以數位口碑為基礎之流行性商品銷售預測

Abstract

市場瞬息萬變，企業必需不斷調整腳步，才有機會取得先機，而企業的採購與生產計劃源頭即是銷售預測，足見銷售預測的重要性。即使如此，企業在銷售起伏較大的流行性商品上的銷售預測仍不是非常準確。 從文獻探討中可知，口碑是影響消費者購買的動機之一，消費者可由身邊的親朋好友口中獲得口碑，亦可利用網際網路取得數位口碑。但過去只對口碑形成的因素與影響層面做研究，而未實際量化口碑對銷售數字的影響，故本研究欲將數位口碑實際應用在銷售預測上。 首先自動抓取網路中的商品相關討論文章，利用簡單貝氏分類器判斷文章的評價，並將其轉換成量化評價以用作預測。除了建立多個目標函式以釐清口碑與銷售量的關係外，同時亦嘗試各種時期的口碑作為自變數，以決定口碑與銷售量的期間差距為何。最後找出所有組合中預測誤差最小，且具效度的模型用作預測，以提升流行性商品銷售預測的準確度。 以台灣知名連鎖藥妝店的十大熱銷商品為樣本，發現本模型適用具話題性、能在網路上引起足夠的討論量商品，且其預測誤差皆低於常見的移動平均法、指數平滑法、趨勢指數平滑法等。可知將本模型應用於流行性商品銷售預測時，可有效地增進預測準確度。

Since market is constantly changing, companies need to continuously adjust the pace to get a head start. Forecasts are essential to the business’s decision making and planning processes. Better forecasting can contribute to better price structuring and better inventory management. However, it is a challenging problem owing to the volatility of demand which depends on many factors. And the situation is prominent in fashion product due to its sales versatility. Past research shows that disseminating information through word-of-mouth communication is one of the most effective mediums for relaying important product and company information. It not only plays an important role in the evaluation of products but also plays an important role in society as well. Although many companies have found its effectiveness through lots of literature reviews, most of them are limited in focusing on the the frequency and types of word-of-mouth behavior, or the effects of word-of-mouth behavior on product evaluation. Few of them discuss the relationship with sales volume. In this study, an automatic mining approach is proposed to resolve the aforementioned issues. According to this method, a text mining technique and Naive Bayes classifier will be used to determine the rating of each product-related article extracted from the Internet. Based on the regression model, some target functions have been designed to clarify the relationship between the rating of world-of-mouse and the sales. And a valid forecasting method is generated with the smallest prediction error. Performances of our model are evaluated by using real data from a cosmetic retailer in Taiwan. The experimental results demonstrate that this model is especially suitable for the fashion product with sufficient discussion on the Internet. In addition, our proposed method is proved to outperform several traditional sales forecasting methods such as moving average, exponential smoothing and exponential smoothing with trend. Therefore, we believe that this model can effectively enhance the prediction accuracy when applied to fashion products.