Analysis of the Baltic Exchange Dry Index Using Data Mining Techniques

Abstract

This report discusses the preliminary results for the analysis of the Baltic Dry Index (BDI) using data mining techniques. The overall aim of the project is to investigate ways to identify trends or stages in the BDI associated with economic cycles, as well as other relationships with economic indicators. The dataset consisted of prices for economic indicators from January, 1985 to December, 2008, including monthly average prices of Copper, Oil, Gold, Silver and the Dow Jones Industrial monthly index values.The work described the tasks of Cycle and Stage recognition as a specific type of classification problem and the work of Martin Stopford Maritime Economics (2003) has been used as a starting point. By definition, a complete cycle has four stages. A market trough (stage 1) is followed by a recovery (stage 2), leading to a market peak (stage 3), followed by a collapse (stage 4). Several alternative classification models were developed and were trained to recognize these stages. Furthermore, these models were evaluated according to their classification accuracy measured by their hit ratio. Even though classification capacity was critical to discriminate between models, emphasis was also placed on whether the developed models can be easily understood by the end users.Based on the results reported, we believe that the approach adopted was successful in identifying stages and cycles in the BDI, with an average power of classification of 72% in an out-of-sample set. Additional pieces of interesting knowledge that have been identified include the fact that during the recent years the BDI has been more linked in the model with the copper prices, and that usually stock market prices have a limited effect on it.We believe that the classifiers were able to replicate the approach taken by human beings of interpreting the shapes of the BDI, strength that is remarkable considering the evident cyclic character that the BDI showed from January, 1985 to more or less December, 2000, and the increase that followed after the break point.