Research map / Time series / Foundation models for time series
UniTS: A Unified Multi-Task Time Series Model
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 UniTS, a single shared transformer for forecasting, classification, anomaly detection and imputation over heterogeneous multi-domain time series, using task tokens and masked-reconstruction pre-training, evaluated on 38 datasets.
Problem
Time series models are task-specific and dataset-specific, and existing unified or LLM-based approaches need extensive fine-tuning or task-specific modules; a single model must handle heterogeneous variables/lengths and both generative and predictive tasks.
Approach
Tokenizes time series samples, learnable prompt tokens, and task tokens (GEN for generative, CLS for predictive) into one format; uses a transformer with two-way time and variable self-attention, a DyLinear interpolated-weight FFN, and gating, with shared GEN/CLS towers and unified masked-reconstruction pre-training so the frozen model can be prompted.
Key results
- Single-task UniTS-ST achieves best MSE on 28 of 32 forecasting datasets and best F1 averaged 89.21 on anomaly detection versus 85.26 for TimesNet.
- Multi-task UniTS-SUP reaches forecasting average MSE 0.439 vs 0.466 for iTransformer, and classification average accuracy 81.6 vs 80.9 for TimesNet.
- Few-shot imputation at 25% mask: UniTS-FT improves MSE over iTransformer by 12.4%; prompted UniTS-PMT forecasting MAE 0.376 vs fully supervised 0.381.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniTS-SUP instead of iTransformer | Multi-task forecasting · 20 forecasting tasks | MSE: 0.466 -> .439 | Helped |
| Use UniTS-SUP instead of iTransformer | Multi-task forecasting · 20 forecasting tasks | MAE: 0.394 -> .381 | Helped |
| Use UniTS-SUP instead of TimesNet | Multi-task classification · 18 classification tasks | accuracy: 80.9% -> 81.6% | Helped |
| Use UniTS-ST instead of TimesNet | Single-task classification · 10 classification datasets | accuracy: 73.6 -> 75.0 | Helped |
| Use UniTS-ST instead of TimesNet | Single-task anomaly detection · 5 anomaly detection datasets | F1: 85.26 -> 89.21 | Helped |
| Train task prompt tokens with the pretrained model frozen instead of supervised UniTS training | Multi-task forecasting · 20 forecasting tasks | MAE: 0.381 -> 0.376 | Helped |
| Use direct multi-step inference instead of sliding-window inference | UniTS · Forecasting at extension +384 | MSE: 0.451 -> 0.442 | Helped |
| Use UniTS-PMT instead of iTransformer-FT | Few-shot transfer · Forecasting and classification, 5% data | MSE: 0.598 -> 0.508 | Helped |
| Fine-tune UniTS rather than use PatchTST-FT | Few-shot anomaly detection · 5 datasets | F1: 84.3 -> 86.3 | Helped |
| Use UniTS-ST instead of the strongest baseline | Single-task imputation · 16 imputation datasets | MSE | Helped |
Surprising findings
- Prompted (frozen backbone) UniTS-PMT slightly outperforms the fully supervised multi-task model on forecasting MAE (0.376 vs 0.381).
Limitations
- Backbone is transformer-based; other backbones such as MLP or Mamba are left for future work.
- Zero-shot learning is significantly more challenging and only initially explored for forecasting.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 ECL Traffic Weather Exchange NN5 ILI Solar-Energy UEA classification datasets
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 ECL Traffic Weather Exchange Solar-Energy SMD MSL SMAP
Metrics: MSE MAE accuracy F1 best count
Baselines: iTransformer TimesNet PatchTST Pyraformer Autoformer GPT4TS DLinear Anomaly Transformer
Builds on: PatchTST TimesNet GPT4TS
Closest papers by meaning
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.88)
- Timer: Generative Pre-trained Transformers Are Large Time Series Models (ICML 2024 · similarity 0.87)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.86)
- TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.86)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.86)
- A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (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.