Méthodes de recherche sur les fluctuations longues

The answer to the question, whether long waves exist or not depends mainly on the statistical methods used for analysis. Since the seminal work of Kondratieff in the 1920th these methods underwent significant changes. The fundamental critic against the traditional approaches in the 1970th lead to the propagation and development of ?ideal? filters which should allow the isolation of predefined frequency components from time series in an exact way. This filter-design-approach was challenged in the 1980th by the unit root hypothesis. If a time series follows a unit root process filters, even if they are ?ideal? lead to spurious long term cycles with duration typical for Kondratieff cycles. Therefore no longer ?ideal? filters were propagated but stochastic models which correspond to the properties of the time series analyzed. This model-based-approach was challenged again in the 1990th by the development of filters which seem appropriate also for time series which contain a unit root. The article will discuss in a first part the historical development of traditional-, filter-designand model-based-approaches comparing their results with regard to the Kondratieff cycles systematically. In a second part the article presents the state of the art of filter-designand model-based-approaches. It will be argued that filters and stochastic models are only adequate if outliers and structural breaks are considered explicitly. If such interventions are modelled correctly it can be shown that filters and models lead to highly similar results. The methodological considerations are illustrated by long term time series which are often used in empirical research on long wave analysis.