A novel approach to quantify volatility prediction

TL;DR

This paper introduces a robust least squares approach with and without LAR for predicting VIX volatility, decomposing it into trend, seasonal, and residual components, and provides generalized prediction equations for practical use.

Contribution

The paper presents a novel, simple robust least squares method with LAR for volatility prediction and offers the first generalized prediction equations for volatility components.

Findings

Robust least squares with LAR outperforms without LAR in RMSE.

Decomposition improves understanding of volatility components.

Generalized equations enable direct application to similar datasets.

Abstract

Volatility prediction in the financial market helps to understand the profit and involved risks in investment. However, due to irregularities, high fluctuations, and noise in the time series, predicting volatility poses a challenging task. In the recent Covid-19 pandemic situation, volatility prediction using complex intelligence techniques has attracted enormous attention from researchers worldwide. In this paper, a novel and simple approach based on the robust least squares method in two approaches a) with least absolute residuals (LAR) and b) without LAR, have been applied to the Chicago Board Options Exchange (CBOE) Volatility Index (VIX) for a period of ten years. For a deeper analysis, the volatility time series has been decomposed into long-term trends, and seasonal, and random fluctuations. The data sets have been divided into parts viz. training data set and testing data set. The validation results have been achieved using root mean square error (RMSE) values. It has been found that robust least squares method with LAR approach gives better results for volatility (RMSE = 0.01366) and its components viz. long term trend (RMSE = 0.10087), seasonal (RMSE = 0.010343) and remainder fluctuations (RMSE = 0.014783), respectively. For the first time, generalized prediction equations for volatility and its three components have been presented. Young researchers working in this domain can directly use the presented prediction equations to understand their data sets.