A Method of Dynamic Model Checking by Fusion Entropy and Residual Correlation Analysis

Correlation analysis on residuals is a classical checking method for dynamic modeling. Motivated by information theory, the paper presents another approach of model checking based on whiteness test and information entropy, and proposes a new diagnostic quantity named "Residual Entropy", which indicates the amount of information left in residual series. On normal distribution hypothesis about residuals, the paper gives the estimation expressions of residual entropy, and illustrates its application. The result of experiment shows that residual entropy is a useful diagnostic quantity in time series modeling. By comparison with the conventional methods in performance, the paper demonstrates that minimum residual entropy can be applied as an effective checking criterion.