A new multivariate non-linear non-parametric regression method, called MARS, is used in a statistical downscaling problem for simulating New Caledonian rainfall anomalies induced by El Niño/Southern Oscillation (ENSO). The use of a non-linear method was motivated by wavelet analyses, pointing to a complex interaction between interannual and intraseasonal rainfall variability. The proposed method is very flexible and computationally efficient. It can be applied to a wide range of problems and its use is not limited to spatial downscaling applications. Furthermore the method provides great potential for comprehensive diagnostics of multivariate problems in climate research. It is shown that this method can take into account the asymmetry between El Niño and La Niña periods and also the type of El Niño.