SimDiff: Simpler Yet Better Diffusion Model for Time Series Point Forecasting

Conference: AAAI 2026 arXiv: 2511.19256
Code: Available
Area: Diffusion Models / Time Series Forecasting / Point Forecasting Keywords: Time Series Forecasting, Diffusion Models, Transformer, Normalization Independence, Median-of-Means, End-to-End

TL;DR

This paper proposes SimDiff — the first purely end-to-end diffusion model to achieve state-of-the-art performance on time series point forecasting. A unified Transformer network serves simultaneously as denoiser and predictor. Combined with Normalization Independence for distribution shift handling and a Median-of-Means ensemble strategy that converts probabilistic samples into precise point predictions, SimDiff achieves 1st place on 6 and 2nd place on 3 out of 9 benchmarks.

Background & Motivation

Diffusion models have demonstrated strong potential for probabilistic modeling in time series forecasting, yet they consistently lag behind regression-based methods in point prediction accuracy. Two core contradictions are identified:

Insufficient contextual bias: Distribution drift commonly exists between the historical and future portions of a time series. Pure likelihood-based methods (e.g., TimeGrad, CSDI) cannot track this drift, leading to unstable training and exploding sampling variance — "diversity so high as to be useless."

Diversity–accuracy trade-off: TimeDiff and mr-Diff stabilize training by providing initial trajectories via pre-trained regressors, but this rigidifies the diffusion process and constrains generative flexibility, effectively reducing these models to regression approaches.

Key question: Can a purely end-to-end diffusion model be designed that neither relies on external pre-trained models nor sacrifices point prediction accuracy, and that instead leverages the intrinsic diversity of diffusion models to improve it?

Method

Overall Architecture

SimDiff is a single-stage end-to-end framework with three core components: - Patch-based Transformer denoising network: serves simultaneously as denoiser and predictor - Normalization Independence (N.I.): past and future are normalized independently during training; only past statistics are used at inference - Median-of-Means (MoM) ensemble: robust point predictions are obtained from multiple inference samples via the MoM estimator

Key Design 1: Normalization Independence

Conventional approaches use statistics \((\mu_X, \sigma_X)\) from the historical sequence \(\mathbf{X}\) to normalize both \(\mathbf{X}\) and \(\mathbf{Y}\). Under distribution shift between past and future:

\[\mathbf{Y}_{\text{norm}} = \frac{\mathbf{Y} - \mu_X}{\sigma_X} = \frac{\sigma_Y \cdot \mathbf{Y}_{\text{real\_norm}}}{\sigma_X} + \frac{\mu_Y - \mu_X}{\sigma_X}\]

The second term \((\mu_Y - \mu_X)/\sigma_X\) constitutes an irremovable bias.

N.I. addresses this as follows: - During training: \(\mathbf{X}\) is normalized via learnable affine parameters \((\gamma, \beta)\); \(\mathbf{Y}\) is independently normalized using its own statistics \((\mu_Y, \sigma_Y)\) - During inference: denoising starts from standard Gaussian noise, and the output is denormalized using \((\mu_X, \sigma_X, \gamma, \beta)\)

This eliminates the bias term at training time, aligning the learned denoising target more closely with the standard Gaussian prior. The learnable parameters \((\gamma, \beta)\) enable the model to infer scale and shift changes in the future from the historical sequence.

Key Design 2: Transformer Denoising Network

A carefully designed lightweight Transformer with several key choices: - Patch tokenization: the time series is divided into overlapping patches as tokens to capture local dependencies - RoPE positional encoding: encodes relative positional information, enhancing attention to temporal patterns; ablation studies demonstrate improvements of 8.3% on NorPool and 1.7% on ETTh1 - Channel independence: each channel is processed independently, increasing effective data volume and simplifying learning - No skip connections: unlike U-ViT, skip connections are found to amplify noise and disturb the diffusion distribution in the time series setting

Key Design 3: Median-of-Means Ensemble

Diffusion models naturally produce diverse probabilistic samples, but extreme values are unavoidable. The MoM estimator proceeds as follows: 1. Draw \(n\) samples, partition them into \(K\) groups of \(B = n/K\) each 2. Compute the mean within each group: \(\hat{\mu}_1, \ldots, \hat{\mu}_K\) 3. Take the median of these \(K\) group means 4. Repeat the shuffle-and-group procedure \(R\) times, and average the \(R\) resulting medians

\[\hat{\mu}_{\text{MoM}} = \frac{1}{R}\sum_{r=1}^{R} \text{median}(\hat{\mu}_1^{(r)}, \ldots, \hat{\mu}_K^{(r)})\]

Compared to simple averaging, MoM preserves subtle temporal patterns rather than smoothing out high-frequency details, while being more robust to outliers. Theoretically, MoM provides tighter finite-sample concentration bounds.

Loss & Training

A weighted MAE loss (rather than standard MSE) is used, normalized by the cumulative noise schedule:

\[L(\theta) = \min_\theta \mathbb{E}_{Y^0, \epsilon, k} \left| \frac{Y^0 - Y_\theta(Y^k, k|c)}{\sqrt{1 - \alpha_{\text{cumprod}}[k]}} \right|\]

The denominator \(\sqrt{1 - \alpha_\text{cumprod}[k]}\) ensures that higher-noise timesteps contribute stronger learning signals.

