應用機器學習與深度學習預測台股指數

Apply Machine Learning and Deep Learning to Forecast Taiwan Stock Exchange Index

陳鼎玟(DING-WEN CHEN)

指導教授 ： 周宗南

朝陽科技大學/管理學院/財務金融系/碩士(2018年)

摘要

隨著金融科技的浪潮與人工智慧的興起，金融市場早已瀰漫著前所未見的變革聲浪，而數據的應用也已成為眾多行業生存的關鍵策略，本研究目的主要是探討景氣指標對股價指數之漲跌影響，以台灣地區國家發展委員會(以下簡稱國發會)每月公布之景氣指標為研究變數，並將變數標準化與離散化兩種方式呈現，透過決策樹、模糊決策樹、模糊遺傳規劃以及深度學習等模型來預測台灣加權股價指數之漲跌關係。本研究發現在標準化與離散化中，景氣指標變數所預測出整體準確率均可達到60%，其中，在標準化的變數中以模糊遺傳規劃之整體預測準確率最高可達66.37%，顯示有66.37%可預測股價之漲跌，而在精確率中，也以模糊遺傳規劃為最佳結果，顯示在標準化所建構之模型以模糊遺傳規劃為最佳。透過將變數離散化後以決策樹整體準確率為最高達至64.6%，其上漲精確率為70.2%，以深度學習精確率為最佳達78%，表示模型預測中有78%的機率能正確預測股價呈現上漲。經由標準化與離散化進行比較分析，其研究結果發現以遺傳規劃為最佳預測結果。

關鍵字

機器學習 ； 深度學習 ； 股市預測

並列摘要

The financial market has produced unprecedented changes along with the rapid development of Financial technology and Artificial intelligence, and Big data analysis has already become a key strategy for survival of numerous industries. The purpose of this study is to explore the impact of the prosperity index on the stock price index. The monthly economic indicators released are research variables by the National Development Council of Taiwan. The variable standardization and discretization are presented, which predicts TAIEX through the model of decision tree, fuzzy decision tree, fuzzy genetic programming and deep learning.
This study found that standardization and discretization, the overall accuracy rate of the indicator of the prosperity index can reach 60%. Among the standardized variables, the overall prediction accuracy of fuzzy genetic programming is up to 66.37%, which shows that 66.37% can predict the rise and fall of the stock price. In the accuracy rate, the fuzzy genetic plan is also the best result. The model constructed in the standardization is optimal for fuzzy genetic programming.
Through the discretizing variables. By discretizing the variables, the overall accuracy of the decision tree is up to 64.6%. The accuracy of the increase is 70.2%, and the accuracy of the deep learning is 78%, which means that 78% of the model predictions are correct. The stock price is forecast to rise. Through comparative analysis of standardization and discretization, the results of the study found that genetic programming was the best predictor.

並列關鍵字

Machine Learning ； Deep Learning ； Stock Prediction

參考文獻

一、 中文部分

李沃牆、聶建中（2009）。遺傳規劃決策樹模型於房貸提前償還之風險管理。住宅學報，18(1)，63-87。

李承璟（2004）。台股趨勢與景氣指標之分析。(未出版碩士論文)，國立台灣大學經濟學研究所，台北。

游皓麟（2017）。實戰R語言預測分析（第一版）。台北：電子工業出版社。

張愷凌（2009)。景氣循環、總體經濟變數與台灣股價指數的關係性研究。(未出版碩士論文)，國立交通大學管理學院財務金融學程，新竹。