Modeling and Forecasting the Distribution of Energy Forward Returns - Evidence from the Nordic Power Exchange

We explore intraday transaction records from NASDAQ OMX Commodities Europe from January 2006 to October 2013. We analyze empirical results for a selection of existing realized measures of volatility and incorporate them in a Realized GARCH framework for the joint modeling of returns and realized measures of volatility. An influential bias in these measures is documented, which motivates the use of a flexible and robust methodology such as the Realized GARCH. Within this framework, forecasting of the full density for long horizons is feasible, which we pursue. We document variability in conditional variances over time, which stresses the importance of careful modeling and forecasting of volatility. We show that improved model fit can be obtained in-sample by utilizing high-frequency data compared to standard models that use only daily observations. Additionally, we show that the intraday sampling frequency and method have significant implications for model fit in-sample. Finally, we consider an extensive out-of-sample exercise to forecast the conditional return distribution. The out-of-sample results for the Realized GARCH forecasts suggest a limited added value from using “traditional” realized volatility measures. For the conditional variance, a small gain is found, but for densities the opposite is the case. We conclude that realized measures of volatility developed in recent years must be used with caution in this market, and importantly that the use of high-frequency financial data in this market leaves much room for future research.