The objective of this study is to investigate the quality of clouds simulated by the National Centers for Environmental Prediction global forecast system (GFS) model and to examine the causes for some systematic errors seen in the simulations through use of satellite and ground-based measurements. In general, clouds simulated by the GFS model had similar spatial patterns and seasonal trends as those retrieved from passive and active satellite sensors, but large systematic biases exist for certain cloud regimes especially underestimation of low-level marine stratocumulus clouds in the eastern Pacific and Atlantic oceans. This led to the overestimation (underestimation) of outgoing longwave (shortwave) fluxes at the top-of-atmosphere. While temperature profiles from the GFS model were comparable to those obtained from different observational sources, the GFS model overestimated the relative humidity field in the upper and lower troposphere. The cloud condensed water mixing ratio, which is a key input variable in the current GFS cloud scheme, was largely underestimated due presumably to excessive removal of cloud condensate water through strong turbulent diffusion and/or an improper boundary layer scheme. To circumvent the problem associated with modeled cloud mixing ratios, we tested an alternative cloud parameterization scheme that requires inputs of atmospheric dynamic and thermodynamic variables. Much closer agreements were reached in cloud amounts, especially for marine stratocumulus clouds. We also evaluate the impact of cloud overlap on cloud fraction by applying a linear combination of maximum and random overlap assumptions with a de-correlation length determined from satellite products. Significantly better improvements were found for high-level clouds than for low-level clouds, due to differences in the dominant cloud geometry between these two distinct cloud types.