CARNet Improves Multivariate Time Series Forecasting with Cycle-Conditioned Aggregation

Multivariate time series forecasting often struggles with accurately modeling complex dependencies between variables, especially when strong periodic patterns are present. Traditional attention-based models, while powerful, suffer from quadratic complexity, making them inefficient for large datasets. Newer attention-free aggregation models offer linear complexity but often overlook the crucial global periodic structures inherent in much time series data. A new framework, CARNet (Cycle-Conditioned Core Aggregation and Redistribution), addresses these limitations. It integrates global recurrent cycle information directly into its core-based interaction modeling through a Multihead Core Aggregation mechanism. This approach allows CARNet to explicitly leverage periodic patterns, which is a common characteristic in many real-world time series. Extensive experiments show that CARNet consistently outperforms both transformer-based and other non-attention baselines across various prediction horizons. Crucially, it achieves these superior results while maintaining a linear complexity for cross-variate dependency modeling, making it more scalable and efficient for practical applications.