标题: 以时间序列对台湾电脑公司进行商业解析及绩效预测
Using Time Series to Conduct Business Analytics and Performance Prediction for Taiwanese Computer Companies
作者: 郑浩宇
王志轩
Cheng, Hao-Yu
Wang, Chih-Hsuan
工业工程与管理系所
关键字: 台湾电脑产业;变数筛选;时间序列;商业解析;绩效预测;动态回归;Taiwanese computer industry;feature selection;time series;business analytics;performance prediction;dynamic regression
公开日期: 2017
摘要: 从电脑问世之后,大家对它的依赖程度与日俱增,但是现阶段却面临着不少的挑战。本研究引用1992年宏碁董事长施振荣先生提出微笑曲线的特性运用在电脑产业,依其特性将产业区分为工业电脑、代工电脑与品牌电脑,透过变数筛选后结合时间序列的分析来预测未来的绩效每股盈余。第一阶段使用商业智慧当中的资料采矿技术及回归方法,分别为分类树、随机森林、多元适应性云形回归与复回归四种方法,使用平均平方误差、平均绝对误差及平均绝对百分差作为评比标准并进行交叉验证,选择出四种方法中误差最低的一种后,进行各子产业的重要变数筛选。并由代表性公司研华、广达、技嘉作为研究对象,首先透过Granger Test检定变数是否有时间递延的存在,分别使用复回归分析、多元适应性云形回归与时间序列中的ARIMA与动态回归对公司建模,最后根据三种误差评判的标准,选出最好的模型,然后针对个别公司进行未来一年的绩效预测及商业解析。
研究结果除了证明挑选的变数能反映产业特性外,也发现动态回归与ARIMA两种方法建模的误差优于多云适应性云形回归及复回归,而动态回归又略优于ARIMA,因此本研究将根据动态回归的模型进行绩效预测,提供给企业及投资人参考。
Since the computer was invented, people have relied on it heavily. However, the challenges it faces are more than ever. This study divides the Taiwanese computer industry into three business model: industrial personal computer (IPC), electronic manufacturing service (EMS) and original brand manufacturing (OBM). The first phase is to extract significant performance indicators by Random Forest and identify the causations between predictors and outcomes by Granger Causality Test. Next, using Dynamic regression to incorporate the temporal impacts of leading predictors on lagging outcomes into the process of performance prediction.
The results indicate that the variables selected by random forest are representative. Besides, the three indices, mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), indicate that each model, dynamic regression in particular, is robust enough. As a result, we believe the forecasting and confidence interval is indeed reliable so that this reference can be an information for executives and investors.
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070453339
http://hdl.handle.net/11536/141025
显示于类别:Thesis