A method of analysis of models of forecast in linear and nonlinear unidimensional chronological successions / Um método de análise e previsão de sucessões cronológicas unidimensionais lineares e não-lineares

AUTOR(ES)
DATA DE PUBLICAÇÃO

2004

RESUMO

The main objective of this work was it of exploring the possibility to use a methodology capable to decompose a temporary series through ondaletas, jointly with the econometrics models and of neural network already existent of forecast and to compare the quality of forecasts obtained for chronological successions no lineal simulated. The proposal was reached mainly by the elaboration of a flowchart for treatment of the forecasts of chronological successions to put a more appropriate quantitative rigidity. The differential of this work was in the accomplishment of the forecasts inside of the sub-series decomposed by an ondaleta in up to two levels, and being obtained the forecast of the original series through reconstruction of the series for models built by generating processes of data of no-lineal chronological successions. They were simulated series of a process ARIMA-GARCH, a process ARIMA, a process bilinear and a series of a movement browniano. The main work was constituted in the elaboration of the pré-processing phase and of the forecasts static in separate for each one of the found sub-series being done for 10 and 200 future observations. Besides the punctual forecasts it was also verified the envelopamento of the data, that it consists of comparing the forecast model through a trust interval for the values foreseen in a thousand simulated series by the same seed. The results appeared that for a model ARIMA(1,0,0)-GARCH(1,1), it can be observed that the pré-processing for the ondaleta went better so much for just a stage of separation of discharges and low frequencies for the correlation as for the criteria of TIC being this reduced and for smaller MAPE for the forecasts of short period. Already for the models of nets neurais an important difference that it should be emphasized between the nets appealing neurais and the nets with retropropagação algorithm is the capacity of forecast of the appealing nets for no-lineal data with 2 pré-processing levels and for forecasts of short period. Though, already for the criterion of the envelopamento, the best results went to the appealing nets in the forecast of the process ARIMA-GARCH and bilinear and pré-processing with 1 level. All the data were also compared with the forecasts done without pré-processing, which were shown inappropriate with MAPE close to 100% for forecasts of long period. It was also checked in this work the alterations that the change of the choice of an ondaleta for other, was able to impactar in the results of the future forecasts. It was verified that the change in the wave way in the pré-processing that if it seems more visually with the form of the data of the series, it reduces the acurácia measures in 48%, leaving evidences that it can have improvements in the results. In the practical analysis for IBOVESPA, the results were not satisfactory, because the best results are for appealing nets with 1 pré-processing level. Likewise, of the analysis of this work, the importance emerges given to the flowchart implemented for the forecasts and the paper of the forecasts in separate for ondaletas as reducers of the mistakes in the processes estocásticos, and of the implementation of the bands of forecasts for appealing nets for no-lineal chronological successions.

ASSUNTO(S)

neural networks wavelets forecasting garch time series sucessões cronológicas envelopamento ondaletas previsão redes neurais recorrentes

ACESSO AO ARTIGO

Documentos Relacionados