Toni,

the answer to your question is very simple if I understand it correctly.  You simply multiply by 100! 

You should generally neglect any negative correlation values, although it might occasionally be the case that a strong negative value indicates that there is potential skill but the forecast model has misplaced the signal geographically.

Alternatively, a common measure is to square the correlation score and again multiply by 100 to give an estimate of the amount of variance explained.  Again beware of negative correlations before you do this of course.

Best regards

Adam

Thank you, Adam! It seems, variance explained and multiplication by 100 are indeed what I was looking for. I suppose it never occurred to me that the solution could be so simple.

Something else that puzzles me is, anomaly correlation is often used with probabilistic forecasts when it is recommended for deterministic forecast by WMO and CPC (see link below). Why is it so common to use it with probabilistic forecasts then? Can they just be treated like deterministic forecasts by using the mean of the spectrum?

http://www.cpc.ncep.noaa.gov/products/ctb/Metrics_climate_models&fcst_15May2015_v2.docx.pdf

Hi Toni!

I think the answer to your second question (usage of anomaly correlation for deterministic/probabilistic forecasts) lies in clarifying the terms:

I suppose you mean ensemble forecasts when writing probabilistic forecasts. But the thing is, just by doing an ensemble prediction you haven't produced a probabilistic forecast, yet.

A probabilistic forecast would be to deduct probabilities for certain events out of your ensemble forecast (the easiest way would be to count the ratio of ensemble members forecasting the event, e.g. temperatures bove normal; e.g. done for calculating the Brier (Skill) Score or the Ranked Probability (Skill) Score) or even consider the whole probability density function (or cumulative distribution function as done when calculating the Continuous Ranked Probabilioty (Skill) Score).

However, many studies employing ensemble predictions consider the ensemble mean only. This makes it a deterministic approach, as eventually there is only ONE forecast considered. For such an approach AC is a useful metric and there is no contradiction between these studies and the WMO/CPC recommendation.

Tim, thanks for clarifying the probabilistic/ensemble difference. This brings up a lot of new questions, like what an actual probabilistic forecast looks like data-wise (if it's not a singular value), but maybe someone already answered them elsewhere. Thanks again!