- 學位論文
建立經濟景氣反轉與匯率上行波動的大蕭條股價預測模型
Build a Stock Price Predictive Model for Economy Reverse and Exchange Rate Up Volatility during Economic Depression
摘要
2019年12月嚴重特殊傳染性肺炎(COVID-19)疫情爆發，對於全球經濟造成巨大的衝擊，而在高度全球化的影響下，各國之間不論是貿易或是金融交易逐漸頻繁，股市之間連動性也越來越高，因此當一個國家匯率改變時，與其有交易關係的國家也會受到影響，當國家的貨幣升值時，是有利於進口，不利於出口貨物；而國家的貨幣貶值時，則是不利於進口貨物，但是有利於出口。而近20年來經濟衰退的發生次數越來越多，當經濟衰退發生時，沒有一個國家能獨善其身，對於投資者而言，會造成投資決策的難度提升，為了應對世界經濟開始衰退甚至是發生類似1930年大蕭條的情況，投資者可以找尋能夠與經濟趨勢有關聯性的經濟指標或是數據，來建立當經濟衰退時期發生時的投資決策。 因此本研究欲探討的是經濟指標的變化對於股市的波動起伏是否會產生影響，以及針對經濟指標在經濟衰退時期以及非經濟衰退時期，對於市場的漲跌趨勢是否也具有影響力，最後探討經濟指標對於中長期市場的漲跌變化是否具有影響性。所以本研究將蒐集美國之消費者物價指數、美元指數，以及美國三大市場指數，標準普爾500指數、道瓊工業平均指數、納斯達克綜合指數和臺灣加權股價指數歷年收盤價為數據，將蒐集的數據整理為日、周、月、季等不同時間周期資料，接著以美國國內生產毛額將資料分割為經濟衰退期和非經濟衰退期兩個資料集，利用機器學習演算法中的支援向量機、隨機森林兩種方法建立預測台灣加權股價指數的漲跌預測模型，藉此探討經濟指標在經濟衰退期和非經濟衰退期以及用不同時間周期的數據時對於股市產生的影響。 研究結果顯示在經濟衰退期和非經濟衰退期的預測結果上，同時間周期的資料比較下，經濟衰退期的表現皆優於非經濟衰退期的結果；在機器學習模型上隨機森林以67.809%高於支援向量機的67.517%；在時間周期資料的比較上，準確率由低到高分別為日資料的準確率最低，其次是周資料，接著是月資料，而季資料的準確率在所有資料中準確率最高，透過結果可以發現當時間周期越短則預測效果越差，反之時間周期越長則預測效果會越佳。
並列摘要
In December 2019, the outbreak of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), known as COVID-19, had a significant impact on the global economy. With the highly globalized nature of today's world, countries have seen an increasing frequency of trade and financial transactions. As a result, there has been a growing interconnectedness between stock markets, and when the exchange rate of one country changes, it affects other countries with trading relationships. When a country's currency appreciates, it is advantageous for imports but detrimental to exports. Conversely, when a country's currency depreciates, it is disadvantageous for imports but beneficial for exports. In the past 20 years, economic recessions have become more frequent, and during such periods, no country can remain unaffected. For investors, economic recessions increase the difficulty of making investment decisions. To prepare for a global economic recession similar to the Great Depression of the 1930s, investors can seek economic indicators or data that are correlated with economic trends to inform their investment decisions during recessions. Therefore, this study aims to explore whether changes in economic indicators have an impact on the volatility of the stock market. It also investigates the influence of economic indicators on market trends during economic recession and non-recession periods. Finally, it examines whether economic indicators have an impact on the fluctuations of the medium to long-term market. To achieve these goals, this research will collect historical closing prices of the U.S. Consumer Price Index, U.S. Dollar Index, and three major U.S. market indices (S&P 500, Dow Jones Industrial Average, and NASDAQ Composite Index), as well as the Taiwan Stock Exchange Weighted Index. The collected data will be organized into different time periods such as daily, weekly, monthly, and quarterly data. Subsequently, using the Gross Domestic Product (GDP) of the United States, the data will be divided into two sets: economic recession period and non-recession period. Support Vector Machines and Random Forest, two machine learning algorithms, will be employed to establish a prediction model for the rise and fall of the Taiwan Stock Exchange Weighted Index. This will allow for the examination of the impact of economic indicators during economic recession periods, non-recession periods, and different time cycles on the stock market. The research results indicate that in terms of prediction accuracy during economic recession and non-recession periods, the performance of the recession period is better than that of the non-recession period when comparing data of the same time cycle. Among the machine learning models, Random Forest performs better with an accuracy of 67.809% compared to Support Vector Machines with 67.517%. When comparing different time cycle data, the accuracy ranks from lowest to highest as follows: daily data, followed by weekly data, then monthly data, and the highest accuracy is observed with quarterly data. From the results, it can be observed that shorter time cycles lead to poorer prediction performance, while longer time cycles yield better predictions.