The operational ocean prediction model for the North and Baltic Seas of the German Maritime and Hydrographic Agency (BSH) is augmented with a multivariate data assimilation (DA) system. We report on the implementation and performance of the scheme which is based on ensemble forecasting. Here we apply the localised Singular Evolutive Interpolated Kalman (SEIK) filter for assimilating the NOAA AVHRR-derived sea surface temperature (SST) data. Results are presented for two periods: October 2007 is used for calibration and March 2011 for the analysis of the performance in a pre-operational phase. The major forecast improvement is found to be a reduction in the local temperature bias. As compared with the regular BSH forecast without assimilation, the root mean square difference between the predicted SST and satellite observations is reduced on average from 0.87. °C to 0.53. °C for March 2011. The quality of the predicted fields that were not assimilated (velocities, sea level and salinity) is preserved as is confirmed by independent in situ data. The results have required adjustment of the conditional data error statistics. The experiments conducted with different timing and frequency of data assimilation and variable forecasting periods show that the DA system corrects systematic model uncertainties and, due to memory to the corrections, improves prediction over periods of up to 5. days. The results also explicitly illustrate a lower quality of the AVHRR daytime product and reveal low informative influence of the data on the forecasting system when daytime SSTs are assimilated additionally to midnight observations. © 2012 Elsevier B.V.