This dissertation addresses the critical issue of volatility forecasting in financial markets. Volatility is a key indicator that reflects the magnitude of fluctuations in asset prices, and accurate forecasting enables investors to respond more effect...
This dissertation addresses the critical issue of volatility forecasting in financial markets. Volatility is a key indicator that reflects the magnitude of fluctuations in asset prices, and accurate forecasting enables investors to respond more effectively to various market risks. In particular, volatility plays a central role in option pricing as a major input for theoretical valuation, and is also essential for portfolio construction strategies aimed at achieving stable returns by considering volatility. Although various methods for measuring volatility exist, most previous studies have focused on realized volatility, which performs strongly according to ex-post evaluation criteria. To predict realized volatility, many approaches based on linear regression models that capture the long-memory property of past volatility have been proposed, and more recently, a variety of models utilizing the predictive power of deep learning have emerged. However, these existing studies have a limitation in that they fail to sufficiently incorporate intraday return information, which constitutes realized volatility. This is largely due to the challenges posed by microstructure noise, fat-tailed distributions, and the large volume of data inherent in intraday returns. To overcome these limitations, this dissertation proposes a novel framework that employs probabilistic modeling to predict the quantiles of next-day intraday returns and, based on these, forecasts realized volatility. For this purpose, a model capable of flexibly learning data-driven distributions is designed, specifically a Neural Quantile Function with Recurrent Neural Network (NQF-RNN) that models quantile functions from time series input. This model demonstrates competitive performance compared to existing probabilistic models in time series with periodic and irregular patterns, and its flexible structure allows for diverse applications. In addition, a Deep Learning Quantile Function (DLQF) model, which has a structure similar to the NQF-RNN, is applied to effectively learn the distribution of intraday returns for realized volatility. The DLQF consists of a BiLSTM model that receives historical time series data and a DNN that outputs quantile values based on compressed time series information and probability inputs. Given the difficulty of obtaining the true distribution of intraday returns, the model is trained using an L2 distance-based loss function between the empirical distribution function and the predicted quantile function. Utilizing the trained model, experiments are conducted on three major U.S. market ETFs (SPY - S&P 500, DIA - Dow Jones Industrial Average, QQQ - NASDAQ 100) using a rolling-window approach, where five years of data are used for training, one quarter for validation, and one quarter for testing, covering Q3 and Q4 of 2022 and Q1 and Q2 of 2023. The proposed framework demonstrates superior predictive performance compared to existing benchmarks in terms of MAE, MSPE, and MAPE. However, in terms of MSE, the model shows less satisfactory performance compared to others. To address this, a correction model is additionally introduced to adjust the predicted values derived from the probabilistic distribution, resulting in improved performance on key risk management indicators such as MSE and QLIKE, which are crucial for measuring the risk of volatility underestimation. The proposed model is trained using both a distribution learning loss function and a distance-based loss function between the final prediction and actual realized volatility, allowing for flexible design according to the model’s objective. Furthermore, economic analysis based on a mean-variance utility function confirms the economic significance and practical utility of the proposed model. This model can be effectively utilized for volatility management and the development of investment strategies aimed at minimizing risk, thereby contributing to safer investment decision-making. In conclusion, this dissertation proposes a novel realized volatility forecasting model that actively incorporates intraday return information, thereby laying the foundation for a new analytical methodology that can advance to both volatility forecasting research and practical investment strategy development.