Research On The HAR Volatility Model In Electricity Markets

The heterogeneous autoregressive HAR model has been widely applied and developed in the modeling and prediction of high-frequency volatility of financial assets.The electricity price volatility has the characteristics of high volatility,extreme jumps,and reverse leverage effect.Existing electricity price HAR model research mostly focuses on the characterization of jump and reverse leverage effect.It is often estimated using raw realized volatility(RV)and the method of ordinary least squares(OLS),and the choice of the estimator and estimation method is often customary.Even if there are improvements in estimators and estimation methods,there is a lack of systematic comparison between related improvements.However,given stylized fact of electricity price RV(such as non-Gaussianity,conditional heteroskedasticity,spikes/outliers)and well-known propperties of OLS(highly sensitive to outliers,suboptimal in the presence of conditional heteroskedasticity),this combination should be far from ideal.Research on the selection of estimators and estimation methods is of great significance.One goal of this paper is to investigate how the predictive accuracy of the HAR model depends on the choice of estimator and estimation method.In terms of estimators,this article compared the effect of RV and its transformations(square root transformation,Box-Cox transformation)on prediction.In terms of estimation methods,this article compared the impact of OLS regression and its alternative regression(neweywest adjusted OLS,weighted least squares WLS,least absolute deviation LAD,and maximum likelihood MLE method with a construction GARCH,EGARCH and FIGARCH model for residuals)on prediction.In addition,to address the problem of sudden changes in the model structure caused by economic policy,emergencies,etc.,this paper first proposes a dynamic model averaging(DMA)approach in the electricity market.The advantage of DMA is that it takes account of the fact that forecasting models and their coefficients can change over time.So the impact of different models in different periods on the prediction is fully considered.The empirical results show that:1)During the stationary period of volatility,square root transformation and Box-Cox transformation of raw RV have an improvement in predictive accuracy over benchmark models.Among several transformations,the square root transformation and Box-Cox transformation(l =0)model have better prediction performance and prediction robustness.2)During the stationary period of volatility,it was found that the LAD method and MLE method with a construction GARCH and FIGARCH model for residuals show better prediction robustness than the benchmark models.3)Neither the transformation of the estimator or the improvement of the estimation method can achieve good prediction performance in the oscillatory period of volatility.The author believed that the assumption of the heterogeneous market hypothesis during the oscillation period is no longer sufficient,and the model is very sensitive to sample data.At this period,the volatility changed drastically.Even if the model is improved to some extent,it is difficult to capture the trend of the real volatility.4)Adopting DMA method has significantly improved the prediction performance of the benchmark models in the stationary period of RV,and at the same time,the DMA model has a very good prediction robustness across the stationary and oscillation periods.