Q: advice on how to evaluate how good a Mean Reversion Jump Diffusion (MRJP) model is
Andrew Coyle
It has been proposed that a "Mean Reversion Jump Diffusion (MRJP) model" be
used for forecasting electricity spot prices. This would be provided as a
black box type model, although it may require some tuning of the input
parameters. Basically I am after advice on how to asses whether this model
will produce good forecasts.
Can anyone provide me with
- Suggested resources which explain what exactly MRJP models are, and how
they work (i understand they are separable, so info on MR or JP models alone
would be fine). I am after both a non-technical explanation, and stuff which
goes into more detail about the individual input parameters. I am in no way
a maths person, but will attempt to struggle through if required.
- Known pros and cons of MRJP models. How to detect if any of the pros and
cons will apply to our data.
- Advice on how to tune the particular parameters. For example, what
distributions (if any) do MRJP assume are present in the data, and what
techniques exist to test and callibrate these assumptions. I suppose I am
wondering what traits (again, if any) a MRJP model assumes, and then how do
you go about testing for these traits (this also applies to the previous
question on how to detect if the pros and cons will apply to our data).
- Advice on good standard statistical methods for assessing how successful
the forecasts are. No doubt we will try to tune the model (if in fact it
requires us to do so) and then backcast over history. What general
techniques are out there for measuring how well a forecast performed against
a desired set of results.
Help on any of the above will be greatly appreciated. The last one is
particularly important, as if all else fails at least I can blindly assess
how good the model seems to be doing.
best regards
Andrew