Live cattle prices forecast at BM&F: a comparison between time series models and neural networks

Abstract

Human forecasting capacity is still very limited. In spite of the extreme efforts of specialists in several different areas for years developing scientific knowledge, forecasting various events, such as climatic conditions at a given time, the evolution of a commodity price in the future, remain subject liable to a considerably high degree of error. Therefore, this paper aims to compare forecast price models for the Live Cattle market at Brazilian Mercantile and Future Exchange (BM&F) using models based in Neural Networks and statistical tools of heteroscedastic times series. The data used correspond to the closing of the live cattle prices, in the period ranging from August 1997 to May 2005, totalizing 1946 observations. The results show the supremacy of neural networks models compared with the AR-EGARCH model, once the Mean Squared Error and the Mean Squared Error Root forecasted were smaller for the neural networks.