摘要
本文探討聯邦公開市場委員會（Federal Open Market Committee，簡稱 FOMC）會議紀要（minutes）的文字風格，尋找不同的利率調整結果（升息、降息、利率不變）時的 FOMC會議紀要的用詞有哪些差異。本文研究資料為1993～2020年的FOMC會議紀要，以及同時期的目標聯邦基金利率（targeted federal funds rate）。透過探索式資料分析（exploratory data analysis）、文字分析（或稱文字探勘）的技術，比較FOMC會議紀要在升息、降息、利率不變時的會議紀要風格。分析發現，目標聯邦基金利率的調整並非隨機（亦即具有自相關性），經常出現連續幾期的升息、降息；文字使用，在不同利率調整時有不少差異，其中升息、降息的報告大多強調美國四大族裔的失業率。另外，由於分析會議紀要的原始文檔為高維度的文檔-詞頻矩陣（document-term matrix），考量多達 4102 個變數，除了具有稀疏矩陣（sparse matrix）的特質外，變數過多也會影響資料分析的效率。因此本文使用倍數指標篩選器、線性降維、非線性降維等方法，透過縮減特徵空間維度以提高執行效率，研究發現倍率指標的為度縮減效果最佳，配合羅吉斯迴歸得出之三分類準確率最高。
並列摘要
In this study, our goal is to explore the writing style of FOMC (Federal Open Market Committee) minutes. In particular, we want to know if the style of minutes shows significant differences when the FOMC decided to raise, lower, or hold interest rates. We applied exploratory data analysis and text mining techniques to the FOMC 1993~2020. We found that the adjustments of targeted federal funds rates are not randomly distributed and they show signs of correlation. For example, among the 39 times of raising interest rate, there was one 17 consecutive intertest increase. Also, the minutes tend to emphasis on the unemployment of four major ethnicities when FOMC decided to raise or lower interest rates. On the other hand, there are 4102 variables involved in exploring the writing study of FOMC minutes. This means that the document-term matrix is a sparse matrix and high dimensionality requires a lot of computation time. Thus, adopted dimensionality reduction techniques: multiplication index, linear reduction and non-linear reduction methods. We found that the multiplication index has the best performance and, together with logistic regression, it has the highest accuracy in classifying the writing style of FOMC minutes in the cases of raising, lowering and holding interest rates.