Pub. Date | : January, 2021 |
---|---|
Product Name | : The IUP Journal of Applied Economics |
Product Type | : Article |
Product Code | : IJAE10221 |
Author Name | : Rama Krishna Yelamanchili |
Availability | : YES |
Subject/Domain | : Economics |
Download Format | : PDF Format |
No. of Pages | : 18 |
This paper analyzes the importance of data frequency, time horizon, and assumption of distribution density in modeling symmetric and asymmetric Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models. The paper uses high frequency (daily) and low frequency (monthly) return series of two Indian stock market indices (Sensex and Nifty) spread over long horizon. It analyzes 7,020 daily observations and 347 monthly observations of Sensex spread over 29 years, and 5,971 daily observations and 287 monthly observations of Nifty spread over 24 years. It estimates several symmetric and asymmetric GARCH models on each return series with normal, student's t, and Generalized Error Distribution (GED). The paper further checks model adequacy with Portmanteau test and model efficiency with information criterion. The results are mixed. First, student's t-distribution is appropriate distribution density for daily return series, whereas normal distribution is appropriate distribution density for monthly return series. Next, leverage effect is visible in daily return series, but not in monthly return series of both indexes. For daily return series, ARMA-GARCH models are adequate models, whereas for monthly return series, basic GARCH specifications suffice. ARMA-APARCH(1,1) model is the best fit model for Sensex daily return series and ARMA-EGARCH(1,1) model is the best fit model for Nifty. These two models capture all the stylized facts of return series. For monthly return series, GARCH(1,1) model is sufficient to capture stylized facts. Depending on the results, it is concluded that data frequency, time horizon, and assumptions of distribution density determine GARCH model specifications.
The Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models are ubiquitous to estimate conditional volatility of economic and financial time series data. The importance of GARCH models continues with extensions to existing models. Currently, many models are available to researchers posing challenge to choose the best fit model. There are multiple factors that determine model selection. Among the factors, frequency of data, time horizon, and assumption of distribution density are primary. Recent studies analyze