半導體產業具有供應鏈前置時間長、外部產能依賴高與需求波動大等特性。對 IC 設計公司而言，業務端提供之需求預測不僅用於銷售規劃，也會影響晶圓投片、備料、封裝測試、在製品安排與交期承諾。若原始需求預測未能即時反映實際需求變化，可能造成需求高估、庫存累積、產能占用與供應鏈資源錯配。因此，如何在不取代既有需求預測流程的前提下，利用歷史資料建立預測修正機制，成為值得探討之議題。 本研究以 A 公司 2022 年至 2026 年之週別需求預測歷史資料與實際銷貨資料為基礎，建構需求預測修正模型，並據以辨識高估暴露。研究首先將不同年度、來源廠區與產品類別之預測資料進行整理與對齊，再以原始需求預測與最終實際銷貨量之差距作為修正目標，建立線性模型、懲罰式迴歸、Huber 穩健迴歸、樹模型與長短期記憶網路序列模型，並以 2026 年資料進行時間外驗證。 研究結果顯示，原始需求預測存在明顯誤差與高估暴露。各模型比較中，Huber 選擇性修正模型表現較穩定，以 2026 年資料進行時間外驗證，整體加權絕對百分比誤差（WAPE）由 1.1956 降至 0.9033，改善約 24.45%，並同時降低大額誤差比例與需求高估所造成之供應鏈資源暴露。研究進一步依實際銷貨資料來源拆分發現，模型改善主要來自對未實現需求高估暴露之辨識與下修。整體而言，本研究提出之模型可作為企業需求預測複核與供應鏈資源配置之輔助檢核機制，協助管理者在投片、備料與產能安排前，辨識未實現需求、高估暴露與大額預測誤差風險。;The semiconductor industry features long supply chain lead times， heavy reliance on external capacity， and volatile demand. For an IC design company， demand forecasts drive not only sales planning but also wafer starts， material preparation， assembly and testing， work-in-process arrangement， and delivery commitment. If the original forecast fails to reflect actual demand changes， it may cause over-forecasting， inventory accumulation， capacity occupation， and supply chain resource misallocation. Building a data-driven forecast correction mechanism without replacing the existing forecasting process is therefore an important issue. This study uses A Company’s weekly demand forecast histories and actual sales data from 2022 to 2026 to develop a demand forecast correction model for identifying over-forecast exposure. The original forecast is aligned with actual sales， and their gap is defined as the correction target. Linear models， regularized and Huber robust regression， tree-based models， and a Long Short-Term Memory sequence model are compared， with 2026 data used for out-of-time validation. The results show that the original forecast contains substantial error and over-forecast exposure. The Huber selective correction model is the most stable: in the 2026 out-of-time validation， the overall weighted absolute percentage error (WAPE) drops from 1.1956 to 0.9033 (about 24.45%)， while also reducing the large-error ratio and the over-forecast resource exposure. The source-split analysis shows the improvement mainly comes from identifying and down-correcting over-forecast exposure related to unrealized demand. The proposed model can thus serve as an auxiliary validation mechanism for demand forecast review and supply chain resource allocation， helping managers spot unrealized demand， over-forecast exposure， and large forecast-error risks before committing wafer starts， materials， and capacity.
