Statistical post-processing of sea surface temperature forecasts

Skillful month-ahead predictions of sea surface temperature and modeling its impact on our weather is widely considered a promising gateway to obtaining seasonal weather forecasts.
In this talk we discuss statistical post-processing of sea surface temperature forecasts on the entire globe issued by the numerical weather prediction model NorCPM.
Challenges are, among others, strong seasonality effects, trends in the bias caused by global warming, a non-stationary spatial error correlation, and a short period of available training data.
We compare various alternatives for bias correction and variance estimation and demonstrate that, despite the high dimensional nature of the problem, principal component analysis can be used to approximate the full error covariance matrix.