The predictive power of log-likelihood of GARCH volatility

Purpose: This paper aims to investigate whether the best statistical model also corresponds to the best empirical performance in the volatility modeling of financialized commodity markets. Design/methodology/approach: The authors use various p and q values in Value-at-Risk (VaR) GARCH(p, q) estimation and perform backtesting at different confidence levels, different out-of-sample periods and different data frequencies for eight financialized commodities. Findings: They find that the best fitted GARCH(p,q) model tends to generate the best empirical performance for most financialized commodities. Their findings are consistent at different confidence levels and different out-of-sample periods. However, the strong results occur for both daily and weekly returns series. They obtain weak results for the monthly series. Research limitations/implications: Their research method is limited to the GARCH(p,q) model and the eight discussed financialized commodities. Practical implications: They conclude that they should continue to rely on the log-likelihood statistical criteria for choosing a GARCH(p,q) model in financialized commodity markets for daily and weekly forecasting horizons. Social implications: The log-likelihood statistical criterion has strong predictive power in GARCH high-frequency data series (daily and weekly). This finding justifies the importance of using statistical criterion in financial market modeling. Originality/value: First, this paper investigates whether the best statistical model corresponds to the best empirical performance. Second, this paper provides an indirect test for evaluating the accuracy of volatility modeling by using the VaR approach.