Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Economics and Finance

Department name when degree awarded


Committee Chair(s)

Christopher Fawson (Co-Chair), Basudeb Biswas (Co-Chair)


Christopher Fawson


Basudeb Biswas


Kenneth Lyon


Seung-Woog (Austin) Kwag


Dwight Israelsen


Yangquan Chen


Volatility is inherently unobservable, and thus the selection of models and their definition is crucial in financial research. This dissertation attempts to check the role of investor sentiment and forecast Value-at-Risk (VaR) of the stock market using both parametric and nonparametric approaches. In the first essay, based on daily return data of three stock indices and four individual stocks from January 1988 to December 2006, the role of day-of-the-week, as well as investor sentiment, is examined using two approaches: linear regression to test investor sentiment effect on stock returns and Logit regression to test the investor sentiment effect on market direction. The results indicate that there is a significant positive role of investor sentiment in the market. However, the outcome also shows that the role of the day-of-the-week effect varies among stocks.

Based on the results presented in the first essay, in the second paper investor sentiment effect was included in both mean and conditional variance equations of GARCH models. By comparing augmented GARCH models considering investor sentiment effect with traditional GARCH models, the result demonstrated that aug-mented GARCH models are signifiantly better than traditional GARCH models where AIC, BIC, log-likelihood, and out-of-sample VaR forecasting were employed. The research indicates that a significant role of investor sentiment in forecasting conditional mean and conditional volatility and the accuracy of GARCH models is improved by accounting for investor sentiment effect.

Compared with the first and second essays employing a parametric method to analyze the stock market, the third paper adopts a nonparametric approach to estimate the conditional probability distribution of asset returns. It is evident that the exact conditional mean and conditional variance is inherently unobservable for time series. In practice, conditional variance is often achieved from different parametric models, such as GARCH, EGARCH, IGARCH, etc., by assuming different distributions such as normal, student's t, or skewed t. Therefore, the accuracy of forecasting strongly depends on the distribution assumption. The nonparametric method avoids the need for a distribution assumption by using a neural network to estimate the potentially nonlinear relationship between VaR and returns. Our results show that the neural network approach outperforms traditional GARCH models.