Predictive analytics of crude oil prices by utilizing the intelligent model search engine

Highlights: • An automated model estimation and selection algorithm is proposed. • The search algorithm utilizes a forward-looking model validation process. • This paper proposes a fair testing procedure for predictive analytics. • The proposed expert system does not require any prior theory or assumption. • This architecture re-estimates and re-searches the model with new data. This paper proposes an intelligent model search engine (IMSE), an integrated model selection algorithm, subject to the out of sample predictive performance and given set of explanatory variables for forecasting crude oil prices. In the conventional applications of energy price forecasting, models are selected based on preliminary assumptions on causality and model structure (e.g. lag length in lagged variables). Relaxation of those assumptions would cause over-fitting and reduce the degree of freedom. Considering the ultimate objective of forecasting models, any variations of models may be tested in the out-of-sample period, and the optimization problem can be redefined as minimization of post-sample error metric in a validation set. By this, data mining would be a legitimate operation for economic forecasting, and it also proves required conditions usually tested by diagnostic tests such as Akaike Information Criterion for model quality. IMSE is a multi-input/single output difference equation based approach which allows users to test various models (for given set of explanatory variables) as well as various order of lagged inputs (lag length) without a priori assumption or theoretical basis except defining set of potential inputs. Finally, it selects the best model subject to predictive accuracy in a validation set. Empirical results indicated that the proposed algorithm significantly outperformed a broad range of benchmark methodologies as well as proving that certain assumptions of econometric approach (e.g. statistical significance of explanatory variables) are independent of predictive performance.