Land-based meteorological measurements at two locations on the Danish coast are used to predict offshore wind speeds. Offshore wind-speed data are used only for developing the statistical prediction algorithms and for verification. As a first step, the two datasets were separated into nine percentile-based bins, with a minimum of 30 data records in each bin. Next, the records were randomly selected with approximately 70% of the data in each bin being used as a training set for development of the prediction algorithms, and the remaining 30% being reserved as a test set for evaluation purposes. The binning procedure ensured that both training and test sets fairly represented the overall data distribution. To base the conclusions on firmer ground, five permutations of these training and test sets were created. Thus, all calculations were based on five cases, each one representing a different random selection from the same data, but maintaining the (approximate) 70-30 split in each bin. This procedure served to ensure that conclusions were not based on a single randomly-selected case. Two statistical methods are employed: multiple linear regression (MLR), and Classification and Regression Trees (CART). MLR produces excellent results using only land-based predictors. The CART results are similar to those from MLR, and tend to be slightly better.