Research map / Time series / Foundation models for time series
CondTSF: One-line Plugin of Dataset Condensation for 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 CondTSF, a one-line plugin that adapts classification-oriented dataset condensation to time series forecasting by decomposing the test-error bound into value and gradient terms and periodically updating synthetic labels with expert predictions.
Problem
Existing dataset condensation methods are designed for classification, where matching labels suffices; time series forecasting requires all predicted points to match, so these methods degrade when applied to TS-forecasting.
Approach
Theoretically decomposes the TS-forecasting condensation objective (via first-order Taylor expansion of linear expert models) into a gradient term, handled by parameter matching, and a value term, handled by an additive update of synthetic labels toward the full-data expert's predictions every G iterations; this is a plug-in requiring no backpropagation for the update.
Key results
- Plugging CondTSF into MTT reduces MSE by 93.7% on ExchangeRate (0.964 to 0.061) and 73.3% on ETTm2 under the main 48-length setting.
- CondTSF improved every backbone method (DC, MTT, PP, TESLA, FTD, DATM) on all eight datasets in the main single-architecture table, with the smallest gain on DC (e.g., 0.7% MAE reduction on Traffic).
- In cross-architecture tests with DLinear-distilled data, MTT+CondTSF achieves lower error than other baselines on MLP, LSTM and CNN test models across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CondTSF to MTT (ablation) | MTT · ExchangeRate | MAE: 0.778$\pm$0.084 -> 0.195$\pm$0.007 | Helped |
| Add CondTSF to MTT (ablation) | MTT · Weather | MAE: 0.509$\pm$0.065 -> 0.326$\pm$0.009 | Helped |
| Add CondTSF to MTT (ablation) | MTT · Electricity | MAE: 0.747$\pm$0.012 -> 0.391$\pm$0.003 | Helped |
| Add CondTSF to MTT (ablation) | MTT · Traffic | MAE: 0.742$\pm$0.010 -> 0.494$\pm$0.022 | Helped |
| Add CondTSF to MTT (ablation) | MTT · ETTm1 | MAE: 0.653$\pm$0.019 -> 0.491$\pm$0.004 | Helped |
| Add CondTSF to MTT (ablation) | MTT · ETTm2 | MAE: 0.685$\pm$0.022 -> 0.347$\pm$0.028 | Helped |
| Add CondTSF to MTT (ablation) | MTT · ETTh1 | MAE: 0.693$\pm$0.009 -> 0.532$\pm$0.014 | Helped |
| Add CondTSF to MTT (ablation) | MTT · ETTh2 | MAE: 0.719$\pm$0.006 -> 0.329$\pm$0.003 | Helped |
Surprising findings
- Plugging CondTSF into DC yields only small gains, which the authors attribute to DC matching gradients rather than parameters, causing accumulated parameter error.
Limitations
- Theoretical analysis holds only for linear expert models (gradient on input equals parameters); non-linear experts are only empirically supported.
- Gains are smaller at higher condensing ratios and smaller for DC.
Experimental setup
Datasets: ExchangeRate Weather Electricity Traffic ETTm1 ETTm2 ETTh1 ETTh2
Benchmarks: ExchangeRate Weather Electricity Traffic ETTm1 ETTm2 ETTh1 ETTh2
Metrics: MAE MSE
Baselines: Random DC MTT PP TESLA FTD DATM KIP FRePo
Builds on: MTT DC DLinear
Closest papers by meaning
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.86)
- Elucidating the Design Space of Dataset Condensation (NeurIPS 2024 · similarity 0.86)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.85)
- SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.85)
- DDN: Dual-domain Dynamic Normalization for Non-stationary Time Series Forecasting (NeurIPS 2024 · similarity 0.85)
- Embarrassingly Simple Dataset Distillation (ICLR 2024 · similarity 0.85)
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.