미국 연방준비제도 텍스트 데이터를 활용한 금리 예측

2022/04/08 ~2022/04/22

최우수팀 선정

• 해당 프로젝트는 '멀티캠퍼스 서비스 산업 데이터를 활용한 빅데이터 분석' 비정형 데이터 활용 분석 프로젝트 결과물입니다.
• 텍스트 데이터 출처: Federal Reserve Board

Abstract

This project was started from the curiousity that if future interest could be predicted using text data. First, we scraped text data such as beigebook and speech from Fed website. Second, we converted text data into numberical data to compare them with interest rates history. We tried several ways to do that. The length of each text data is quite long, so there is difficulty to extract only the necessary parts. Thus, we used question-answering and sentiment analysis from pipelines. Finally, we were able to find that such texts talking about the economy of the time could help expalin the movement of the U.S. interest rates.

Data

• textual data collected from the FED website: minutes, beigebook, speech
• numeric data: prices, unemployment rates
• period: Q3 2010 ~ Q1 2022

Preprocessed: Pipelines

from transformers import pipeline

• question-answering: The lines were extracted from the texts through questions that could confirm the economic situation at the time.
• sentiment analysis: The collected lines were checked where they were positive or negative, and the degrees of sentiment was used for the following analysis.

EDA: Standardization

1. Combine collected data into a single data frame. (Quarterly distribution)

2. Divide the periods and calculate average values for each period.

3. Confirm that the graph for the average values is similar to the graph for interest rates.

Analysis: Regression Model

degree = 1

• Left: minutes, beigebook, speech1
• Center: testimony, minutes, beigebook, speech1, speech2
• Right: minutes, beigebook, speech1, prices, unemployment rate

degree = 2

• Left: minutes, beigebook, speech1
• Center: prices, unemployment rate
• Right: mixed

Conclusion

We verified there is possibility that text data alone can predict interest rates. With unemployment rate and prices, more sophisticated results could be predicted. Therefore, it is expected that it will be possible to predict more diverse economic indicators in the same way. Next, it will be possible to create a website or an application that provides future encomic forecasts. Furthermore, text data can be obtained from various sources to identify the characteristics by each sources. Then, it will be possible to analyze how they interact with certain economic indicators.

Appendix