When Aksel Wiin-Nielsen became director of the European Centre for Medium-range Weather Forecasting (ECMWF), national forecast centres did not make useful operational forecasts past a few days. Aksel said, no doubt from his experience in the states at the Joint Numerical Weather Prediction Unit of the U.S. Weather Bureau, that medium range (10 day) forecasts could be made successfully with state of the art computing and other technology, with the best and the latest in physics and numeric techniques including the best observation assimilation, and with the best scientists from the member states. Aksel's idea of developing a numerical weather prediction model from “square 1” is certainly an advantage, and having the European atmospheric community's research converted to operations has given ECMWF the lead in NWP modelling skill since its inception. We note, comparing the NCEP and ECMWF forecast systems, that it is no secret that skill for synoptic and planetary model forecast measures are better for ECMWF forecasts than with the NCEP GFS. In addition, on approximately a monthly basis, poor forecasts or skill score “Dropouts” plague GFS performance and are responsible for a ~10% skill loss. The ECMWF forecasts often do not exhibit this loss of skill even though the raw observation feeds are similar. The climatology of these events indicates differences between mass and motion fields that are due to quality control (QC) and treatment of conventional and non-conventional observations in the assimilation system. This report investigates how the NCEP global forecast system (GFS) model skill compares with ECMWF by examining the influence of observations and assimilation. We attempt to quantify the differences between the GFS and ECMWF treatment of observations to detect issues that when addressed will improve GFS forecasts. A GFS analysis is made using ECMWF initial conditions converted to simulated or “pseudo” RAOB observations, and inserted into the NCEP GFS grid point statistical interpolation (GSI) to create a new analysis from which new forecasts are made. These are called ECM experiments and are compared with operational forecasts. This same process can be repeated for the output of GFS operations, instead of the ECMWF output, to make controlled runs to compare with the ECM experiments. The ECM results show improvement in 5-day skill scores in typical and practically all Southern and most Northern hemispheric dropout (worst) cases. Areas in the initial condition that influence the sensitivity of forecast skill can be confirmed by creating a hybrid initial condition, selectively overlaying the ECMWF pseudo-analysis over the GFS initial condition described above. The regions used are “patches” over special areas, e.g., where there is ambiguity in observation quality or large differences between centre initial conditions, to isolate problems that alter downstream 5-day forecasts to determine if problems originate from particular observation types.