents of products to derive the search indexes. The search indexes of the terms manually decided may be limited in approximating consumers’ preferences. In this work, we utilize the popularity of celebrities words appeared in article titles to predict the sales of magazine, and the relative popularity of title words needs to be measured for all the title words of magazine articles across different issues of magazine. The search indexes of title words derived from Google insights or Google search engine denote consumers’ interests in the contents of the magazines. However, only five search terms are allowed to obtain their relative search volumes from Google insights. To solve this problem, we propose a method to compare the search volumes of multiple words across different issues of magazine to measure the popularity of title words. The popularity index of each magazine issue is derived from the popularity of title words appeared in that issue. We integrate ARMA time series model with the popularity indexes of magazines to enhance sales forecasting. Our experiment results show that our proposed approach outperforms conventional approaches (ARMA) and can improve prediction accuracy by using the popularity indexes of magazines.