Integration Advanced Neural Networks Technologies: a new Opportunity for economic Time Series Forecasting

This paper presents the results of integrating transformer outputs into advanced adaptive architectures of convolutional neural networks (CNN) and describes the advantages of this approach compared to baseline models. The novelty of this work lies in the approach to creating models capable of leveraging the strengths of both CNNs and transformers, dynamically adjusting their architecture based on cross-validation results. The strengths of CNNs in handling local patterns through causal convolutions and activation layers are complemented by the transformers' capabilities in processing long-term dependencies, which is crucial for forecasting financial time series that often exhibit noise and unpredictable behavior. Combining these technologies results in robust hybrid forecasting architectures that not only improve validation metrics, indicating higher accuracy in predictions but also reduce model execution time, making them highly practical for forecasting highly dynamic financial series. Additionally, the adaptive approach leads to more stable performance across multiple neural network models with identical parameters, although absolute resolution of this task requires further research.