AverageTime: Enhance Long-Term Time Series Forecasting with Simple Averaging

Multivariate long-term time series forecasting aims to predict future sequences by utilizing historical observations, with a core focus on modeling intra-sequence and cross-channel dependencies. Numerous studies have developed diverse architectures to capture these patterns, achieving significant improvements in forecasting accuracy. Among them, iTransformer, a representative method for channel information extraction, leverages the Transformer architecture to model channel-wise dependencies, thereby facilitating sequence transformation for enhanced forecasting performance. Building upon iTransformer’s channel extraction concept, we propose AverageTime, a simple, efficient, and scalable forecasting model. Beyond iTransformer, AverageTime retains the original sequence information and reframes channel extraction as a stackable and extensible architecture. This allows the model to generate multiple novel sequences through various structural mechanisms, rather than being limited to transforming the original input. Moreover, the newly extracted sequences are not restricted to channel processing; other techniques such as series decomposition can also be incorporated to enhance predictive accuracy. Additionally, we introduce a channel clustering technique into AverageTime, which substantially improves training and inference efficiency with negligible performance loss. Experiments on real-world datasets demonstrate that with only two straightforward averaging operations, applied to both the extracted sequences and the original series. AverageTime surpasses state-of-the-art models in forecasting performance while maintaining near-linear complexity. This work offers a new perspective on time series forecasting: enriching sequence information through extraction and fusion. The source code is available at https://github.com/ UniqueoneZ/AverageTime.

💡 Research Summary

AverageTime is a novel framework for multivariate long‑term time‑series forecasting that builds on the channel‑extraction concept of iTransformer but dramatically simplifies the downstream processing pipeline. The authors first extract channel‑wise information using two parallel branches: a Transformer encoder and a multilayer perceptron (MLP). Both branches operate on the transposed input (channels become the sequence dimension), thereby learning complementary representations of inter‑channel dependencies. The outputs of these branches are new auxiliary time‑series that are immediately averaged with the original series, constituting the first fusion step.

Next, each fused series (the original and each auxiliary series) is fed into an independent MLP‑based predictor that forecasts the future horizon. The predictions from all predictors are then averaged a second time to produce the final forecast. Remarkably, only two straightforward averaging operations are required to integrate the rich channel information, avoiding the need for complex gating, attention weighting, or additional learnable fusion parameters.

To address scalability when the number of channels C is large, the authors introduce a channel‑clustering module. They compute Spearman rank correlations on the training set, threshold the correlation matrix (default threshold T = 0.8), and apply a label‑propagation community‑detection algorithm to group highly correlated channels. Channels within the same group share a single set of MLP parameters, turning the per‑channel computation into a per‑group computation. This reduces both memory footprint and runtime, especially in high‑dimensional settings, while incurring negligible loss in predictive accuracy. The lightweight variant, LightAverageTime, demonstrates 1.5–2× speed‑up over the full model.

Beyond channel extraction, the framework is designed to be extensible. The authors integrate a modified Variational Mode Decomposition (VMD) called FVMD, which adds a frequency‑separation penalty to the VMD objective, encouraging the decomposed modes to have distinct central frequencies. The resulting mode‑level components are treated as additional auxiliary series and fused through the same averaging mechanism. Experiments show that FVMD‑augmented models achieve further gains on datasets with strong non‑stationarity or multi‑frequency dynamics.

Complexity analysis reveals that AverageTime operates in near‑linear time O(N·C), where N is the input length and C the number of channels. This is substantially lower than the O(N·C·log N) or O(N·C·d) complexities of many Transformer‑based forecasters. Empirical evaluation on seven public benchmarks—including ETT, Traffic, Electricity, Weather, and others—shows that AverageTime consistently outperforms state‑of‑the‑art models such as iTransformer, Informer, TimesNet, and various linear‑MLP baselines. Relative improvements in Mean Squared Error (MSE) and Mean Absolute Error (MAE) range from 3 % to 7 %, with the most pronounced gains on high‑dimensional datasets where channel clustering is most beneficial.

In summary, AverageTime contributes four key ideas: (1) multi‑branch channel extraction (Transformer + MLP), (2) a two‑stage simple averaging fusion, (3) correlation‑based channel clustering with shared MLPs for efficiency, and (4) optional frequency‑aware decomposition (FVMD) as auxiliary information. Each component independently improves performance, and their combination yields a forecasting system that is both accurate and computationally efficient. By demonstrating that sophisticated forecasting can be achieved with minimalistic operations—essentially two averages—the paper challenges the prevailing belief that ever‑more intricate attention mechanisms are required for state‑of‑the‑art results, offering a compelling direction for future research and practical deployment.