Research map / Time series / Foundation models for time series
TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data
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 TimePFN, the first multivariate time-series Prior-data Fitted Network, pretrained on LMC-Synth synthetic multivariate data (Gaussian-process kernel synthesis plus linear coregionalization) with a convolution-plus-channel-mixing patch transformer, for zero- and few-shot MTS forecasting.
Problem
Multivariate time-series forecasting when domain-specific training data is scarce, requiring strong zero-shot and few-shot generalization; existing transformer MTS models underperform when data is limited.
Approach
Generate large-scale synthetic multivariate series by combining KernelSynth Gaussian-process latent functions with a linear model of coregionalization (LMC-Synth) to model inter-channel dependencies, and train a PFN-style transformer with 1D convolutional filtering, overlapping patches, and channel mixing across all patch tokens.
Key results
- In zero-shot MTS forecasting, TimePFN achieves best MSE on ECL (0.315), Weather (0.209), Traffic (1.108), and ETTh1 (0.453) among compared methods.
- With a 50-sample fine-tuning budget, TimePFN outperforms all transformer baselines and DLinear on most datasets, e.g. ECL MSE 0.235 vs 0.278 for iTransformer.
- In univariate zero-shot forecasting, TimePFN-36 has average MSE 0.752 on ECL and best on Traffic/ETTh2 among compared models, with Chronos-small better at short horizons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add convolutional filtering before patching (ablation) | TimePFN · Zero-shot MTS benchmarks | MSE: 0.315 -> 0.653 (ECL) | Helped |
| Use channel-mixed TimePFN architecture instead of PatchTST (ablation) | TimePFN · Zero-shot MTS benchmarks | MSE: 0.315 -> 0.470 (ECL) | Helped |
| Use generative coregionalization rather than independent variates only (ablation) | TimePFN · Zero-shot MTS benchmarks | MSE: 0.315 -> 0.350 (ECL) | Helped |
| Fine-tune TimePFN with a budget of 50 examples rather than zero-shot | TimePFN · ECL | MSE: 0.315 -> 0.235 | Helped |
Surprising findings
- Solar-Energy and Exchange are exceptions where TimePFN is not best; the Exchange dataset is won by naive last-value forecasting in zero-shot.
Limitations
- Fixed input and forecast lengths of 96 due to architecture.
- Fails to anticipate sudden spikes (e.g., Solar-Energy), attributed to the synthetic prior lacking changepoints.
- Univariate results require padding input to 36 steps for ForecastPFN comparison.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 ECL Weather Traffic Solar Energy Exchange
Benchmarks: ECL Weather Traffic Solar Energy Exchange ETTh1 ETTh2 ETTm1 ETTm2
Metrics: MSE MAE
Baselines: iTransformer PatchTST DLinear FEDformer Autoformer Informer ForecastPFN Chronos-small
Builds on: Prior-data Fitted Networks PatchTST KernelSynth Chronos
Closest papers by meaning
- TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting (NeurIPS 2026 · similarity 0.90)
- ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning (ICLR 2024 · similarity 0.89)
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.88)
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.88)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.88)
- CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables (ICML 2024 · similarity 0.88)
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.