結合統計迴歸及機器學習對台灣半導體產業之績效分析與模擬

Abstract

台灣的半導體產業在全球排名數一數二，但近年來受到中國紅色供應鏈以及全球經濟發展趨緩的影響；因此台灣半導體產業不論是在上游的IC設計、中游的晶圓代工、下游的封裝測試，都正面臨巨大的轉變，而公司也想藉由企業績效診斷從激烈的市場中存活下來，且從中取得更多的獲利，現今不論從各產業的大公司都積極著墨商業智慧的領域，可知其中龐大的商機。本研究以平衡計分卡為架構來衡量台灣半導體產業在財務、學習與成長、顧客及內部流程構面的表現，在此四大構面之下總共高達25項的變數，因此找出會影響每股盈餘(EPS)及總資產報酬率(ROA)的關鍵變數為優先重要的課題。本研究結合統計迴歸與資料探勘的方法進行變數篩選，在迴歸的部分則採用複迴歸(MLR)、迴歸決策樹(CART)、多元適應性雲型迴歸(MARS)、隨機森林(RF)四種方法來預測非線性結構的資料。研究結果發現迴歸決策樹及隨機森林的表現最佳，故利用此兩最佳方法找出的重要變數交集挑選出對影響每股盈餘及總資產報酬率的重要指標前五名，最後再透過機器學習中支撐向量迴歸(SVR)來針對此五重要變數進行改善後的績效模擬，並提出公司績效管理的意涵。

Taiwanese semiconductor industry has occupied an important position in the world. Number of Taiwanese companies, however, is fewer and fewer because of rising of Chain. Meanwhile, many software vendors are spending their efforts on business intelligence to help firms keep more competitive and earn more profit. In order to help Taiwanese IC design industries find key performance indicators (KPIs). In other hand, large corporations would like to improve themselves for worldwide market as well. This study uses the balanced scorecard (BSC) to set 25 predictors. In particular, a two-phase approach is adopted to conduct feature selection. In performance regression, Multiple Linear Regression (MLR), Multivariate Adaptive Regression Splines (MARS), Classification and Regression Tree (CART), and Random Forest (RF) are compared. Then uses the best performance regression which has the lowest Mean Squared Error (MSE). Furthermore, performance simulation is conducted by using Support Vector Regression (SVR), respectively.