文字探勘對量化交易策略報酬之影響：人民幣兌美金外匯保證金商品

本論文使用2015/09/14~2018/12/31的人民幣兌美元之外匯保證金商品報價，及透過爬蟲爬取相同時間區間中、英文兩種語言之新聞，來源包括鉅亨網、FXWeek、Currency News、Daily Beast、Reuters等新聞網站。藉由Ckiptagger、NLTK等python套件，對文章進行分詞處理，再透過TF-IDF與滾動線性迴歸等量化方式，製作多種語言情緒指標。最後利用情緒指標將消息面的資訊導入交易策略中，並基於新聞來源及策略不同，建立多個交易策略模型，並嘗試使用除了獲利以外的衡量指標，來比較策略模型的績效。實證結果發現，中文情緒指標建立之策略較英文情緒指標建立之策略績效較於優異，而英文情緒指標策略在新聞來源不同時，多個新聞來源較單一新聞來源策略績效優異；若情緒指標運用在出場策略，普遍能降低最大策略虧損，若將情緒指標運用在進場策略，普遍能提高勝率。

關鍵字

爬蟲；文字探勘；量化交易；情緒指標；外匯保證金

並列摘要

This paper uses the foreign exchange margin commodity quotes of USD/CNY from 2015/09/14 to 2018/12/31, and web crawls the news in both English and Chinese. Sources include Juheng.com, FXWeek Currency news, daily beasts, Reuters. Through Ckiptagger, NLTK, TF-IDF, rolling linear regression and other quantitative methods to produce a variety of language sentiment indicators. Finally, the sentiment indicators are used to import the information on the news surface into the trading strategy, and compare the performance of the strategy model.The empirical results found that the strategy established by the Chinese sentiment indicator is superior to the English sentiment indicator, and multiple English news sources perform better than the single; The appearance strategy can generally reduce the maximum strategy loss. If the sentiment index is used in the entry strategy, it can generally improve the winning rate.