本文提出以k個鄰近樣本法為基礎的模式，利用預售資料變動型態來預測一項服 務的最終銷售量。我們根據現行預售規則建立期望銷售曲線，針對單一列車服務品項(班 次與服務起迄點之組合)進行數值波動與時間特性之關聯分析，發現時間特性對於曲線 分佈型態有明顯影響。據此本研究整合群集分析與k個鄰近樣本法之優點提出一個新 的預售-銷售預測流程，並與四種傳統方法進行兩種資料維度間的預測绩效比較。結果 發現本研究所提出的整合模式在服務需求低且距離服務提供時間近的情境下預測绩效 表現明顯優於其他競爭模式。然當情境轉移至服務需求高且具服務提供時間近時，則 以傳統的銷售-預售模式預測绩效表現較佳。而當情境轉移至當距離服務提供時間有一 段期間時，無論需求高低則以傳統模式表現優於銷售-預售模式。因此在不同資料情境 採用不同類型預測模式會比使用單一預測模式獲得較佳的绩效。此一結果說明管理者 必須對其服務需求型態有足夠的認識並據此選擇適用的預測模式。

This paper proposes a k nearest neighbor model to predict service sales volumes by using reservation information. The concept of expected sales curves is first introduced based on the reservation rules of the studied case. The analysis of curves indicates that day-of-week, holidays, and vacations are crucial factors for influencing curve patterns. In order to deal with the heterogeneity, cluster analysis is applied to divide raw data into several groups before constructing forecasting models. After that, the proposed k nearest neighbor method is utilized for modeling a specific group of curves and the predictive performance is compared with that of four conventional benchmarks in terms of two dimensions (days before service time and demand strength). The empirical study shows that using one particular model in all data scenarios is not an appropriate decision. Managers should select suitable forecasting models from a bunch of potential alternatives based on the understanding of their services/products in order to achieve promising predictive performance.