Shipping market time series forecasting via an Ensemble Deep Dual-Projection Echo State Network

The Echo State Network (ESN) is a specific type of recurrent neural network characterized by random initialization of recurrent and input connections. The reservoir structure effectively reduces computational workload. However, the redundancy in the reservoir network can degrade predictive accuracy, especially in time series forecasting tasks. The challenge of extracting dependable and meaningful features from high-dimensional reservoirs remains an open question. Additionally, the presence of high-dimensional reservoirs in small datasets can lead to ill-conditioned problems. Therefore, the objectives of this study are twofold: to reduce reservoir dimensionality and extract meaningful features concurrently. We first employ a dual-projection module to fine-tune the reservoir and build an ensemble deep architecture to achieve this. Empirical experimentation confirms the superior predictive performance of our proposed approach when applied to shipping market datasets, outperforming several baseline models and traditional benchmarks.