Testing for normality with neural networks

In this paper, we treat the problem of testing for normality as a binary classification and construct feedforward neural network that can act powerful test. We show by changing its decision threshold, control frequency false non-normal predictions thus make more similar to standard statistical tests. also find optimal thresholds minimize total error probability each sample size. The experiments conducted on samples with no than 100 elements suggest our method is accurate selected tests almost all types alternative distributions sizes. particular, was most fewer 30 regardless distribution type. Its accuracy increased Additionally, when decision-thresholds were used, very larger 250–1000 elements. With AUROC equal 1, overall. Since data an assumption numerous techniques, constructed in study has high potential use everyday practice statistics, analysis machine learning.