Breaking Silos: Adaptive Model Fusion Unlocks Better Time Series Forecasting

TL;DR

TimeFuse addresses the fragmentation in time-series forecasting by showing that no single model dominates every input. It introduces a sample-level adaptive fusion framework composed of a meta-feature extractor and a fusor that outputs per-sample weights to combine base model predictions, enabling dynamic integration of a diverse model zoo. Trained with cross-task meta-training and data mixing, the fusor generalizes to unseen datasets and delivers strong improvements over state-of-the-art baselines and static ensembles across long- and short-term forecasting tasks, with interpretable fusion weights. The approach is lightweight, scalable, and adaptable as new models emerge, offering practical benefits for real-time and multi-task forecasting scenarios.

Abstract

Time-series forecasting plays a critical role in many real-world applications. Although increasingly powerful models have been developed and achieved superior results on benchmark datasets, through a fine-grained sample-level inspection, we find that (i) no single model consistently outperforms others across different test samples, but instead (ii) each model excels in specific cases. These findings prompt us to explore how to adaptively leverage the distinct strengths of various forecasting models for different samples. We introduce TimeFuse, a framework for collective time-series forecasting with sample-level adaptive fusion of heterogeneous models. TimeFuse utilizes meta-features to characterize input time series and trains a learnable fusor to predict optimal model fusion weights for any given input. The fusor can leverage samples from diverse datasets for joint training, allowing it to adapt to a wide variety of temporal patterns and thus generalize to new inputs, even from unseen datasets. Extensive experiments demonstrate the effectiveness of TimeFuse in various long-/short-term forecasting tasks, achieving near-universal improvement over the state-of-the-art individual models. Code is available at https://github.com/ZhiningLiu1998/TimeFuse.

Figures (6)

• Figure 1: Sample-level inspection reveals that each time series model excels in a considerable fraction of test samples, highlighting their unique strengths for certain types of input. TimeFuse adaptively leverages the strengths of different models for each input time series, achieving dynamically fused forecasting that outperforms the best individual model on up to 95.1% samples.
• Figure 2: The TimeFuse framework for ensemble time-series forecasting, best viewed in color.
• Figure 3: Comparison between TimeFuse and static ensemble methods with validation top-k model filtering.
• Figure 4: Visualization of the average model fusion weights by TimeFuse across datasets, with variation bars showing standard deviations between samples within the dataset. TimeFuse adaptively produces diverse ensemble strategies, effectively supporting tasks that benefit from either a broad model ensemble (e.g., ETTh1/m1) or a more selective integration of specific models (e.g., Electricity/Traffic).
• Figure 5: TimeFuse achieves increasingly better forecasting performance as more base models are included in the model zoo.