利用機器學習配置台灣指數多空模型

Using Machine Learning to Build Taiwan Stock Index Bull and Bear Model

邱正凱(Zheng-Kai Chiu)

指導教授 ： 石百達

共同指導教授 ： 莊文議(Wen-I Chuang)

國立臺灣大學/管理學院/財務金融學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202001097

摘要

本篇論文透過主成分分析(principal component analysis, PCA)的方式，處理預測變數，並利用簡單線性回歸模型(simple linear regression)計算個股的期望報酬率，並使用下而上(bottom-up)的方式，建構市場期望報酬，並以此計算市場期望指數，搭配技術分析的方式，通過乖離率、布林通道等核心概念，配置台灣加權的多空頭模型。並透過機器學習的方式，以支持向量機迴歸(support vector regression, SVR)，優化預期報酬的計算，並提升多空模型的績效表現。並以此作為十三個產業進出場的時間點，檢視各產業的報酬以及勝率表現。
十三個產業名單如下：塑膠類(M1300)、紡織纖維(M1400)、化學工業(M1721)、生技醫療(M1722)、半導體(M2324)、電腦及週邊設備(M2325)、光電(M2326)、通信網路葉(M2327)、電子零組件(M2328)、電子通路(M2329)、其他電子(M2331)、金融保險(M2800)、油電燃氣(M9700)。
在市場多頭的部分，採用台灣技術指標以及美國經濟數據做為預測變數，其配置出的模型，透過SVR優化後的結果，其勝率可達100%；在市場空頭的部分，分為台灣技術指標及美國經濟數據個別做為預測變數的兩個模型，經SVR優化後的結果，在前者的部分，勝率可達100%，後者則為71.43%。在此，空頭的部分，採取聯集的方式，作為進出產業的時間依據。
在產業多頭模型表現上，除紡織纖維(M1400)、其他電子(M2311)、油電燃氣(M9700)、生技醫療(M1722)，其餘產業皆可達80%以上之勝率；在產業空頭模型表現上除電腦及週邊設備(M2325)、電子通路(M2329)、半導體(M2324)、油電燃氣(M9700)、生技醫療(M1722)、通信網路業(M2327)，其餘產業皆可達75%以上之勝率。

關鍵字

主成分分析 ； 預期報酬 ； 布林通道 ； 機器學習 ； 支持向量機迴歸

並列摘要

This thesis used principal component analysis (PCA) to process predictive variables, and used simple linear regression models to calculate the expected return of stocks, and used bottom-up method to construct market expected returns. In this way, the market expectation index was calculated, combined with technical analysis, such as deviation rate and Bollinger Bands to build the signal. Through machine learning, support vector regression (SVR) was used to optimize the calculation of expected returns and improved the performance of bull and bear models. We examined the performance of thirteen industry's returns and win rates by market’s signal.
The list of thirteen industries is as follows: Plastics (M1300), Textiles (M1400), Chemical (M1721), Biotech. Med. (M1722), Semiconductor (M2324), Computer and Per. (M2325), Optoelectronic (M2326), Comm. Internet (M2327), Elec. Parts (M2328), Elec. Products (M2329), Other Electronic (M2331), Finance (M2800), Oil, Gas, and Elec.(M9700).
In the bull model, Taiwan’s technical indicators and US economic data were used as predictive variables. The win rate of model became 100% after SVR optimization, and we used it as bull signal. For the bear models,Taiwan’s technical indicators and the US economic data were used as predictive variables individually. The results of SVR optimization could achieve a win rate of 100% in the former and 71.43% in the latter. We took union of two models to build market bear signal, and then applied bull and bear signal into thirteen industries.
For the performance of the industry bull model, except for Textiles (M1400), other Other Electronic (M2331), Oil, Gas, and Elec.(M9700), Biotech. Med. (M1722), the remaining industries could achieve a win rate of more than 80%. For the perfomance of industry bear model performance, except for Computer and Per. (M2325), Elec. Products (M2329), Semiconductor (M2324), Oil, Gas, and Elec.(M9700), Biotech. Med. (M1722), and Comm. Internet (M2327), the rest of the industries could reach 75% win rate.

並列關鍵字

Principal component analysis ； expected return ； Bollinger bands ； machine learning ； support vector machine regression

參考文獻

Balsara, N. J., Chen, G., Zheng, L. (2007). The Chinese stock market: An examination of the random walk model and technical trading rules. Quarterly Journal of Business and Economics, 43-63.

Bollinger, J. (1992). Using bollinger bands. Stocks Commodities, 10(2), 47-51.

Brock, W., Lakonishok, J., LeBaron, B. (1992). Simple technical trading rules and the stochastic properties of stock returns. The Journal of finance, 47(5), 1731-1764.

Chen, S. S. (2009). Predicting the bear stock market: Macroeconomic variables as leading indicators. Journal of Banking and Finance, 33(2),211-223.

Cortes, C., Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3), 273-297.