Weekly Time-Series Paper Digest (Sep 20‑26, 2025)

TIMED: Adversarial and Autoregressive Refinement of Diffusion-Based Time Series Generation

Generating high‑quality synthetic time series is required for tasks such as prediction and anomaly detection, but real data can be scarce, noisy, or expensive. TIMED addresses three modeling needs: (1) capturing the marginal distribution of observations, (2) modeling conditional temporal dependencies, and (3) ensuring temporal smoothness and fidelity.

The framework combines four components that share a mask‑attention backbone designed for sequence modeling:

DDPM core – a denoising diffusion probabilistic model provides a forward‑reverse diffusion process that learns the global structure of the series.

Teacher‑forced autoregressive network – a supervised predictor trained with next‑step teacher forcing learns short‑term autoregressive dependencies.

Wasserstein critic – an adversarial discriminator supplies a Wasserstein loss that penalises temporally incoherent generations, encouraging smoothness.

MMD loss – a maximum‑mean‑discrepancy term aligns the real and synthetic distributions in a learned feature space, improving diversity and sample quality.

All modules are trained jointly, enabling both unconditional and conditional generation. Experiments on several multivariate time‑series benchmarks report that TIMED produces sequences that are more realistic and temporally coherent than existing state‑of‑the‑art generators. (Paper: http://arxiv.org/pdf/2509.19638v1)

IConv: Focusing on Local Variation with Channel Independent Convolution for Multivariate Time Series Forecasting

Real‑world multivariate series often exhibit non‑stationarity: changing trends, irregular seasonality, and residual components that differ across channels. Recent MLP‑based forecasters capture long‑range dependencies efficiently but, because of their linear nature, struggle to model channel‑wise distribution differences, leading to missed local patterns.

IConv resolves this by coupling an MLP for long‑term trend extraction with a novel convolution architecture that processes each channel independently:

Channel‑independent convolution – each channel passes through a convolution with a large kernel, allowing fine‑grained local variation to be modeled without interference from other channels.

Inter‑channel interaction across layers – after the independent convolutions, subsequent layers aggregate information across channels, preserving cross‑channel relationships while keeping computational cost low.

The design yields reduced parameter count compared with standard multi‑channel CNNs and improves the ability to capture diverse local temporal dependencies. Extensive experiments on multiple public multivariate forecasting datasets demonstrate that IConv consistently outperforms prior methods. (Paper: http://arxiv.org/pdf/2509.20783v1)

DS‑Diffusion: Data Style‑Guided Diffusion Model for Time‑Series Generation

Existing diffusion‑based generators for time‑series require retraining the entire model when a new conditioning signal is introduced, and they often exhibit distribution bias between generated and real data, which can propagate bias to downstream tasks. Moreover, the latent diffusion process is difficult to interpret.

DS‑Diffusion introduces two mechanisms to overcome these limitations:

Style‑guidance kernel – a lightweight kernel injects style information (e.g., data source or regime) into the diffusion process, eliminating the need to retrain the full model for each new condition.

Temporal‑hierarchical denoising (THD) – a hierarchy of denoising steps that leverages explicit temporal information reduces the distribution gap between synthetic and real series.

Generated samples carry an explicit style label, making the inference process more transparent. On several public datasets, DS‑Diffusion lowers the prediction score by 5.56 % and the discrimination score by 61.55 % relative to the strongest baseline (ImagenTime), indicating a substantial reduction in distribution bias and improved realism. (Paper: http://arxiv.org/pdf/2509.18584v2)