A model-free approach to do long-term volatility forecasting and its variants.

Area of Science:

• Financial econometrics
• Time series analysis
• Statistical modeling

Background:

• Volatility forecasting is crucial in financial econometrics, primarily relying on Generalized Autoregressive Conditional Heteroskedasticity (GARCH)-type models.
• Traditional GARCH models exhibit limitations, including instability with highly volatile or short datasets and difficulty in uniform cross-dataset application.
• Existing methods struggle with the complexities of financial market volatility, necessitating more robust prediction techniques.

Purpose of the Study:

• To evaluate the effectiveness of the Normalizing and Variance Stabilizing (NoVaS) method for long-term volatility forecasting compared to standard GARCH models.
• To investigate the performance of NoVaS, particularly on short and highly volatile financial datasets.
• To introduce and assess a novel variant of the NoVaS method designed for improved forecasting accuracy.

Main Methods:

• Extensive empirical and simulation analyses were conducted to compare forecasting performance.
• The study utilized the model-free NoVaS method, originally derived from an ARCH model inverse transformation.
• A new, more complete variant of the NoVaS method was developed and tested.

Main Results:

• The NoVaS method demonstrated higher-quality long-term volatility forecasting than standard GARCH models.
• This advantage was particularly pronounced for datasets characterized by high volatility and short time series.
• The proposed NoVaS variant generally outperformed the existing state-of-the-art NoVaS method.

Conclusions:

• NoVaS-type methods exhibit uniformly superior performance, supporting their widespread adoption in volatility forecasting.
• The flexibility of the NoVaS approach allows for further development and application in addressing specific prediction challenges.
• The study validates NoVaS as a robust and accurate alternative for financial volatility prediction, especially in difficult data conditions.