Research map / Time series / Foundation models for time series
AdaPTS: Adapting Univariate Foundation Models to Probabilistic Multivariate 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 AdaPTS, feature-space adapters (linear/dropout/VAE autoencoders, PCA) that let a frozen univariate time series foundation model (MOMENT-small) produce accurate probabilistic multivariate forecasts.
Problem
Univariate time series foundation models cannot directly handle multivariate inputs with inter-feature dependencies or provide calibrated uncertainty for multivariate forecasting.
Approach
Insert an invertible-in-spirit learned encoder/decoder (adapter) around a frozen univariate FM, applied channel-wise, with probabilistic variants (VAE, MC-dropout) to give uncertainty estimates; trained with forecasting loss plus ELBO-style regularization.
Key results
- On Illness (H=24), VAE adapter reduces MOMENT MSE from 2.902 to 2.461 (about 15% improvement).
- On ETTh1 H=96, dropoutLinearAE achieves 0.395 MSE vs 0.411 for MOMENT (about 8% improvement).
- Adapters improve MOMENT in 5 of 8 task settings, match in 2, and degrade in 1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a dropout LinearAE adapter | Moment-small · ETTh1 (H=96) | MSE: 0.411_{\pm 0.012} -> 0.395_{\pm 0.003} | Helped |
| Add a VAE adapter | Moment-small · Illness (H=24) | MSE: 2.902 -> 2.461 | Helped |
| Add an adapter | Moment-small · Weather (H=96) | MSE: 0.177_{\pm 0.010} -> 0.156_{\pm 0.001} | Helped |
| Add a LinearAE adapter | Moment-small · ExchangeRate (H=96) | MSE: 0.130_{\pm 0.011} -> 0.130_{\pm 0.011} | Mixed |
| Add an adapter | Moment-small · ExchangeRate (H=192) | MSE: 0.210_{\pm 0.002} -> 0.304_{\pm 0.005} | Hurt |
| Increase beta regularization (ablation) | VAE adapter · Illness (H=24) | MSE | Helped |
| Increase beta and log sigma^2 (ablation) | VAE adapter · Illness (H=24) | ECE | Helped |
| Keep sigma fixed during training instead of optimizing it (ablation) | VAE adapter · Illness (H=24) | MSE | Helped |
| Add the decoder component (ablation) | LinearAE · ETTh1, Weather, and ExchangeRate | forecasting error | Helped |
| Add the encoder component (ablation) | LinearAE · Weather | forecasting error | Helped |
Surprising findings
- PCA adapter often fails to improve over the no-adapter baseline, and degrades performance on several datasets.
- On ExchangeRate, adapters including VAE substantially degrade performance at longer horizons (VAE MSE 0.607 vs 0.210 baseline at H=192).
Limitations
- Only MOMENT-small is evaluated; generalization to other univariate FMs is untested.
- Calibration degrades at longer horizons, with overconfident forecasts.
- Normalizing-flow adapters were hard to optimize and are left to future work.
Experimental setup
Datasets: ETTh1 Illness ExchangeRate Weather synthetic sinusoid data
Benchmarks: ETTh1 Illness Weather ExchangeRate
Metrics: MSE MAE ECE reliability diagram calibration
Baselines: MOMENT-small without adapter PCA adapter
Builds on: MOMENT RevIN beta-VAE PCA
Closest papers by meaning
- ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning (ICLR 2024 · similarity 0.89)
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.88)
- Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Models (NeurIPS 2026 · similarity 0.88)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.88)
- SEMPO: Lightweight Foundation Models for Time Series Forecasting (NeurIPS 2025 · similarity 0.87)
- Multivariate Probabilistic Time Series Forecasting with Correlated Errors (NeurIPS 2024 · similarity 0.87)
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.