Using Neural Networks for Forecasting of Commodity Time Series Trends

Abstract

Time series of commodity prices are investigated on two scales - across commodities for a portfolio of items available from the database@ of the International Monetary Fund on monthly averages scale, as well as high quality trade event tick data for crude oil futures contract from the market in Japan. The degree of causality is analyzed for both types of data using feed-forward neural network architecture. It is found that within the portfolio of commodities the predictability highly varies from stochastic behavior consistent with the efficient market hypothesis up to the predictability rates of ninety percent. For the crude oil in Japan, we analyze one month (January 2000) series of a mid-year delivery contract with 25,210 events, using several schemes for causality extraction. Both the event-driven sequence grid and second-wide implied time grid are used as the input data for the neural network. Using half of the data for network training, and the rest for validation, it is found in general that the degree of trend extraction for the single next event is in the sixty percent range, which can increase up to the ninety percent range when the symbolization technique is introduced to denoise the underlying data of normalized log returns. Auxiliary analysis is performed that incorporates the extra input information of trading volumes. The time distribution of trading event arrivals is found to exhibit interesting features consistent with several modes of trading strategies.