How Much Can Time-related Features Enhance Time Series Forecasting?

TL;DR

This work argues that explicit encoding of time-related features can significantly boost long-term multivariate time series forecasting. By introducing TimeStamp Forecaster (TimeSter) and combining it with a simple linear Backbone (TimeLinear), the approach achieves strong, data-efficient performance, especially on datasets with clear periodicity, while maintaining computational efficiency. Through extensive ablation studies and analyses, the authors show when and why time-related features help, including feature selection, encoder/decoder design, and robustness across seeds. The work also highlights limitations, such as limited modeling of cross-variable time interactions, and outlines directions for integrating such interactions in future work. Overall, TimeSter provides a practical, plug-and-play mechanism to leverage time semantics in forecasting systems.

Abstract

Recent advancements in long-term time series forecasting (LTSF) have primarily focused on capturing cross-time and cross-variate (channel) dependencies within historical data. However, a critical aspect often overlooked by many existing methods is the explicit incorporation of \textbf{time-related features} (e.g., season, month, day of the week, hour, minute), which are essential components of time series data. The absence of this explicit time-related encoding limits the ability of current models to capture cyclical or seasonal trends and long-term dependencies, especially with limited historical input. To address this gap, we introduce a simple yet highly efficient module designed to encode time-related features, Time Stamp Forecaster (TimeSter), thereby enhancing the backbone's forecasting performance. By integrating TimeSter with a linear backbone, our model, TimeLinear, significantly improves the performance of a single linear projector, reducing MSE by an average of 23\% on benchmark datasets such as Electricity and Traffic. Notably, TimeLinear achieves these gains while maintaining exceptional computational efficiency, delivering results that are on par with or exceed state-of-the-art models, despite using a fraction of the parameters.

Figures (10)

• Figure 1: (a) Comparison of forecasting using only time-related features, historical observations, and other baselines on Electricity with a look-back window of 96 and a future horizon of 720. (b) 100 people's (indexed as 190-290 on Electricity) average electricity consumption at noon and midnight every Monday during the first half of 2018. (c) Temperature prediction using time-related features, where each hour and season is associated with some temperature values.
• Figure 2: Overview of the proposed method. (a) The TimeSter module encodes historical time-related features and predicts future values. (b) The BonSter, i.e., any backbone model, utilizes historical observations of multivariate time series to generate predictions. Their outputs are weighted and summed to yield the final prediction. The model after Norm could be replaced with any backbone. Here, we use a linear layer and name the whole model TimeLinear.
• Figure 3: Distribution of training data across datasets at noon on Mondays. From left to right, ETTh2: Variate HUFL, Electricity: Variate 1, Weather: Variate p (mbar), and Traffic: Variate 1. From this, we can see a significant timestamp-related distribution.
• Figure 4: Accuracy, parameter count, and training time of TimeLinear and other models on ETTh2, with historical window $L=96$ and future horizon $T=720$.
• Figure 5: Performance promotion with longer historical windows. The forecasting length is $96$.