晶圓代工廠的需求預測模型-以ARIMA模式分析-

Abstract

在以製造業為主的台灣產業結構裡，「需求預測」這件事情不論是在各行各業都缺少不了，需求預測影響擴及企業財務規劃、庫存管理、生產計劃、配銷計劃、行銷計劃及客戶管理等各層面，因此在企業每日、每週、每月、每季及每年的大大小小會議裡，需求預測都是一個最關鍵也是最基本的數字，不論這個數字是如何產生，大家都只有一個要求，那就是愈準確愈好。那麼不準確的需求預測會帶來什麼樣的結果呢？過度樂觀（預測＞實際）會造成過多的庫存、庫存成本資金積壓、產生轉運成本、報廢核銷庫存、降低獲益率，過度保守（預測＜實際）會造成訂單修改成本、較高的產品成本、錯失銷售機會成本、錯失搭售產品的銷售機會、客戶滿意度降低。因此如何提高需求預測的正確性，成為企業經營中一個重要的課題，因為它會直接影響到企業經營的績效，不過也因為它的不確定性，造成在決策上的因難。

要如何管理這個不確定性呢？對於具備歷史資料的產品來說，利用過去的資料進行預測是最簡便的方法，因為只要有一位瞭解時間序列預測模式的分析人員，就可以產生客觀的預測結果。鑒於此，本研究應用時間序列分析法之ARIMA預測模式，以單變量及多變量方法探討樣本晶圓代工廠未來之晶圓需求量，利用1997年1月至2007年5月晶圓之歷史銷售量為樣本內資料，提出該晶圓廠未來需求量之ARIMA單變量及多變量預測模型，並將此需求預測模型所產生之預測結果與2007年6月至2007年12月之該晶圓代工廠晶圓銷售數字(樣本外資料)做比較，結果顯示，本研究所使用之ARIMA預測模型由MAPE值判斷可得到合理之預測績效，整體而言也會比現行樣本晶圓代工廠之業務單位、區域規劃單位、總部規劃單位之需求預測績效來得好，足見其確實可幫助企業提高需求預測的準確度。

在真實的世界□，要百分之百準確的預估未來幾乎是不可能的事，但我們仍可以藉由各種之需求預測模型的分析，在日常的管理中，盡力將預測曲線推向準確的方向，長期下來，將可提高公司競爭力與獲利能力。希望透過本論文之實證研究結果及相關之需求預測模型，能做為產業界在從事需求規劃時的一個參考，進而能提昇晶圓代工產業之需求預測品質，增進供應鏈管理之效率。

In the industrial structure of Taiwan, demand forecast is very critical and necessary for business financial planning, inventory management, manufacturing plan, distribution plan, marketing and customer service and the forecast number is the basic and crucial item for every person on daily operation or meeting. No matter how the forecast number comes from, the only requirement is its accuracy. How it will happen when the accuracy of demand forecast is low to a company? As we can understand, Over-forecast will induce many unnecessary inventories and operation cost. Under-forecast also may have a company lose its opportunities and customer satisfactions. For the reason of that, it is an important topic to improve the accuracy of demand forecast and management the uncertainty of demand.

How to manage the uncertainty of demand? Many studies focus on analyzing historical data and using time series forecast model to do demand forecast to management the uncertainty of demand. Hence, the main purpose of this research is to find out a useful ARIMA model by adopting the company’s historical sales data from Jan. 1997 to May 2007. The forecast performance of the ARIMA model in this research is evaluated by MAPE value comparing to the company’s actual sales volume for the following several months (2007.6~2007.12). Based on the MAPE value, the outcome or the forecast performance of this research is very reasonable and accurate. Also, it has better performance than the forecast performance of Sale Division, Regional Planning Division and Headquarter Planning Division in the company.

In the real world, it is impossible to have a absolutely precise forecast result. But we can do our best to make the distortion smaller by using forecast models and thus company will improve its competitiveness and gain more profit. Hopefully, this research can be helpful to all the wafer foundries when they are working on demand forecasting and planning.