Forecasting German energy market index within an SVM model & Hilbert Huang transformation

The German wholesale electricity market is the most important in Europe with an annual trade volume of 1,1TWh for 2012. As participants focus on making short-term contracts, mainly via the day-ahead market, the construction of a forecasting model for the directional change of the following day’s electricity prices is important to market participants for short-term decision making.

Deriving data from EPEXSpot, we developed a model to forecast the rise or fall of the Phelix day peak index. The calculation of this index is based on the average electricity prices during the peak-load period of the electricity transmission system. The methodology employed for this purpose in the Support Vector Machines (SVM). SVMs are supervised machine learning methods used for binary classification. Given a finite set of training examples, an SVM estimates an optimal classification hyperplane. The use of kernel based techniques extends the abilities of SVMs to non-linearly separable data mapping the data of the input space to a higher dimensional space. The generalization ability of the model is validated based on the degree on which the SVM model accurately classifies the out-of-sample data.

Non-stationarity, non-linearity and multiple seasonality (hourly, daily, weekly and monthly), are some of the characteristics of electricity prices that render forecasting a demanding task. The Hilbert Huang Transformation (HHT) is an empirical method used to analyze signals that have a non-linear and non-stationary behavior. By combining Hilbert Spectral Analysis with the Empirical Mode Decomposition algorithm, the initial time-series is decomposed in intrinsic mode functions (IMFs), which represent various oscillation modes of the initial data. By applying the HHT method on the Phelix peak-day index, we extracted the various oscillation modes coexisting in the time series of electricity’s daily prices. The IMFs lags were used as input variables in the SVM model and we managed to achieve a 84% forecast accuracy in out-of-sample data during a 200 day period. Moreover, we investigate to which extent the introduction of explanatory variables can improve the accuracy of the model’s forecasts.

JEL code: C53, C49, G17