engleski

Using Support Vector Regression Machines to Forecast Natural Gas Prices

With the substitution of coal and wide proliferation of natural gas usage the forecasting of gas prices has become one of the most critical issues in utility’ planning and operations. Accurate natural gas price forecasting and the direction of price changes are paramount since these forecasts are used in commodity trading, electricity production planning and regulatory decision making, covering both the supply and demand side of natural gas market. This paper presents the results of short-term forecasting of US Henry Hub spot gas prices based on the strategic seasonality-adjusted support vector regression machines (SSA-SVR). Several improvements to the forecasting method based on SVR are introduced. The improvements we propose are made by taking into account that the successfulness of any method (including the SVR) to forecast natural gas prices depends on the available data. An additional objective of the proposed approach is to reduce the human input in the building of the forecasting model. The proposed use of FS algorithms for automatic model input selection as well as the use of the global optimization PSwarm method to optimize the SVR hyper parameters further reduces the human input. We generate the inputs of the model in such a way that it enables us to capture the seasonal effects. Adding the binary variables, such as days in a week and months in a year, as inputs often enhances the model accuracy. Furthermore, the use of FS algorithm improves the forecasting accuracy compared to expert based input selection. The forecasting results of 1- day ahead prices show a clear dominance of Steepwise FS algorithm both in terms of RMSE and MAPE statistics. In the case where we allowed ten variables in the FS algorithm Steepwise FS and GA come to a tie. In the five variable FS case there is only a marginal improvement of RMSE and MAPE statistics obtained by adding days, months and temperatures. We can conclude that there is no real gain in adding additional variables to the basic set of variables. In the ten variable FS case there is a more noticeable gain in accuracy especially in adding days in a week and temperatures variables to the basic set. The forecasting results for 1- week ahead gas prices display greater difference among the analyzed FS algorithms but again Steepwise FS clearly dominates over the competition in both the RMSE and MAPE categories. The superiority of Steepwise FS is especially pronounced in the case when ten variable FS is considered. Unlike the 1-day ahead forecasts there is no gain in adding months to the basic set of variables, since the addition worsens the RMSE and MAPE statistics. Our results point to the conclusion that the SVR results reported in the literature often over exaggerate the successfulness of these models since we record only marginal improvements over a classical, simplistic approach to short term asset price forecasting.

Forecasting, Henry Hub, Natural Gas, Support Vector Regression Machines

nije evidentirano