The feasibility of utilizing infrared spectroscopy for the prediction of haze formation in white wines resulting from heat and colloidal stability tests was investigated. One-hundred eleven white wines, representing multiple regions and varieties from the 2008 California vintage, were collected and analyzed. The near and mid-infrared spectra were measured and heat and colloidal (ethanol addition) stability tests were performed on the same wines. Partial-least squares regression analysis was then used to construct models predictive of the resulting nepholometric turbidity to the acquired spectra. Preliminary models obtained following application of spectral pretreatments today considered as "classical" (e.g., derivatives, standard normal variate, vector normalization, constant offset elimination) lacked robustness; two alternative algorithms designed to remove spectral information unrelated to the turbidity were then employed (orthogonal signal correction; direct orthogonal signal correction). While OSC pretreatment did not result in more robust models, DOSC considerably enhanced the goodness of the PLS model constructed to predict the ethanol test turbidity. Predictive modeling of the short-NIR spectra, following DOSC preprocessing, allowed the prediction of colloidal stability on an unknown test set with an R(2) = 0.80 and a RMSEP = 10.12 using three latent variables. When the data set was restricted to Chardonnay wines alone, the predictive ability improved, with R(2) = 0.85 and RMSEP = 8.90. (C) 2010 Elsevier Ltd. All rights reserved.