Research map / Time series / Foundation models for time series

TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization

NeurIPS 2026 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 0 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 TimeTok, a hierarchical tokenizer plus block-wise VAR transformer that generates time series at any target granularity, unconditionally or from coarse inputs, by controlling the number of token blocks.

Problem

Existing time-series generative models cannot control temporal granularity of conditioning inputs or generated outputs, so they cannot refine coarse sketches or generate at a user-specified level of detail.

Approach

A hierarchical tokenizer with register tokens and FSQ, whose leading tokens encode coarse structure, decoded by a flow-matching decoder trained under varying token budgets; a VAR-style transformer then generates token blocks coarse-to-fine, so the number of blocks sets output granularity.

Key results

• On ECG5000 standard generation, TimeTok reaches FID 0.004 and ACC 0.948 versus 0.020 FID and 0.940 ACC for the best baseline ImagenTime.
• In conditional GC-TSG on ECG5000, TimeTok achieves CRPS_sum 1.587 versus 1.737 for the best baseline TimeGAN, with C-Cons 0.010 vs 0.011.
• Coarse-input classification on ECG5000 rises from 0.314 accuracy (coarse series) to 0.918 accuracy after TimeTok refinement, against an oracle of 0.959.

Surprising findings

• Level detection via reconstruction-error tradeoff gives CRPS_sum 1.587 versus 1.509 with ground-truth levels, a small gap.
• Uniform and exponential token allocations produce nearly identical losses across levels, suggesting fewer tokens suffice for coarser levels.

Limitations

• Foundation-scale tokenization evaluation is described by the authors as preliminary.
• Granularity levels in the single-dataset setting depend on a chosen hierarchy of information-reducing operators (Gaussian smoothing), which is not uniquely invertible.
• C-Cons is only defined for Gaussian smoothing operators.

Experimental setup

Datasets: ECG5000 ItalyPowerDemand Nasdaq ETTh1 UTSD

Benchmarks: ECG5000 ItalyPowerDemand Nasdaq ETTh1

Metrics: FID ACC F1 MSE CRPS_sum C-Cons

Baselines: RNN(TF) RNN(PF) TSGDiff TimeGAN TimeVQVAE DiffusionTS TOTEM ImagenTime

Builds on: VAR TimeVQVAE TOTEM Conditional Flow Matching Diffusion-TS

Closest papers by meaning

• Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.90)
• TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series (NeurIPS 2024 · similarity 0.87)
• Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models (NeurIPS 2026 · similarity 0.87)
• Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.86)
• TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.86)
• MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process (ICLR 2024 · similarity 0.86)