We apply the CARDS solar forecasting tool, developed at the University of South Australia, to forecasting of solar radiation series at three sites in Guadeloupe in the Caribbean. After performing the model estimates at each individual site, forecast errors were tested for cross correlation. It was found that on an hourly time scale, there was small but significant correlation between sites, and this was taken into account in refining the forecast. Cross correlation was found to be insignificant at the ten minute time scale so this effect was not included in the forecasting. Also, the final error series in each case was tested for an ARCH effect, finding that to construct prediction intervals for the forecast a conditional heteroscedastic model had to be constructed for the variance. Note that cross correlation between sites has to be included for this procedure as well as in the forecasting of the radiation. (C) 2014 Elsevier Ltd. All rights reserved.