Inicjalizacja ukrytych modeli Markowa z wykorzystaniem analizy skupień
Initialization of hidden Markov models by means of clustering analysis

Summary/Abstract: The following paper addresses some major problems which arise in the estimation of Hidden Markov Models, and delves in to the possibilities of augmenting traditional algorithms such as Baum-Welch (EM), with methods known from cluster analysis. Author conducted thorough tests of the proposed heuristics not only using simulated data, but also shows a successful applications on a real world, macroeconomic time series from labor market and business tendency surveys. Low estimation and prediction errors which are presented in the article, testify in favor of proposed method and give good reason to believe that clustering algorithms can constitute either a sound initialization method for Baum-Welch procedure, or a standalone estimation method where likelihood maximization can lead to overfitting.

• Details
• Contents