摘要
Forecasting mineral commodity(MC) prices has been an important and difficult task traditionally addressed by econometric, stochastic-Gaussian and time series techniques. None of these techniques has proved suitable to represent the dynamic behavior and time related nature of MC markets. Chaos theory(CT) and machine learning(ML) techniques are able to represent the temporal relationships of variables and their evolution has been used separately to better understand and represent MC markets. CT can determine a system's dynamics in the form of time delay and embedding dimension. However, this information has often been solely used to describe the system's behavior and not for forecasting.Compared to traditional techniques, ML has better performance for forecasting MC prices, due to its capacity for finding patterns governing the system's dynamics. However, the rational nature of economic problems increases concerns regarding the use of hidden patterns for forecasting. Therefore, it is uncertain if variables selected and hidden patterns found by ML can represent the economic rationality.Despite their refined features for representing system dynamics, the separate use of either CT or ML does not provide the expected realistic accuracy. By itself, neither CT nor ML are able to identify the main variables affecting systems, recognize the relation and influence of variables though time, and discover hidden patterns governing systems evolution simultaneously. This paper discusses the necessity to adapt and combine CT and ML to obtain a more realistic representation of MC market behavior to forecast long-term price trends.
Forecasting mineral commodity(MC) prices has been an important and difficult task traditionally addressed by econometric, stochastic-Gaussian and time series techniques. None of these techniques has proved suitable to represent the dynamic behavior and time related nature of MC markets. Chaos theory(CT) and machine learning(ML) techniques are able to represent the temporal relationships of variables and their evolution has been used separately to better understand and represent MC markets. CT can determine a system's dynamics in the form of time delay and embedding dimension. However, this information has often been solely used to describe the system's behavior and not for forecasting.Compared to traditional techniques, ML has better performance for forecasting MC prices, due to its capacity for finding patterns governing the system's dynamics. However, the rational nature of economic problems increases concerns regarding the use of hidden patterns for forecasting. Therefore, it is uncertain if variables selected and hidden patterns found by ML can represent the economic rationality.Despite their refined features for representing system dynamics, the separate use of either CT or ML does not provide the expected realistic accuracy. By itself, neither CT nor ML are able to identify the main variables affecting systems, recognize the relation and influence of variables though time, and discover hidden patterns governing systems evolution simultaneously. This paper discusses the necessity to adapt and combine CT and ML to obtain a more realistic representation of MC market behavior to forecast long-term price trends.
