Multi-Period High Dimensional Data Modeling Using Hybrid Zero-Convolution CNN-LSTM for Improved Crude-Oil Price Prediction

The prediction of crude oil prices poses a difficult yet pertinent task across various domains. Therefore, there are dozens of studies in this field to predict its price as accurately as possible. In the majority of these studies, the input signal is univariate. However, when the input signal becomes multivariate and the number of channels (dimension) increases, the error is more likely to rise for many reasons, e.g., overfitting. Moreover, when the available data spans multiple periods (e.g., daily, weekly, and monthly), utilizing these diverse time spans for enhanced forecasting accuracy presents a notable challenge. Addressing these complexities, the present study proposes a hybrid zero-convolutional CNN-LSTM model. This model first involves training a CNN-LSTM network for each distinct period, where the CNN component serves to compress the multivariate signal into a univariate format, effectively mitigating overfitting. Subsequently, the CNN outputs are fed into the LSTM for the final forecasting stage. Furthermore, to leverage the entirety of the available periods for enhanced prediction, the weekly and monthly models are integrated with the daily model using zero convolutional layers to protect the daily model against the initial noise while enhancing its performance utilizing different periods' data. The model was trained on a dataset in which the input signal has 39 channels (dimensions) in different periods (i.e., daily, weekly, and monthly), and the target was the West Texas Intermediate (WTI) crude oil price changes. Ultimately, it was compared to some famous financial time series prediction models and the benchmark using several metrics, which showed the superiority of the introduced model.