Regular supplies of crude oils during a 3 year period were used to develop and test a partial least-squares multivariate model for determining the total acid number (TAN) based on near-infrared spectra. Several models built automatically by chemometrics software were selected for testing in a series of cross-validations and via large, independent prediction data sets. Cross-validation errors were largely consistent regardless of whether 2% or 20% of the data was left out of the calibration. Two of the three separate prediction sets were also predicted very satisfactorily, but one of the prediction sets, covering a full year of spectra and being the same size as the calibration model, showed some outliers and an obvious deterioration in prediction errors. Nevertheless, two of the selected models held satisfactorily after a small number of outliers were removed and thus proved very effective for determining the TAN in crude oil for all the spectra involved. The spectra from the poorest-performing prediction set were compared with the spectra from the calibration set and no meaningful spectral differences were identified.