ITS-Mina: All-MLP Framework for Multivariate Time Series Forecasting

A new all-MLP framework called ITS-Mina has been proposed for multivariate time series forecasting. The framework integrates three key innovations: an iterative refinement mechanism that progressively enhances temporal representations through a shared-parameter residual mixer stack, an external attention module replacing traditional self-attention with learnable memory units to capture cross-sample global dependencies, and a Harris Hawks Optimization-based hyperparameter tuning strategy. The paper, published on arXiv (2604.27981), demonstrates that MLP-based models can achieve competitive or superior performance compared to Transformer-based architectures with significantly reduced computational cost. The method is designed for applications such as financial analysis, energy management, and traffic planning.