Key Experimental Results

Main Results: Multivariate Point Forecasting MSE (Selected from Table 2)

Method	NorPool	Electricity	Traffic	ETTh1	ETTm1	Avg. Rank
SimDiff	0.534(1)	0.145(1)	0.383(2)	0.394(1)	0.322(1)	1.33
PatchTST	0.547(2)	0.147(2)	0.385(3)	0.405(2)	0.337(3)	3.22
mr-Diff	0.645(4)	0.155(5)	0.474(8)	0.411(5)	0.340(4)	4.00
TimeDiff	0.665(6)	0.193(7)	0.564(10)	0.407(3)	0.336(2)	5.67
TMDM	0.681(8)	0.267(14)	0.513(9)	0.535(13)	0.436(14)	12.00
TimeGrad	1.152(22)	0.736(23)	1.745(24)	0.993(24)	0.874(23)	21.89

SimDiff achieves an average rank of 1.33 among 25 methods, ranking 1st on 6 of 9 datasets.

Ablation Study: Component Contributions (Tables 4 & 5)

Component	ETTh1	Weather	NorPool	Impact
Full SimDiff	0.394	0.299	0.534	—
w/o N.I.	0.405(+2.8%)	0.328(+9.7%)	0.555(+3.9%)	Significant
w/o RoPE	0.401(+1.8%)	0.310(+3.7%)	0.582(+9.0%)	Significant
1 inference (no ensemble)	0.408	0.317	0.548	MoM reduces error by 3.4–6%
Simple average (non-MoM)	0.398	0.305	—	MoM outperforms averaging

Key Findings

1. Overwhelming inference efficiency: SimDiff requires only 0.22–0.46 ms per inference on ETTh1, compared to 67–380 ms for CSDI and 295–2312 ms for TimeGrad — 100–5000× faster
2. Sound distributional modeling: SimDiff also achieves state-of-the-art probabilistic metrics (CRPS: Electricity 0.22, Traffic 0.16), indicating that its point prediction capability stems from high-quality distributional modeling
3. N.I. yields the largest gains on datasets with severe distribution shift (Weather, NorPool), consistent with the design motivation
4. MoM vs. simple averaging: MoM consistently outperforms simple averaging across all datasets, as averaging smooths out high-frequency temporal patterns
5. End-to-end vs. pre-trained conditioning: SimDiff exhibits higher but more meaningful sampling variance — pre-trained conditioning models constrain the exploration space of diffusion

Highlights & Insights

1. First purely end-to-end diffusion model to achieve SOTA on time series point forecasting: challenges the prevailing paradigm that diffusion-based time series forecasting necessarily requires a pre-trained regressor
2. N.I. is simple yet effective: requires only a lightweight affine layer with negligible computational overhead, yet substantially improves adaptation to distribution shift
3. MoM bridges probabilistic and point forecasting: elegantly transforms the probabilistic sampling advantage of diffusion models into precise point predictions with theoretical guarantees
4. Decisive speed advantage: single-pass inference is over 1000× faster than TimeGrad; even with 100 samples plus MoM aggregation, SimDiff remains faster

Limitations & Future Work

1. Multiple inference passes required: although each forward pass is extremely fast, tens to hundreds of samples are typically needed for robust MoM estimation
2. Channel independence assumption: may discard inter-variable correlations, potentially underperforming in strongly coupled multivariate settings
3. Removal of skip connections: while validated for time series, this limits the model's ability to capture multi-scale features
4. Unevaluated on other modalities: the framework is designed for time series; generalization to spatial data, images, and other modalities remains unexplored
5. MoM hyperparameter sensitivity: the optimal choices of group count \(K\) and repetition count \(R\) still require manual tuning

Related Work & Insights

Method	Stage	Pre-training Required	Inference Speed	Point Quality	Probabilistic Quality
TimeGrad	1	✗	Very slow	Poor	Moderate
CSDI	1	✗	Very slow	Poor	Moderate
TimeDiff	2	✓ (DLinear)	Moderate	Good	—
mr-Diff	2	✓ (DLinear)	Moderate	Good	—
TMDM	2	✓ (Autoformer)	Slow	Moderate	Good
SimDiff	1	✗	Very fast	Best	Good

SimDiff is the only diffusion model that achieves SOTA point forecasting performance without relying on any pre-trained model.

Rating

• Novelty: ⭐⭐⭐⭐ (unique combination of end-to-end diffusion point forecasting, N.I., and MoM)
• Technical Contribution: ⭐⭐⭐⭐ (multiple carefully designed components, each supported by ablation studies)
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ (25 baselines, 9 datasets, dual evaluation of probabilistic and point forecasting)
• Writing Quality: ⭐⭐⭐⭐ (problem motivation is clear, though the main text is somewhat redundant)
• Practical Impact: ⭐⭐⭐⭐ (establishes a new paradigm for diffusion models in time series forecasting)
• Overall Recommendation: ⭐⭐⭐⭐ (4/5)

Related Papers

• [AAAI 2026] TSGDiff: Rethinking Synthetic Time Series Generation from a Pure Graph Perspective
• [ICLR 2026] Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting
• [AAAI 2026] Difficulty Controlled Diffusion Model for Synthesizing Effective Training Data
• [AAAI 2026] Self-NPO: Data-Free Diffusion Model Enhancement via Truncated Diffusion Fine-Tuning
• [NeurIPS 2025] A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking