新商品銷售預測解決方案之研究

Abstract

在競爭日益激烈的環境下，企業常常透過開發新商品來開拓或鞏固市場，並提升競爭力。然而新商品必須能快速反應市場需求，不準確的新商品銷售預測會導致錯誤的商業決策，造成過多存貨或嚴重缺貨，甚至使得企業失去市場競爭力。因此，準確的新商品銷售預測能提供經理人適當的資訊以作出正確決定，進一步創造競爭優勢。 由於新商品缺乏銷售資料，進行銷售預測時，大部份傳統時間序列預測方法都不能使用，所以只能依賴經理人的經驗，主觀地決定未來需求。因此本研究提出一套新商品銷售預測流程，協助對新商品缺乏經驗的經理人客觀地為每種新商品決定要使用的預測方法及參數。本研究亦從流程的功能面建構一新商品銷售預測解決方案，以解決新商品銷售預測中的各種問題，並建立智慧學習平台，根據最新相關資訊，在六個預測方法中自動選取最佳方法及參數，進行未來的銷售預測。 本研究提出之解決方案，包括資料處理啓發式演算法、修正的預測方法及啓發式學習演算法。資料處理啓發式演算法，目的為整理新商品之資料，同時在新商品資料不足時，引用分類架構、試賣活動等資料進行調整；解決方案中採用六個預測方法，包含時間序列預測方法及啓發式預測方法，但啓發式演算法有許多使用情境的限制，因此本研究對這些方法作出修正，讓他們能適用於更多情境；啓發式學習演算法，目的是根據已建立的相關法則，考量各種因素，在適用的方法中為新商品決定最佳之預測方法及參數。 在情境分析中，本研究所提出之解決方案有相當良好的表現，與一般實務上所採用的一期或多期移動平均作比較，本研究之解決方案在大部份情境下皆有較準確的預測結果。此外，為了測試本解決方案的可行性，本研究分別針對茶產業、彩妝產業及清涼飲料產業的成功新商品之真實資料進行測試，結果顯示本研究之解決方案的預測結果顯著優於一般預測方法。因此，本研究之新商品銷售預測解決方案能適用於各種產業。

Developing new Product is not only a competitive strategy for a business to enlarge its market share but also a survival strategy required to respond quickly to the market. Inaccurate forecasts for new products may lead to a false business decision that results in a large inventory or a significant shortage, which then costs a corporation to lose its competitive ability in the market. Therefore, an accurate sales forecast for new products is very important for managers to deal with uncertainty and make the right decision. Two issues need to be addressed in the problem of new product sales forecast: limited data and selection of forecasting model. In the past, new product sales forecast usually relies on the experience and the expertise of a manager to select the forecasting model with only limited data available. Instead of human judgments, this study proposes a New Product Forecast Procedure (NPFP) and a New Product Forecast Solution (NPFS) to solve the new product sales forecast problem. NPFP and NPFS together provide an auto-learning platform to select the best parameter for each of the six forecast methods and finally the best forecast technique using rule-based weighted MAPE standard. The future forecasts will then be made according to the selected method. NPFS includes four modules: Data Handling, Forecast Model, Learning, and Forecast Calculation. A five-step Heuristic Data Handling Algorithm (HDHA) is developed to clean and generate data for the new product using classification and pretest data. The Forecasting Model module includes three classical and three modified heuristic forecast methods. Each method has its own applicable scope, so that the solution can be applied to different kinds of sales pattern. The Learning module chooses the best forecast method according to the Heuristic Rule-based Learning Algorithm (HRLA) that identifies applicable methods and considers the number of actual data points. Finally, Forecast Calculation Module computes forecasts for new product using the selected best method. Among most of the twenty seven designed scenarios, the NPFS has better performance than the methods that widely adopted in practice, such as moving average. The NPFS is also applied to three real-world cases for products of tea, cosmetics, and soft drinks. In each case, several successful new products in different classifications are identified. As expected, the NPFS has better performance than commonly used methods. In conclusion, the NPFS can improve the accuracy of new product sales forecast and can be easier adopted that the commonly-used methods without relying too much on human judgments.