TIMED: Generative Time Series Framework

• TIMED is a generative framework for multivariate time series that integrates diffusion-based generation, autoregressive supervision, and kernel matching to produce realistic synthetic sequences.
• It employs a masked-attention Transformer architecture that enforces both global distribution modeling and causal temporal dynamics for improved sequence coherence.
• Empirical evaluations across datasets like Stocks, ECG, and Energy demonstrate substantially lower predictive and discriminative errors compared to established baselines.

TIMED is a generative framework for multivariate time series that combines diffusion-based generation, autoregressive temporal refinement, adversarial feedback, and kernel-based distribution matching within a single masked-attention architecture (EskandariNasab et al., 23 Sep 2025). It is motivated by settings in which real time series are scarce, noisy, privacy-sensitive, or expensive to collect, and it targets applications such as forecasting, anomaly detection, and downstream classification. Its central premise is that realistic synthetic sequences require modeling both the joint distribution of entire sequences, $p(\mathbf{X}_{1:T})$, and the temporal dependencies $p(X_t \mid \mathbf{X}_{1:t-1})$; TIMED addresses these jointly rather than treating time series generation as a static-data problem (EskandariNasab et al., 23 Sep 2025).

1. Problem formulation and position within time-series generation

TIMED is introduced against three limitations in prior generative modeling for time series. First, classical autoregressive models such as RNNs and transformers are effective at conditional dynamics but are described as weak generators because they suffer from exposure bias, error accumulation, and limited diversity. Second, GAN-based approaches such as TimeGAN, ChronoGAN, and SeriesGAN rely heavily on adversarial training and can be unstable and prone to mode collapse. Third, standard diffusion models such as DiffTime and vanilla DDPM model global distributions well and avoid adversarial instability, but they typically rely on CNN or U-Net backbones designed for spatial images and do not explicitly enforce autoregressive sequence dynamics (EskandariNasab et al., 23 Sep 2025).

TIMED addresses these issues by combining five design choices in one framework: a denoising diffusion probabilistic model as the generation backbone, masked Transformer architectures instead of CNN or U-Net backbones, an autoregressive supervisor trained with teacher forcing, a Wasserstein critic with gradient penalty, and a Maximum Mean Discrepancy loss with an RBF kernel (EskandariNasab et al., 23 Sep 2025). All four learnable components are optimized jointly.

This design places TIMED between purely unconditional distribution matching and strictly causal sequence prediction. A plausible implication is that its contribution is not a single new module but the integration of several established objectives into a time-series-specific training regime in which unconditional fidelity, conditional dynamics, temporal smoothness, and diversity are optimized simultaneously.

2. Core architecture and composite objective

TIMED has four main components: a DDPM module, an autoregressive supervisor $f_{\text{AR}}$, a Wasserstein critic $D_\phi$, and an MMD loss term (EskandariNasab et al., 23 Sep 2025). The DDPM starts from Gaussian noise at diffusion step $T_{\text{diff}}$ and denoises it into an initial synthetic sequence $\hat{\mathbf{x}}_0$. The supervisor then refines $\hat{\mathbf{x}}_0$ into the final synthetic sequence $\hat{\mathbf{x}}_{\text{synth}}$. The critic compares real and synthetic sequences and supplies adversarial feedback, while MMD aligns real and synthetic feature distributions.

The composite training objective is

$$\mathcal{L}_{\text{TIMED}} = \mathcal{L}_{\text{DDPM}} + \lambda_{\text{AR}}\mathcal{L}_{\text{AR}} + \lambda_{\text{MMD}}\mathcal{L}_{\text{MMD}} + \lambda_{\text{W}}\mathcal{L}_{\text{W}}.$$

This decomposition corresponds directly to the three modeling requirements emphasized in the framework: global structure is handled by DDPM and MMD, local temporal dynamics by the autoregressive supervisor, and adversarial realism by the Wasserstein critic (EskandariNasab et al., 23 Sep 2025).

