In time series analysis, linear models that have been broadly used, such as linear regressions and auto-regressive models, are not able to capture the some times non linear nature of some data, offering poor estimation results. Financial series, for instance, show that kind of behavior. Over the last years, a great number of non linear models have been developed in order to analyze time series, some of them statistical, others based on computational intelligence techniques such as neural networks. The purpose of this dissertation is to analyze the performance of the STAR-Tree model under distinct scenarios that differ in model specification, parameterization and estimation methodologies. This class of models splits time series data into individual regions which fulfill the criteria set up by functions called pertinences. A linear model then is selected for each one of those regions. Each estimated data point can belong to one of the mentioned regions with some degree of pertinence, supplied by the above mentioned pertinence functions. Aspects like the proximity between regions, the smoothness of the pertinence functions and the lack of diversity in real data can significantly affect the estimation of models. In order to evaluate the quality of the estimations under the different proposed scenarios, a software was developed with the capabilities of generating artificial time series, importing external series, estimating them under the STAR-Tree model, and generating Monte Carlo simulations that evaluate the quality of parameter estimation and the tree structure detection capability of the model. The software was used as the single tool to generate this dissertation’s analyses, and allowed that different model specifications and methods could be tested without difficulty.