標題: 結合資料解析與時間序列進行台灣半導體公司之績效預測Combining Data-Analytics and Time Series to Forecast Semiconductor Companies of Taiwan 作者: 姜穎憲王志軒Chiang, Ying-HsienWang, Chih-Hsuan工業工程與管理系所 關鍵字: 半導體產業;隨機森林;資料採礦;格蘭傑因果關係;時間序列;semiconductor industry;random forest;data-mining;Granger causality test;time series 公開日期: 2017 摘要: 台灣半導體產業在世界上占據了舉足輕重的地位，半導體產業的總產值逐年增加，並且多年間都處於全球第二位，僅次於美國，然而整個半導體產業的公司卻因為中國半導體產業的崛起而使數量卻年年遞減，近年來半導體產業已經形成大者恆大、小者越小的兩極差異；因此越來越多的中小型公司希望透過資料解析來得知公司的營運以及績效指標，藉此改善公司的營運狀況；而大企業也希望利用商業解析促使自己更進一步的成長並立足國際。 本研究將分成兩大面向；第一以產業為面向，將產業以及隨機森林個別找出半導體設計、製造、封測產業中會對每股盈餘有重要影響的關鍵性變數，並以各個子產業別的代表性公司(聯發科技、台積電、矽品)，以複線性迴歸(MLR)、多元適應性雲型迴歸(MARS)等兩種方法預測公司的每股盈餘；第二以時間序列為面向，利用格蘭傑檢定法來判斷重要變數是否對每股盈餘有落後時間的影響，並且以單根檢定判斷各個公司是否符合穩態性，再分別以差分整合移動平均自迴歸模型(ARIMA)以及向量自迴歸模型(VAR)等兩種方法，透過考量落後期數的重要變數來建立時間序列模型，並利用平均平方誤差(Mean Squared Error, MSE)、平均絕對誤差(Mean Absolute Deviation, MAD)以及平均絕對百分比誤差(Mean Absolute Percentage Error, MAPE)等三指標衡量模型的好壞，再利用其模型預測三家代表性公司的未來績效預測，並評估公司未來之營運狀況。Taiwanese semiconductor industry has occupied an important position in the world. Total value of out-put of semiconductor of Taiwan follows behind the United States. Number of Taiwanese companies, however, is fewer and fewer because of rising of Chain. From now on, some of companies which have scale are extremely higher than others. Thus, the small and median scale companies would like to figure problems out by data-analytics and KPIs. In other hand, large corporations would like to improve themselves for worldwide market as well. This research has two parts. In the first place, this research would use random forest to find out important variables which highly affected EPS. In addition, this research would use MLR and MARS models to predict leading companies which included MTK, TSMC and SPIL. In the second place, this research would use Granger causality test and unit root test to check if important variables are lagging for EPS and if models are stationary and use ARIMA and VAR models to predict each leading companies and check model fitting by MSE, MAD and MAPE. At last, this research would forecast the future EPS of leading companies by ARIMA and VAR models and evaluate the operation of companies. URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070453301http://hdl.handle.net/11536/140934 顯示於類別： 畢業論文