Abstract
In recent years, transformer-based models have gained prominence inmultivariate long-term time series forecasting (LTSF), demonstratingsignificant advancements despite facing challenges such as high computationaldemands, difficulty in capturing temporal dynamics, and managing long-termdependencies. The emergence of LTSF-Linear, with its straightforward lineararchitecture, has notably outperformed transformer-based counterparts,prompting a reevaluation of the transformer's utility in time seriesforecasting. In response, this paper presents an adaptation of a recentarchitecture termed extended LSTM (xLSTM) for LTSF. xLSTM incorporatesexponential gating and a revised memory structure with higher capacity that hasgood potential for LTSF. Our adopted architecture for LTSF termed as xLSTMTimesurpasses current approaches. We compare xLSTMTime's performance againstvarious state-of-the-art models across multiple real-world da-tasets,demonstrating superior forecasting capabilities. Our findings suggest thatrefined recurrent architectures can offer competitive alternatives totransformer-based models in LTSF tasks, po-tentially redefining the landscapeof time series forecasting.