A thorough investigation of the performance of broadband direct irradiance predictions using 21 solar radiation models, along with carefully measured radiation data and ancillary meteorological data, is detailed here. A sensitivity study and a detailed error analysis show that precipitable water, and even more so, turbidity, are the two most critical inputs, whose accuracy conditions the resulting uncertainty in irradiance predictions. Large prediction uncertainties result from the use of time/space interpolated or extrapolated data of precipitable water and turbidity. So that the results of performance assessment Studies like this one can be of any significance, it is necessary to rely on highly accurate precipitable water and turbidity data from collocated instruments with an appropriate sampling rate. An experimental assessment of the performance of all models has been conducted, using nearly 5000 data points from five different sites covering a large range of Geographical and climatic conditions. Direct irradiance measured with first-class instruments at these sites are compared to model predictions where precipitable water and turbidity are determined from collocated sunphotometric measurements. This experimental assessment is found to be less stringent than the theoretical assessment (in Part 1 of this investigation), while confirming its main results. The same four high-performance models as in Part 1 are finally recommended: CPCR2, MLWT2. REST and Yang (in alphabetical order). Remarkably, they can predict direct irradiance under a variety of atmospheric conditions within the uncertainty of modern and well-maintained pyrheliometers, provided that good quality inputs of precipitable water and turbidity are used. The MLWT2 model produces the best results, with the lowest bias and variance for any irradiance value. (C) 2003 Elsevier Ltd. All rights reserved.