Component	Function	Training signal
DDPM module	Global distribution modeling	$\mathcal{L}_{\text{DDPM}}$
AR supervisor	Next-step temporal refinement	$p(X_t \mid \mathbf{X}_{1:t-1})$0
Wasserstein critic	Realism and smoothness	$p(X_t \mid \mathbf{X}_{1:t-1})$1
MMD term	Feature-space alignment and diversity	$p(X_t \mid \mathbf{X}_{1:t-1})$2

The diffusion component follows the standard forward-reverse DDPM construction. For a multivariate time series $p(X_t \mid \mathbf{X}_{1:t-1})$3, the forward process applies Gaussian corruption with a linear variance schedule:

$p(X_t \mid \mathbf{X}_{1:t-1})$4

Training uses the standard DDPM noise-prediction loss:

$p(X_t \mid \mathbf{X}_{1:t-1})$5

Sampling begins from $p(X_t \mid \mathbf{X}_{1:t-1})$6, iterates the reverse process to obtain $p(X_t \mid \mathbf{X}_{1:t-1})$7, and passes that output to the autoregressive supervisor (EskandariNasab et al., 23 Sep 2025).

3. Autoregressive refinement, causality, and masked attention

The autoregressive supervisor is introduced because diffusion alone does not explicitly enforce time-lagged dependencies such as $p(X_t \mid \mathbf{X}_{1:t-1})$8 depending on $p(X_t \mid \mathbf{X}_{1:t-1})$9 and earlier observations. TIMED therefore adds a strictly causal module trained by next-step prediction with teacher forcing (EskandariNasab et al., 23 Sep 2025). In pretraining, the last $f_{\text{AR}}$0 steps are masked and predicted from ground-truth history:

$f_{\text{AR}}$1

After diffusion sampling, the refinement step is

$f_{\text{AR}}$2

All major modules use Transformer-style masked attention tailored to sequences. For the DDPM module, the initial hidden representation is

$f_{\text{AR}}$3

For the supervisor, positional embeddings encode sequence order and a strictly causal mask enforces autoregression:

$f_{\text{AR}}$4

where $f_{\text{AR}}$5 if $f_{\text{AR}}$6 and $f_{\text{AR}}$7 if $f_{\text{AR}}$8 (EskandariNasab et al., 23 Sep 2025). Transformer blocks include masked self-attention, a feed-forward network

$f_{\text{AR}}$9

residual connections, and LayerNorm.

TIMED’s adversarial component is a Wasserstein critic trained with gradient penalty:

$D_\phi$0

with $D_\phi$1 set to $D_\phi$2. The MMD component uses an RBF kernel,

$D_\phi$3

and computes empirical squared MMD after flattening $D_\phi$4 into $D_\phi$5, $D_\phi$6 (EskandariNasab et al., 23 Sep 2025).

This architecture makes the causal structure explicit in a way that standard image-style diffusion backbones do not. This suggests that TIMED’s main technical claim is not only that Transformers outperform CNN/U-Net-style backbones for these tasks, but that masked attention is the mechanism through which diffusion, supervision, and adversarial refinement can be made sequence-aware.

4. Training stages and implementation characteristics

Training proceeds in three stages: supervised pretraining of the autoregressive supervisor, diffusion pretraining with the denoising loss, and joint optimization of all modules under the full objective (EskandariNasab et al., 23 Sep 2025). The pretraining stage for $D_\phi$7 uses masked suffix prediction with teacher forcing. The diffusion stage uses the standard DDPM mean-squared error objective on noisy sequences and timestep conditioning. The final stage jointly optimizes DDPM, autoregressive, MMD, and Wasserstein objectives.

The implementation described for TIMED uses a 4-layer Transformer with 8 attention heads and hidden dimension 128, 500 diffusion steps linearly scheduled from $D_\phi$8 to $D_\phi$9, Adam optimizer, learning rate $T_{\text{diff}}$0, and pretraining plus joint training over several hundred epochs (EskandariNasab et al., 23 Sep 2025). The DDPM, supervisor, and critic all share the masked attention design pattern.

These implementation choices are technically consistent with the framework’s stated aim of sequence-native generation. A plausible implication is that the computational burden of TIMED is driven jointly by iterative diffusion sampling and Transformer memory cost, rather than by any single auxiliary loss.

5. Empirical evaluation and ablation structure

TIMED is evaluated on four multivariate time series datasets: Stocks, Sines, ECG, and Energy (EskandariNasab et al., 23 Sep 2025).

