Forecasting and selling futures using ARIMA models and a neural network
Loading...
Files
Date
1997-01-01T00:00:00Z
Authors
Holens, Gordon Anthony
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This study involves comparing the forecasting and trading performance of an ARIMA model and a neural network model. The optimal ARIMA model is selected by choosing the combination of sample size and forecast ahead period that produce the minimum forecast error. Weekly data for two contracts traded on the futures exchanges are used. Results suggest that a mid range sample size together with the minimum forecast ahead period produces the lowest forecast error. Secondly, a neural network using the optimal sample size and forecast ahead period chosen from above is compared to the ARIMA model. It turns out that the neural network is able to lower the forecast error. This study also checks for the ability of both the ARIMA and neural network models to detect turning points in the market. It turns out that both models for both commodities are able to predict turning points with about the same degree of accuracy. Lastly, the optimal ARIMA model together with the neural network model are used to trade futures contracts using a given trading strategy. The models all produce negative profits but the neural network suffers smaller losses per trade and trades slightly more often. Neither the neural network or the ARIMA models were able t sell at a significantly higher price than the overall average selling price. Overall, the negative profits produced by the models together with the low percentage of profitable trades may indicate that the trading regime is not appropriate. It may also suggest that the neural network is over fitting the data or that the ARIMA model is not well specified.