This study investigates the potential of predicting local precipitation over northern Taiwan using statistical downscaling of large-scale circulation variables from global climate models (GCMs). Historical hindcast data of 500 hPa geopotential height (Z500) and sea level pressure (SLP) from six different GCMs, with the target season of being that of June, July, and August (JJA), are used as predictors for downscaling. Singular value decomposition analysis (SVDA) using observational data reveals that the rainfall over northern Taiwan is strongly coupled with a prominent tripole pattern of Z500 (SLP) field over the western North Pacific/East Asian coast. SVDA using model SLP or height field and station rainfall as input also gives similar results, indicating that most models can capture this mode of covariability. SLP and Z500 from models are then used for local rainfall prediction based on their relationship, which is drawn from the SVDA. For every station considered in this study, downscaled prediction shows considerable improvement when compared with model output. In particular, downscaling is able to correct the erroneous sign of model rainfall prediction. However, a few models show very low skill in their downscaled precipitation. For these models, the correlation between observed rainfall and simulated Z500 (SLP) leading SVD patterns is found to be weak. The performance based on the average of downscaled prediction using Z500 and SLP is also evaluated. In general, the average prediction is more stable and skillful when compared with results based on one predictor. Overall, this study demonstrates that useful regional climate information can be obtained from downscaling using large-scale variables from coarse-resolution GCM products.