With very noisy data, having plentiful samples eliminates overﬁtting in nonlinear regression, but not in nonlinear principal component analysis (NLPCA). To overcome this problem in NLPCA, a new information criterion (IC) is proposed for selecting the best model among multiple models with diﬀerent complexity and regularization (i.e. weight penalty). This IC gauges the inconsistency I between the nonlinear principal components (u and ˜u) for every data point x and its nearest neighbour ˜x, with I = 1 − correlation(u, ˜u), where I tends to increase with overﬁtted solutions. Tests were performed using autoassociative neural networks for NLPCA on synthetic and real climate data (tropical Paciﬁc sea surface temperatures and equatorial stratospheric winds), with the IC performing well in model selection and in deciding between an open curve or a closed curve solution.