Comparaison des méthodes de sélection de structures de modèles non-linéaires en prédiction de séries temporelles: Application à la prévision des cycles endogènes des séries de la production industrielle en Tunisie

摘要

The nonlinear forecast of univariate time-series with discrete time is an exciting area of research. To develop efficient forecasting models, it must be able to understand precisely the related issues, unresolved in their particularities. The major difficulty is to choose from a possible family of future data generating process. However, normally in Econometrics, a real data generation process of observed time series is modeled with a stochastic process. But more important here is to assume that time series are generated by a deterministic dynamical system rather than a random process. In this case, the real issue mentioned is the dichotomy between deterministic and residual part of a data generating process.This thesis is organized as part of the econometrics of non-linear processes and chaos theory and admits on the analysis and prediction of the situation in Tunisia. Our guiding reference simply comes from the need to study and predict the contemporary phenomenon of endogenous instability of fluctuations of variables macroeconomic. Thus our original contribution is an explanation of endogenous cycle fluctuations of the index of industrial production in Tunisia and forecast analysis by threshold models with smooth transition of STAR Type (Smooth transition autoregressive) univariate.From then on, trying to make some theoretical and empirical insights on this subject, our results advocate the idea that the instability of the dynamic balance which is an endogenous weakening mechanism is causing dynamic cyclic structures by bifurcation (endogenous cycles) following a change of qualitative properties of the equilibrium at the level of the dynamics of nonlinear econometric model with a smooth transition of regime shift. But also, this growing fragility can sometimes result in massive and persistent asymmetries that reflect the sensitivity to initial conditions (Butterfly effect), causing endogenous instability of nonlinear model via a chaotic dynamics which is the major source of its inefficiency in anticipation of materials especially over a long period.The work presented in this thesis, with the goal of understanding a better representation of nonlinear dynamical fluctuations of self-sustained endogenous cycles of series of industrial production in Tunisia using the STAR univariate modeling, come from a very interesting personal reflection highlighting the contribution of the various techniques of analysis and forecasting of the economy, knowing that not only the growing interest but also the relevance of the results they provided for the prediction in an area of instability of dynamic structures. However, this presentation is still insufficient to provide a comprehensive and final study of, first, the dynamics of nonlinear models in the econometrics of time series, and secondly of predictive performance models to change plans with smooth transition so specified, to answer definitively many question marks that remain open until now.