Research map / Time series / Causal and generative time series
FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation
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 FIDE, a conditional diffusion model for time series that inflates high-frequency Fourier components, conditions on block maxima, and adds a GEV negative log-likelihood loss to better preserve extreme values.
Problem
Standard time series diffusion models fail to preserve the distribution of extreme values (block maxima) in generated samples, which matters for risk-critical applications.
Approach
Frequency-domain inflation of the top-κ high-frequency Fourier coefficients before the forward noising process so block maxima (tied to abrupt high-frequency changes) decay more slowly; plus block-maxima conditioning and a GEV likelihood regularizer on the denoised estimate of block maxima.
Key results
- On block-maxima distribution metrics, FIDE achieves the best JS, KL and CRPS on all 5 datasets in Table 1 (e.g., AR1 KL 0.0030 vs 0.0054 for Diffusion-TS).
- On all-values distribution, FIDE is comparable to DDPM and Diffusion-TS rather than uniformly best (e.g., AR1 KL 0.0012 vs DDPM 0.0010).
- Predictive score: FIDE is best on 3 of 5 datasets and second on the remaining 2 according to the paper's claim.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add high-frequency inflation (ablation) | FIDE · AR1, Stock, Energy, Temperature, and ECG block maxima | JS Divergence: 0.0006+0.0001 -> 0.0004±0.0001; 0.0898+0.0054 -> 0.0700±0.0061 | Helped |
| add conditional block maxima input (ablation) | FIDE · AR1, Stock, Energy, Temperature, and ECG block maxima | JS Divergence: 0.0007+0.0001 -> 0.0004±0.0001; 0.1054+0.0089 -> 0.0700±0.0061 | Helped |
| add GEV loss (ablation) | FIDE · AR1, Stock, Energy, Temperature, and ECG block maxima | JS Divergence: 0.0005+0.0001 -> 0.0004±0.0001; 0.0922+0.0056 -> 0.0930±0.0082 | Mixed |
Surprising findings
- FIDE does not beat DDPM and Diffusion-TS on all-values KL and CRPS, indicating a trade-off for extreme-value focus.
- Predictive score remains relatively unchanged when any single component is removed, suggesting redundancy among components.
Limitations
- Theoretical analysis relies on strong assumptions (PSD modeled as generalized Gaussian, stationary noise PSD).
- Gains on all-values metrics are not consistent; the paper notes comparable rather than superior performance.
- Evaluated on only five datasets with a single univariate setting and no stated computational-cost analysis.
Experimental setup
Datasets: AR(2) synthetic Google stock prices 2004-2019 UCI Appliances energy Daily minimum temperature Melbourne ECG5000
Benchmarks: AR1 Stock Energy Temperature ECG
Metrics: JS divergence KL divergence CRPS predictive score (MAE)
Baselines: beta-VAE conditional beta-VAE TimeVAE TimeGAN cGAN RealNVP Fourier-Flows DDPM Diffusion-TS
Builds on: DDPM Diffusion-TS TimeGAN TimeVAE
Closest papers by meaning
- Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting (ICML 2025 · similarity 0.89)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.89)
- E4GEN: Event-level Explainable Extreme-Enhanced Time-series Generation (NeurIPS 2026 · similarity 0.88)
- Population Aware Diffusion for Time Series Generation (AAAI 2025 · similarity 0.88)
- Probabilistic Time Series Modeling with Decomposable Denoising Diffusion Model (ICML 2024 · similarity 0.87)
- ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space (NeurIPS 2026 · 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.