Page 335 - Data Science Algorithms in a Week

316 Mayra Bornacelli, Edgar Gutierrez and John Pastrana

Comparison of Neural Networks and Regression Trees for Predicting
the Price of Thermal Coal

Table 2 represents the error rate calculated for predicting the price of thermal coal
with neural networks (8 and 12 input variables) and regression trees, where we can see
how the prediction of the neural networks with 12 input variables indicated the best
prediction.

Table 2. Prediction errors for the neural networks and regression trees

CONCLUSION

According to the consensus (based on the Delphi methodology), we obtained 25
variables, that were considered the most important ones for the price of thermal coal.
These variables and their potential trends were used to train neural networks and
regression trees. The utilization from correlations and cross validations with the neural
network architectures and the processes of MARS provided the following variables in
order of importance:

 Price of Oil,
 Development of Renewable energy in China,
 Oversupply of the thermal coal market,
 China’s economy (ratio of the Yuan/US dollar),
 Development of Renewable Energy in the United States and
 Transportation Costs of the thermal coal.

We also found how each of these variables model the price of coal using neural
networks and regression trees. Neural networks have the best prediction for the price of
thermal coal. Trends are very important to consider too.