Research map / Time series / Long-term forecasting
Breaking Silos: Adaptive Model Fusion Unlocks Better Time Series Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TimeFuse, a sample-level adaptive ensemble that uses a learned linear fusor over 24 time-series meta-features to weight predictions from 13-14 forecasting models, improving over individual SOTA models on 16 benchmarks.
Problem
No single time-series forecasting model consistently wins across test samples; static ensembles require per-dataset model selection and fail to adapt per input.
Approach
Train a single-layer softmax fusor on task-agnostic meta-features (statistical, temporal, spectral, multivariate) of each input to predict per-sample fusion weights over a model zoo, meta-trained on validation-set predictions pooled across datasets so it can transfer zero-shot to unseen datasets.
Key results
- TimeFuse achieves average MSE on long-term forecasting of 0.257 vs. 0.265 for the best static-ensemble baseline (forward selection) and reduces MSE by 3.61% relative to TimeXer across seven datasets.
- On PEMS short-term forecasting, TimeFuse reduces MAPE versus the top individual models by 20.46% (TimeXer), 9.89% (iTransformer), and 15.37% (LightTS).
- Zero-shot TimeFuse (trained without the target dataset) remains competitive with the best individual model on most datasets, e.g., ETTh1 MSE 0.3707 vs. 0.3770 for best individual.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use sample-level adaptive fusion instead of the strongest individual model | Time-series forecasting · Long-term benchmarks (ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, Traffic) | MSE: average MSE reduction of 3.61%/6.88%/11.77%/8.39% across seven datasets w.r.t TXer/Tmixer/PAttn/iTF | Helped |
| Use sample-level adaptive fusion instead of the strongest individual models | Time-series forecasting · PEMS datasets | MAPE: 20.46%, 9.89%, and 15.37% | Helped |
| Use sample-level adaptive fusion instead of static ensemble strategies | TimeFuse · Long-term forecasting datasets | forecasting error | Helped |
| Use zero-shot rather than normal meta-training | TimeFuse · ETTm2 | MSE: 0.1701 -> 0.1721 | Hurt |
| Use zero-shot rather than normal meta-training | TimeFuse · ETTh1, ETTh2, ETTm1, Weather, Electricity, Traffic, PEMS03, PEMS04, PEMS07, PEMS08 | MSE or MAE | Mixed |
| Replace the 24-feature set with TSFEL features (165 variables) (ablation) | TimeFuse · Long-term average (7 datasets) | MAE: 0.2887 -> 0.2898 | No clear effect |
| Replace the 24-feature set with TSFEL features (165 variables) (ablation) | TimeFuse · Short-term average (PEMS) | MAE: 17.334 -> 17.367 | No clear effect |
| Add diverse base models to the model zoo | TimeFuse · Long-term forecasting datasets | MAE and MSE | Helped |
| Increase the input lookback length | TimeFuse · Long-term forecasting datasets | MSE and MAE | Mixed |
Surprising findings
- The best individual model wins on only up to 23.2% of test samples, suggesting no universal winner even among SOTA models.
- Chronos-Bolt and AutoGluon show abnormally high errors on long-term forecasting, which the authors attribute to the model's pretraining prediction horizon limit.
Limitations
- Requires training and storing many base models and their validation predictions, so cost scales with model zoo size.
- AutoGluon and foundation-model comparisons were limited by time and only tested at short horizons.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic PEMS03 PEMS04 PEMS07 PEMS08 EPF-NP
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic PEMS03 PEMS04 PEMS07 PEMS08 EPF NP
Metrics: MSE MAE RMSE MAPE
Baselines: TimeXer TimeMixer PAttn iTransformer PatchTST TimesNet DLinear AutoGluon Chronos-Bolt
Builds on: TSLib Huber loss Caruana ensemble selection
Closest papers by meaning
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.87)
- SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion (NeurIPS 2024 · similarity 0.87)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts (ICLR 2025 · similarity 0.86)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.86)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.