In this paper，an iterative shrinkage window-bootstrapping soft shrinkage（ISWBOSS） algorithm was proposed to address problems of poor prediction performance of the calibration models based on near-infrared（NIR） spectroscopy due to their redundant variables in NIR spectra．The variables was divided by windows in the method，and the windows were randomly selected，in which the sub-models were built with the variables．The soft shrinkage of the variable space was gradually achieved by calculating the normalization of the regression coefficients of the variables in the window，and continuing the weighted sampling as weights．Meanwhile，the window size was continuously shrunk during the iterative process to perform an accurate search of the feature variables．It was validated on a corn dataset，and compared with the partial least squares models established by the full-spectrum method，genetic algorithm，competitive adaptive reweighted sampling，and bootstrapping soft shrinkage approach．The results showed that the new method had significant advantages in terms of both accuracy and stability．Taking corn protein content prediction as an example，the root mean square error of prediction of ISWBOSS was reduced from 0.041 8 to 0.010 3，compared with the bootstrapping soft shrinkage approach．Moreover，the new method required fewer iterations and higher operational efficiency to reach the optimal model，which was a guideline for improving the performance of NIR spectral calibration models．