Ghosh, Sudipto2026-04-292026-04-292026-04-292026-04-10https://hdl.handle.net/10012/23115This thesis investigates the role of positive ("good") and negative ("bad") realized semivariances in modeling and forecasting financial market volatility. The central contribution is the development and empirical evaluation of threshold-based realized volatility frameworks that allow for regime dependence, asymmetry, and time variation in the influence of volatility components. By combining high-frequency information with flexible nonlinear volatility dynamics, the thesis provides new insights into how economically meaningful downside and upside risks affect future volatility. The first part of the thesis develops a threshold realized GARCH framework that explicitly distinguishes between positive and negative realized semivariances. Closed-form expressions for the cross-moment conditions are derived, yielding a computationally feasible setting for regime-specific volatility dynamics. Monte Carlo simulations demonstrate favorable finite-sample properties of the proposed model. Empirical applications to 26 Dow Jones Industrial Average constituents and the S&P 500 index from 1997–2013, as well as an extended sample from 2014–2024, show that future volatility is driven predominantly by negative realized semivariance. This dominance is especially pronounced when regimes are selected from the left tail of the return distribution. The proposed model delivers superior out-of-sample volatility forecasts relative to standard realized GARCH specifications. The second part extends the framework to allow for smooth, time-varying regime shifts, addressing parameter instability in long financial time series. Model parameters evolve gradually over time, capturing extended periods of structural change. Simulation results confirm the reliability of the estimation approach in finite samples. Empirical findings reveal pronounced regime-dependent asymmetries between positive and negative volatility components. While negative realized semivariance dominates during the Global Financial Crisis, the COVID-19 period exhibits a relatively stronger contribution from positive realized semivariance, underscoring important differences in volatility dynamics across crisis episodes. The final part of the thesis examines volatility forecasting under alternative, threshold-based definitions of realized semivariance that extend beyond the conventional zero threshold decomposition. Using high-frequency data for large-cap U.S. equities, forecasting performance is evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Likelihood Ratio (LR) tests. The results show that forecast accuracy remains robust even as increasing thresholds exclude a substantial fraction of low-magnitude intraday returns, indicating that predictive information is concentrated in relatively few large price movements. Further extensions incorporating intraday segmentation and multiscale averaging reveal that negative semivariance during the market closing period contains the strongest predictive content. Overall, the thesis demonstrates that modeling volatility as distinct positive and negative components within threshold-based, regime-dependent frameworks yields substantial gains in interpretability and forecasting performance. The findings highlight the central role of downside and upside risk in volatility dynamics while showing that the informational content of realized semivariance is robust to alternative threshold definitions and market conditions.enrealized semivariancerealized volatilityvolatilityforecastingHARGARCHTRESVAR-GARCHDoctoral Thesis