An improved classical least-squares (CLS) method that can correct the detrimental effects of traditional spectroscopic data analysis CLS models has been developed. An additional noise (introduced by unmodeled concentrations) model is constructed, and the original mixed spectrum is purified using least squares to eliminate the additional noise signals. The extracted spectral signals are then used to augment the concentration predictive power. An application study is presented, which involves data recorded from an experiment of multicomponent concentration determination using Raman spectroscopy, and two designed strategies are exploited to test the predictive power of the proposed method. Results indicate that this method is an applicable spectral analysis tool for compound concentration determination against traditional CLS model ineffectiveness rendered by unmodeled concentrations, and the predictive power of the proposed method is comparable to existing methods.