I am using the 2006 till 2009 data for the New England pool of day-ahead weighted average prices. I am not very fluent with econometrics, but I read in some papers that GARCH has been used successfully to forecast electricity wholesale prices. When I train a GARCH model on one year worth of data, and forecast for the last 3 days of training data, I get a mean absolute error of 3.6642%. The error increases to 27.826% when I use two years worth of data, and decreases to 17.123% when using three years worth of past data. Is this expected?

Also, when I plot the actual and fitted data against time, the GARCH model seems to have done a really bad job, compared to a default ARIMA model. I'm guessing this might be because people are actually using ARIMA models with (added) GARCH errors, so a simple GARCH model-based forecast isn't doing exactly what they have done. Am I right? Why would the ARIMA be a better fit than GARCH?

One author mentioned that they took a log of the prices (in their case it was hourly prices) before fitting a GARCH model. In your opinion, is that an important factor in the kind of errors I am getting?

Thanks and best regards

--
Muhammad Saqib Ilyas
PhD Student, Computer Science and Engineering
Lahore University of Management Sciences