The procedure of business cycle turning points identification based on hidden Markov models

In the paper the procedure, based on hidden Markov chains with conditional normal distributions and uses algorithms such as time series decompositions (STL), Baum-Welch algorithm, Viterbi algorithm and Monte Carlo simulations, is proposed to analyze data out of the business tendency survey conducted by the Research Institute for Economic Development, Warsaw School of Economics. There are considered three types of models, namely, with two-state, three-state and four-state Markov chains. Results of the procedure could be treated as an approximation of business cycle turning points. The performed analysis speaks in favor of multistate models. Due to, an increasing with the number of states, numerical instability, it is not obvious which model should be considered as the best one. For this purpose various optimization criteria are taken into consideration: information criteria (AIC, BIC) and the maximum-likelihood, but also frequency of obtaining a given set of parameters in the Monte Carlo simulations. The results are confronted with the turning points dated by OECD. The tested models were compared in terms of their effectiveness in detecting of turning points. The procedure is a step into automation of business cycle analysis based on results of business tendency surveys. Though this automation covers only some models from millions of possibilities, the procedure turns out to be extremely accurate in business cycle turning points identification, and the approach seems to be an excellent alternative for classical methods.