<p>Data normalization is a pre-processing technique which involves scaling or transforming data so that each feature can contribute equally. In order to produce a generalized prediction model of a classification problem, the machine learning algorithms rely on the quality of data. Numerous research works have demonstrated the value of data normalization in enhancing data quality, and in turn, improves the performance of machine learning algorithms. In this paper, a new normalization technique based on piecewise-continuous, symmetric estimation function and <i>tanh</i>-transformation has been proposed. The breakpoints <i>a</i> and <i>b</i> are the harmonic mean and arithmetic mean of a particular feature for a given dataset. The work aims to investigate the impact of the proposed normalization technique on the datasets having varying sizes, classes, data ranges, fields. We have compared its performance with the existing thirteen normalization methods on the basis of classification measures and statistical tests- Wilcoxon Signed Rank test, Friedman test and Mann Whitney-U test. Experiments are performed on twenty publicly available real valued datasets with different properties to validate the distribution-free characteristic of the proposed normalization. Also, the outcomes are analyzed on the basis of the area under the receiver operating characteristic curve, Cohen’s Kappa coefficient and statistical measures. The changes in the statistical measures of central tendency and dispersion for <i>tanh</i>-normalization and the proposed normalization are compared. It has been observed from the experimentation that the proposed normalization provides superior results over the other normalization methods for the given datasets. The statistical tests show that the proposed normalization attains better rank as compared to the existing thirteen normalization methods.</p>