A fast and qualitatively robust non-parametric modeling method was employed in the modeling of the secondary coating line for the production of fiber optic cable. The system had 24 explanatory variables and one response variable. Variable discrimination in wavelet domain was used in discriminating variables while stepwise regression was used in the determination of model coefficients. The remarkable results showed that in spite of the crudeness of the data (very noisy), de-noising in wavelet proved very effective. Nine key parameters significantly affected the excess length; while two other parameters likely did so, but not within the natural variations of the system.