Time series structure of news sentiment

Although news sentiment estimation is a hot topic in financial studies, no previous research investigated the time series structure of news sentiment and how to predict it. In this paper, we construct the pessimism index from 2 columns of the New York Times over a 100-year horizon and study the time series structure of the pessimism index. We find that structural breaks in mean and variance happened over the time horizon, and there is a “day-of-the-week” effect on the pessimism index. We find a total of 10 breaks in mean and 18 breaks in variance in history, and the dates of breaks are related to major events in the financial market. The upward breaks in the mean of the pessimism index usually indicates downside risk in the market, whereas the downward breaks usually indicates stock market recovery from hard times. The breaks in variance usually related to uncertainty change in the market. We also find that the effect of the pessimism index on the financial market is over-estimated if structural breaks are not considered. After normalizing the pessimism index with local means and variances, we find the autoregressive moving average (ARMA) model fits the pessimism index. We tried ARMA model, artificial neural network model, and a hybrid model to predict the sentiment, and the results indicate that hybrid model performs significantly better than ARMA model, indicating the existence of the nonlinear phenomena in the pessimism index prediction.雖然新聞情緒的估計是金融研究中的一個熱門話題，但是之前從未有過關於時間序列結構和預測的研究。在本文中，我們構造了《紐約時報》中的兩個欄目100 年來的悲觀指數，來研究該指數的時間序列結構。我們發現該時間序列中均值和方差的結構性斷裂的確發生，而且該指數存在“周內效應”。我們一共發現了10 個均值的結構性斷裂和18 個方差的結構性斷裂，而且結構性斷裂的日期與金融市場相關的大事件有關。向上的均值結構性斷裂通常可以預示金融市場中的下行風險，而向上的均值結構性斷裂通常可以表明股票市場從困難時期的恢復。方差的結構性斷裂則通常和市場中不確定性的變化相關。我們同時發現，如果不考慮悲觀指數中的結構性斷裂，那麼該指數對於市場回報率的影響會被高估。在我們通過局部均值和標準差來歸一化悲觀指數後，我們發現自回歸滑動平均模型適用於該指數。同時我們使用了回歸滑動平均模型，人工神經網路模型和一種混合模型來預測該悲觀指數，結果顯示雖然三種模型都可以預測，但這種混合了回歸滑動平均模型和人工神經網路模型的方法最好。該結果揭示了悲觀指數預測中存在的非線性關係。Yao, Ming.M.Phil. Chinese University of Hong Kong 2020.Includes bibliographical references (leaf 56).Abstracts also in Chinese.Title from PDF title page (viewed on October 11, 2022)