A non-parametric Approach to simulate Panel Data.

Real world multivariate data mostly contains correlation structure because generally some variables tend to have a similar behavior or some dependency structure. This is due to the nature of data generation process. The variables can contain cross-sectional correlation, correlation between variables at a given time stamp, and temporal correlation, correlation between various observations of a given variable at various timestamps. Also, there are datasets which contain both cross-sectional and temporal correlations. Modeling of such correlation structure in the data is important because it provides us the predictive power and affects the Machine Learning algorithms in various ways, due to relationships within variables. We propose a methodology to simulate a cyber activity dataset, containing both cross-sectional and temporal correlations, also called Panel dataset. The proposed methodology uses non-parametric approach to induce correlations among feature vectors and across time without disturbing the marginal distributions of features.