| 摘要: | 供需失衡導致NAND型Flash價格驟變，半導體供應鏈中的所有成員都期望能夠掌握市場脈動及價格變化的趨勢，從業人員僅能從該產業中的專家實際經驗，參考相關佐證資料，以主觀或輔以實際資料的方式進行簡單分析後進行價格趨勢分析，爾後據此決定該產品的採購及存貨政策，但通常所產生之結果較容易產生偏差及在缺乏科學依據的情況下，讓公司的資產無形曝露於風險之中。 本研究欲發展出簡單且有效的預測模型，幫助業界對NAND型Flash價格及下游交運量在非線性趨勢及極端不規則變化之下，對價格走勢作出正確的預測，搭配對下游客戶的交運量預測，藉以擬定相關避險及因應策略，持續增加企業的獲利水準及營運效能，降低公司蒙受損失的可能性。 在價格預測的實證中，在2004年及2005年間，其原始數列呈現穩定的趨勢變化時，以時間序列分析法中的天真預測法、移動平均法及指數平滑法及灰色預測GM(1,1)模型預測該年10月至12月的價格變化時，皆發揮不錯的預測效果，其MAPE值皆低於10%以下，達到高準確度的預測能力，且維持不錯的獲利水準；在2006年間，原始數列呈現不規則地變動時，各預測模型在預測該年10月至12月的價格變化時，發生較大的差異，其中以指數平滑法( =0.3時)預測效果最好，賺得368.50仟美元的利潤，但MAPE值仍高達27.72%，本研究透過預測組合的方式將該期各模式最佳的預測值予以平均，如此求算的結果可將MAPE值降低至23.34%。在交運量預測的實證中，因交運量的變化起伏很大，明顯高於價格趨勢之變化，各模式在預測各年10月至12月的交運量準確性上並不理想，其MAPE值皆高達20%以上。 將上述驗證結果，藉由瞭解管理當局的風險承受程度，透過風險分析，制定該產品的風險評估矩陣，持續監控高風險的部份，藉由決策的運用，以期能在最低風險下追求最高的獲利水準。後續在應用上，將各預測模式建制在資訊決策系統中，利用該公司Oracle資料庫之歷史數據加以運算分析，藉以計算出其預測值，除可用以比較分析各預測的結果，加以解釋及了解差異外，進而能夠運用其中有用的資訊並預測出每日、每週的價格變化情形，透過集體小組的意見及討論來掌握每個趨勢轉折點，據以掌握先機。 Because the imbalance of supply and demand often leads to the sudden change of the Nand flash prices, members of the semiconductor supply chain hope that they can control the market volatility and the change of price trend. However, they merely perform some simple analyses according to the experiences of experts in the industry or with their own subjective judgment in order to find out the price trend. They often determine their purchase and inventory policy based on these simple analyses. However, such analyses usually lack of scientific basis and decisions based on which could result in error and expose a company asset to risks. The purpose of this study is to develop a simple and effective pattern of price forecast, and to help the people in this industry make an accurate forecast of both the price trend of the Nand flash and its delivery amount to final customers under the situation of extremely irregular changing trend, and thus, make a policy to avoid risk, continuously maximize company profit, improve operational efficiency, and reduce the possibility of losses. In the price forecast analysis, evidence shows that, in 2004 and 2005, when the original figures presented a stable trend change, among all time-series analysis methods, the naïve method, moving average method, exponential smoothing method and grey prediction of GM(1,1) method had good results in the forecast of price change from October to December of the year. The MABE value was lower than 10% which indicates a high forecasting accuracy,and maintain a relatively high level of profit. When the figures aroused irregular changes for the period of October to December 2006, various forecast methods indicated larger differences in the forecast results. Among them, the exponential smoothing method ( =0.3) had best forecast ability,and earning a profit of US$368,500, but the MAPE value still reached to 27.72%. By combining the forecast methods, this study averaged the forecast values of various kinds of forecast methods, and the MAPE value can be reduced to 23.34%.The results of the delivery quantity forecast analysis in this study indicates that, due to the delivery quantity is highly volatile, evidently higher than the change of price trend, the accuracy of various kinds of forecast in the forecast of delivery amount from October to December of the year was far from ideal. All MAPE values were higher than 20%. According to the foregoing findings in this study, a hazard estimation of a product is formulated after performing hazard analysis and giving fully consideration of the management’s risk-bearing capacity. Through continuous monitoring on highly hazardous parts and mature policy making, it is expected that highest profits can be achieved under the lowest hazardous conditions. In actual applications and future practice, various kinds of predictive models should be established in the policy making data system, and the historic figures in the oracle data system of the company can be utilized for computation and analysis to derive the prediction value. In addition to comparing and analyzing the various prediction results to find out their differences, the daily and weekly trend of price change can be predicted and controlled through the review of the collective team members on the point of every trend change. |
