Ulrica Müntzing and Joakim Hansson

 present their master thesis

Predicting the Directions of Forward Price Movements at
the Nordic Power Exchange 

 Abstract


 
This thesis is about predicting price movements of Nord Pool's yearly
electricity forward during their trading period. The main idea is to
use a mean-reverting SDE to model the price movements. Out of this
model an expression for th expected value of the forward price F_t can
be calculated. To be able to use that expression the price,P_T , towards
which  F_t reverts has to be approximated. This is done using an additive
model alternatively to create a response surface between the expected
value of P_T and different external variables, such as for example the
hydrological situation in Norway and Sweden and time to
maturity. During the iteration a mean-reverting coefficient alpha and
a diffusion coefficient sigma  have to be estimated. This is done using
least squares estimation. As a compliment to this method another method
which has a similar SDE but with a know distribution, which  makes it
possible to use maximum likelihood estimation instead of least squares
estimation to approximate alpha and sigma, is also investigated. Using
additive models with many cross-terms requires a lot of data. Because
of this we also tried to predict the price  movements using a
classification tree method. Another reason why we tested this method
was that it does not need any economic theory in form of the SDE. Over
50% of our predictions of the direction of the price movements are
correct for the whole trading period, which we have split in three
periods. This true for all three methods.The model used for the period
closest to maturity shows the best residuals of the three periods, but
it surprisingly predicts the directions slightly worse.The reason for
this is that the first year of the trading period, that shows bad
residuals caused by over-estimation of the forward price, has an
up-going trend which  implies that the over-estimation gives correct
predictions. The model for the middle year has good residuals and the
best result. The SDE model, with or without maximum-likelihood
estimation, gives better results than the classification tree.