Research map / Time series / Foundation models for time series

Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation

AAAI 2025 · Time series › Foundation models for time series · approach: Time series foundation models · 9 citations (OpenAlex)

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 TAFAS, a model-agnostic test-time adaptation framework for time series forecasting that uses periodicity-aware partially-observed ground truth and gated calibration modules to adapt frozen source forecasters to shifting test distributions.

Problem

Pre-trained time series forecasters degrade under non-stationarity and distribution shifts at test time, and classic test-time adaptation methods rely on entropy losses and IID assumptions that do not fit regression-based forecasting.

Approach

Adapt a frozen source forecaster at test time by training lightweight input/output gated calibration modules with MSE on partially-observed ground truth whose length is set from the dominant FFT period of the look-back window, then replace delayed predictions with the adapted ones.

Key results

• TAFAS reduces test MSE for iTransformer and DLinear by 4.95% and 5.20% on average at H=336, and by 5.76% and 6.30% at H=720.
• TAFAS improves Chronos forecasting error on ETT datasets by up to 45% (Chronos-small/base/large, H=96).
• On FreTS + RevIN on ETTm1 (H=96), TAFAS improves MSE by 66.39%, from 1.071 to 0.360.

Measured effects reported in the paper

Surprising findings

• Normalization modules pre-trained for non-stationarity (RevIN) can substantially increase test MSE in some settings, e.g., FreTS + RevIN on ETTm1 rising from 0.367 to 1.071 at H=96.
• Adapting the internal modules of the forecaster (norm, encoder, decoder) can degrade performance below baseline, likely from overwriting core semantics.

Limitations

• Gains are smaller in short-horizon settings and sometimes negligible (e.g., Illness at some horizons).
• Adaptation relies on partially-observed ground truth becoming available with delay, so it does not apply where ground truth never arrives.
• TAFAS underperforms the baseline on some cases such as Exchange at H=720 for certain models, and on ETTh1 H=720 vs. FSNet/OneNet in the comparison table.

Experimental setup

Datasets: ETTh1 ETTm1 ETTh2 ETTm2 Exchange Illness Weather

Benchmarks: ETTh1 ETTm1 ETTh2 ETTm2 Exchange Illness Weather

Metrics: MSE MAE

Baselines: iTransformer PatchTST DLinear OLS FreTS MICN RevIN Dish-TS SAN FSNet OneNet

Builds on: TENT RevIN Dish-TS SAN FSNet

Closest papers by meaning

• Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning (ICML 2024 · similarity 0.86)
• TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.86)
• Less is More: Unlocking Specialization of Time Series Foundation Models via Structured Pruning (NeurIPS 2025 · similarity 0.86)
• Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting (NeurIPS 2026 · similarity 0.86)
• Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.85)
• AdaPTS: Adapting Univariate Foundation Models to Probabilistic Multivariate Time Series Forecasting (ICML 2025 · similarity 0.85)