基於Google Trends的投資人情緒萃取,本文提供一具情緒表徵學習能力的集成預測框架。以隨機森林模型建構台灣加權指數期貨量化交易策略為例,本文探究輔以情緒萃取的分類器特徵生成之於模型預測能力及其量化交易策略之影響。本文的研究發現,輔以市場負面情緒(FEARS指數)以及股市關注度(Company_SVI)特徵生成,能有效提高隨機森林模型之陰性預測能力,其量化交易策略於測試區間之累積損益與風險比率皆勝出於大盤。特別地,我們發現2020年新冠疫情之後,輔以情緒特徵生成之模型預測能力及交易策略績效都能夠有效提升,在獲得與大盤相同績效的同時,承受虧損的幅度以及時間皆呈現大幅縮減。另外,當允許市場情緒萃取作近一步正負面之區分,本文發現陰性預測率雖能更有效提升,然而對下方風險的趨避能力下降,從而減損其量化交易策略之績效。
By extracting public investor sentiment from Google Trends, this thesis provides an ensemble prediction framework that allows for sentiment representation-learning. Based on random forest models, TAIEX futures trading strategies are devised to examine the impacts of the added sentiment dimension on the random forest models’ predictive abilities and the trading strategies’ risk-reward performances. Our numerical findings show that, sentiment assisted representation-learning, when attributed by FEARS and Company_SVI indices, can effectively improve the downside predictive ability of random forest models, resulting in higher cumulative returns and better risk-return profiles relative to simple buy-and-holds. Further evidence suggests that, adopting sentiment assisted representation learning, especially during the post-pandemic era (after 2020), helps to maintain a comparable risk-return profile relative to that of a buy-and-hold while at the same time significantly reduces the extent of losses and the time endured for losses. In addition, upon further categorizing market sentiment as positive or negative, the random forest models’ downside predictive power is found to increase while the strategies’ downside-risk-aversive ability seems to decrease, leading to an overall detrimental effect on trading performance.