Abstract
•We build-up a vector ANN to be exploited as a nonlinear VAR model.•The scheme is improved with a wavelet-based data preprocessing task.•A comparison to VAR models from which data were simulated shows its superiority.•The model is validated for real-world extremely fluctuating data.•The wavelet-based vector ANN outperforms the ANN-based NAR model.
The paper proposes a parsimonious nonlinear framework for modeling bivariate stochastic processes. The method is a vector autoregressive-like approach equipped with a wavelet-based feedforward neural network, allowing practitioners dealing with extremely random two-dimensional information to make predictions and plan their future more and more precisely. Artificial Neural Networks (ANN) are recognized as powerful computing devices and universal approximators that proved valuable for a wide range of univariate time series problems. We expand their coverage to handle nonlinear bivariate data. Wavelet techniques are used to strengthen the procedure, since they allow to break up processes information into a finite number of sub-signals, and subsequently extract microscopic patterns in both time and frequency fields. The proposed model can be very valuable especially when modeling nonlinear econophysical systems with high extent of volatility.