Accurate prediction models for fat and total protein content in raw milk have been built on visible and short-wave near infrared spectra (400-1100. nm) of a designed sample set with systematically varied nutrient composition and homogenization degree. Unlike the conventional approach to the spectroscopic milk analysis, exploiting the components' absorbance, the present method basically relies on the phenomenon of light scatter by fat globules and protein micelles.It has been shown that partial least-squares (PLS) regression on raw spectral data results in higher prediction accuracies of fat and total protein content compared to scatter-corrected spectra. Interpretation given to individual PLS factors confirms a dominating role of scatter in the modeling and aids in general understanding of quantitative multivariate analysis based on complex optical responses resulting from multiple scattering by different particles at significantly varying size distributions. This new approach can be used in the raw milk laboratory or field analysis as well as for in-line monitoring of various milk transfer and production stages.Practical modeling issues, i.e. experimental design of the sample set, model validation and refinement, have been also elaborated in this study. Variable selection by interval PLS (iPLS) regression essentially improved the model performance. Selected intervals can be utilized for technical simplification of suggested method of milk analysis.