Dataset	Seq length / features	Domain
Stocks	24 / 6	Google stock indicators
Sines	24 / 4	Synthetic sinusoidal signals
ECG	24 / 3	Synthetic ECG-like signals
Energy	24 / 5	UCI appliances energy data

The baselines are DiffTime, Vanilla DDPM, AVATAR, TimeGAN, and Teacher Forcing (T-Forcing). Evaluation uses three categories of metrics: a discriminative score based on a GRU classifier trained to distinguish real from synthetic sequences, a predictive score based on GRU forecasting MAE from synthetic-to-real transfer, and distributional alignment assessed with t-SNE and PCA visualizations (EskandariNasab et al., 23 Sep 2025).

Quantitatively, TIMED consistently achieves the lowest discriminative and predictive scores across all datasets. Averaged over datasets, the discriminative score is reduced by approximately $T_{\text{diff}}$1 versus DiffTime, $T_{\text{diff}}$2 versus AVATAR, and $T_{\text{diff}}$3 versus TimeGAN, while the predictive score is reduced by approximately $T_{\text{diff}}$4 versus DiffTime, $T_{\text{diff}}$5 versus AVATAR, and $T_{\text{diff}}$6 versus TimeGAN (EskandariNasab et al., 23 Sep 2025). Vanilla DDPM performs worst, which the paper attributes to the mismatch between CNN-based U-Net inductive bias and temporal sequence structure.

Qualitatively, t-SNE and PCA plots show TIMED’s synthetic sequences tightly co-located with real sequences, and sample-level plots across all four datasets match trends, periodicity, and variability of real samples (EskandariNasab et al., 23 Sep 2025). This suggests that the framework improves both embedding-level overlap and trajectory-level plausibility rather than optimizing only one of those criteria.

The ablation study removes the autoregressive supervisor (w/o ASL), MMD, masked attention (w/o MA), and Wasserstein critic (w/o WC). Removing masked attention causes the largest degradation, with discriminative score increasing by about $T_{\text{diff}}$7 on average and predictive score by about $T_{\text{diff}}$8. Removing ASL also causes substantial degradation, with discriminative score increasing by about $T_{\text{diff}}$9 and predictive score by about $\hat{\mathbf{x}}_0$0. Removing MMD yields moderate degradation, and removing the Wasserstein critic has the smallest effect, though still some degradation (EskandariNasab et al., 23 Sep 2025).

These ablations indicate a hierarchy among TIMED’s components: masked attention and autoregressive supervision are the primary contributors, MMD provides additional distributional alignment, and the Wasserstein critic acts as auxiliary refinement rather than the main performance driver.

6. Uses, limitations, and significance

TIMED is presented as suitable for realistic synthetic time series generation in data augmentation, privacy-preserving analytics, rare-event or imbalanced settings, forecasting model training, anomaly detection training, simulated sensor streams, and benchmark creation without exposing sensitive raw data (EskandariNasab et al., 23 Sep 2025). Its intended utility is specifically tied to scenarios where both high-fidelity global distributions and coherent temporal dynamics matter.

Its main limitations are computational and architectural. DDPM inference requires 500 reverse steps, so sampling is slow; the paper explicitly identifies DDIM as a future direction for faster inference. Transformer layers with multiple heads are memory-intensive as sequence length and feature dimension grow. The weighting coefficients $\hat{\mathbf{x}}_0$1, diffusion schedule, number of steps, and Transformer depth and head count require tuning. The current framework focuses on unconditional sequence generation and does not include explicit conditioning on labels or covariates, making conditional generation a natural extension. The paper also points to adaptive attention for long-range dependencies as a future direction (EskandariNasab et al., 23 Sep 2025).

Within diffusion-based time-series synthesis, TIMED matters because it argues that global distribution modeling is insufficient on its own, and that sequence-native generation requires explicit causal refinement and temporal inductive bias (EskandariNasab et al., 23 Sep 2025). More specifically, it shows that replacing image-oriented backbones with masked Transformers, then coupling diffusion with teacher-forced autoregressive refinement and distribution-matching regularizers, yields synthetic sequences that are more realistic and temporally coherent than the state of the art reported in its benchmark suite.