1. Asymmetries in Forecasting Energy Product Prices APF Conference, Deutsche Bundesbank Frankfurt, Germany 2004 M. E. Malliaris Loyola University Chicago S. G. Malliaris Massachusetts Institute of Technology

2. Introduction Fluctuations in energy costs can cause significant volatility in earnings The ability to forecast energy prices will aid industry in managing energy cost fluctuation and thus the risk associated with energy Energy products rate second only to financial product in amount of trading on futures contracts and that energy commodity price risk has a dominant role in the energy industry

3. Products The five energy-related products – Heating oil – Propane – Natural gas – Gasoline – Crude oil

4. Inter-Relationships Energy products are inter-related Some can be substituted for others when prices rise One of the primary determinants of the price of natural gas is the price of oil Propane is a by-product of both oil refining and natural gas processing

5. Variables daily closing spot price change in daily closing spot price from the previous day standard deviation over the previous 5 days standard deviation over the previous 21 days for each of the five markets

6. Models A multiple regression model was built for each of the five energy markets with the spot price 21 trading days into the future as the dependent variable.. A neural network model was developed for each of the five markets.

7. Regression Results CLHOHUNG Multiple R0.9481540.9496550.9192180.913553 R Square0.8989970.9018450.8449620.83458 Adjusted R Square0.8979770.9005520.8433960.833077 Standard Error2.1999530.0660210.0799850.6553

8. Neural Network Results CLHOHUNGPN Minimum Error-11.524-0.199-0.317-2.461-0.215 Maximum Error10.4030.3990.4804.7460.389 Mean Error-2.152-0.036-0.024-0.1570.01 Mean Absolute Error4.7320.0990.0960.6410.061

9. Conclusions While the nonlinear models derived by the neural network provided superior forecasting in a majority of cases, there remain undeniable asymmetries in the predictive abilities of the three models examined. For crude oil, heating oil, gasoline, and natural gas, the neural network gave the best results, consistently boasting a mean squared error less than half that of the regression or simple predictions.

10. Conclusions, Cont. However, with propane, the neural network gave the least accurate prediction. Also surprising were the results of the simple model, which predicted that the commodity price 21 trading days into the future would hold unchanged from the present day’s price. Except in the case of gasoline, the mean squared error of the simple prediction was lower than that of the regression model.