Forecasting Price Distributions in the German Electricity Market

Abstract

Electricity price distributional forecasts are crucial to energy risk management. In this paper we model and forecast Value at Risk (VaR) for the German EPEX spot price using variable selection with quantile regression, exponential weighted quantile regression, exponential weighted double kernel quantile regression, GARCH models with skewed t error distributions, and various CAViaR models. Our findings are; (1) exponential weighted quantile regression tends to perform best overall quantiles and hours., and (2) different variables are selected for different quantiles and different hours. This is not surprising since the there is a non-linear relationship between fundamentals and the electricity price. This non-linear relationship is different between the different hours as the dynamics of the intra-daily prices are different. Quantile regression has the feature of capturing these effects. As the input mix has changed in Germany over the last years, exponential weighted quantile regression allowing for time-varying parameters can also capture the effect of changing quantile sensitivities over time. Exponential weighted quantile regression is also easy model to implement relative to the other models investigated in this study. Thus, we recommend this model together with carefully selecting fundamentals for given hours and quantiles when the aim is to forcast VaR for German electricity prices.