Tendências comuns em modelos estruturais de séries de tempo: uma aplicação ao preço da soja no Brasil e nos Estados Unidos

Resumo

The multivariate approach in Structural Time Series Models (STSM) fashion permits an empirical investigation of the presence of common trends which can be modeled using the Kalman Filter to decompose its components. This process define the Seemingly Unrelated Time Series Equations (SUTSE) (Harvey 1989) and permits to specify a long run relationship between variables or the presence of cointegration. The SUTSE models differentiate from the univariate approach in permiting that each series to be modeled in the same manner that in a univariate approach, but extending the results from components decomposition to more complex relationships between the variables explained by the presence of common trends, which can produce more accurate forecasts. The goal of this paper is then investigate the presence of common trends and its influence on forecasts for the Brazilian and U.S. soybean price, using for this the test developed by Nyblom and Harvey (1997) and Nyblom and Harvey (1999) and the SUTSE models. Our results suggestthe existence of cointegration between the variables and more efficient forecasts to the SUTSE models than the similar univariate models.