Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84149
Title: | 以褐皮書資料預測FOMC利率決策 Beige Book for Predicting FOMC Interest Decision |
Authors: | 陳威宇 WEI-YU CHEN |
Advisor: | 石百達 PAI-TA SHIH |
Co-Advisor: | 蔡政安 CHEN-AN TSAI |
Keyword: | 褐皮書,利率預測,Bert,文字探勘,深度學習, Beige Book,Interest Predict,Bert,Text Mining,Deep Learning, |
Publication Year : | 2022 |
Degree: | 碩士 |
Abstract: | 利率,代表著貨幣的時間價值,在金融活動中,是不可或缺的重要指標。而隨著當今美國金融業的強勢,又以美元的利率最受投資人關注。本研究主要目的為試圖找出褐皮書的文字資料與FOMC利率決策之相對關係,以1983年到2020年所發佈的褐皮書資料,透過NLP的方式進行分析與預測,其中NLP模型選用Bert作為訓練的模型,試圖以褐皮書中對經濟狀況的文字表述,萃取出其可能對整體利率變化的影響,進而達到預測利率變動決策的成果。
進行褐皮書文字資料分析時,以Bert模型為主要模型,結合TF-IDF、Random Forest、 Self-labeling等處理,分別進行升息模型以及降息模型的訓練。升息模型以使用Self-labeling的模型為最佳,整體準確率達66.7%(Total Accuracy),而對於升息決策的預測亦有66.7%的召回率(Recall rate);而在降息模型中則是以Self-labeling與TF-IDF的結合使用效果最好,有78.3%的整體準確率(Total Accuracy),此外對於降息事件的預測則是40%的召回率(Recall rate)。 研究顯示,在升降息模型的比較中,升息模型的預測普遍表現的比降息模型要好。其主因推測為,升息時往往是市場景氣的升溫,進而較容易在褐皮書中找到經濟成長的脈絡;而相對的,當市場趨向降息時,往往是發生了突發事件,對經濟形成一定的衝擊,進而需要降息刺激,然而此時由於事件的突發,導致難以從褐皮書中找到脈絡,進而使得降息模型的表現普遍較差。而在幾種資料處理方式中,Self-labeling能對預測結果造成明顯的提升,而TF-IDF、Random Forest提供的改善則較為有限。 Interest, in other word is the cost of money. In financial world, interest is one of most important factors that everybody cared. And because of the domination of US dollar in the world, the interest of US dollar is also the leading indicator in financial world. On the other hand, FOMC (Federal Open Market Committee) is the Committee that Fed decide the national monetary policy which includes the change of interest. And the Beige Book is report that Fed gathers anecdotal information on current economic conditions, and those economic conditions may be the trigger of interest change decision. so, we can maybe use the beige book data to predict the FOMC interest decision, then can be well prepared for the market change. When analyzing the text data of the Beige Book, the Bert model is used as the main model, combined with TF-IDF, Random Forest, Self-labeling, etc., to train the interest increasing model and the interest decreasing model respectively. The model with Self-labeling is the best in interest increasing predicting, with an Total Accuracy of 66.7%, and the Recall rate in predicting interest increasing event is also 66.7%; in the interest decreasing model, it is The combination of Self-labeling and TF-IDF is the best, with Total Accuracy 78.3%, and the Recall rate in predicting interest decreasing event is 40%. Research shows that interest increasing models generally perform better than interest decreasing models. The main reason is speculated that when the interest rate is raised, the market prosperity is often warmed up, and it is easier to find the context of economic growth in the Beige Book; on the other hand, when the market tends to decrease interest rates, emergencies often occur, which have a negative impact on the economy. However, due to the suddenness of events at this time, it is difficult to find the context in the Beige Book, which in turn makes the performance of the interest decreasing models generally poor. Besides, among several data processing methods, Self-labeling can significantly improve the prediction results, while the improvement provided by TF-IDF and Random Forest is relatively limited. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84149 |
DOI: | 10.6342/NTU202201236 |
Fulltext Rights: | 同意授權(限校園內公開) |
metadata.dc.date.embargo-lift: | 2027-07-01 |
Appears in Collections: | 統計碩士學位學程 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-110-2.pdf Restricted Access | 1.85 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.