Power study of statistical tests for network traffic change detection

Analysis of large data volumes calls for reliable and robust methods. Statistical features specific to telecommunication network traffic - such as self-similarity and heavy-tailed distributions - ruin some of the traditional analysis tools. We emphasize the role of statistics in data analysis and present how Monte Carlo simulation can help when analytical solutions are not available. Our main result is that median works well as a test statistic with network traffic data, which we prove by conducting a power study with data collected from a live network.