Does Scaling Law Apply in Time Series Forecasting?

TL;DR

Does the parameter-scaling intuition hold for time series forecasting? The paper introduces Adaptive Linear (ALinear), an ultra-lightweight model that uses horizon-aware adaptive decomposition and progressive frequency attenuation to forecast with far fewer parameters than large Transformers. Across seven benchmarks and ultra-long horizons, ALinear consistently outperforms state-of-the-art baselines while using less than 1% of their parameters, aided by a parameter-aware metric (PNP) that quantifies efficiency. The results show that the relative importance of trend and seasonal components is data-dependent, supporting adaptive design choices and challenging the assumption that bigger models are always better for time series forecasting.

Abstract

Rapid expansion of model size has emerged as a key challenge in time series forecasting. From early Transformer with tens of megabytes to recent architectures like TimesNet with thousands of megabytes, performance gains have often come at the cost of exponentially increasing parameter counts. But is this scaling truly necessary? To question the applicability of the scaling law in time series forecasting, we propose Alinear, an ultra-lightweight forecasting model that achieves competitive performance using only k-level parameters. We introduce a horizon-aware adaptive decomposition mechanism that dynamically rebalances component emphasis across different forecast lengths, alongside a progressive frequency attenuation strategy that achieves stable prediction in various forecasting horizons without incurring the computational overhead of attention mechanisms. Extensive experiments on seven benchmark datasets demonstrate that Alinear consistently outperforms large-scale models while using less than 1% of their parameters, maintaining strong accuracy across both short and ultra-long forecasting horizons. Moreover, to more fairly evaluate model efficiency, we propose a new parameter-aware evaluation metric that highlights the superiority of ALinear under constrained model budgets. Our analysis reveals that the relative importance of trend and seasonal components varies depending on data characteristics rather than following a fixed pattern, validating the necessity of our adaptive design. This work challenges the prevailing belief that larger models are inherently better and suggests a paradigm shift toward more efficient time series modeling.

Figures (6)

• Figure 1: Comparison of model's parameter efficiency on ETTm1 dataset when prediction length = 960. Figure \ref{['fig:introduction']} (a) shows the metrics below different number of layers of Encoder/Decoder of the state-of-the-art (sota) time series forecasting model. Figure \ref{['fig:introduction']} (b) illustrates the comparison of sota models in terms of performance and parameter quantity.
• Figure 2: ALinear model. The model separates the input time series into trend and seasonal components through a decomposer with adaptive kernel size, then separately through component-specific projections, and combines the predictions through adaptive trend-seasonal balancing. The right side of the figure reflects our design concept. As the predicted length varies, the importance of periodicity and trend (the depth of the colors in the figure) is constantly changing.
• Figure 3: Efficiency analysis and forward pass algorithm of ALinear
• Figure 4: Parameter-Normalized Performance (PNP) comparison across prediction lengths. The bar charts show MSE-based (left) and MAE-based (right) PNP scores for each model.
• Figure 5: Dataset-dependent evolution of trend (dark blue) and seasonal (light blue) components across prediction horizons.