🤖 machine learning

When Denoising Hurts: Rethinking the Terminal Step of Diffusion Time Series Forecasters -- Extended Version

This paper challenges the assumption that full iterative denoising always improves diffusion-based time series forecasts by demonstrating that low-noise refinement can degrade accuracy, leading to a novel label-free global stopping criterion and a specialized training sampler that collectively enhance both inference speed and predictive performance across multiple real-world datasets.

Original authors: Dat Nguyen-Cong, Luong Tran, Tung Kieu

Published 2026-08-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dat Nguyen-Cong, Luong Tran, Tung Kieu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the future, not with a crystal ball, but with a very smart computer that learns by "un-learning" mistakes. This is the world of time series forecasting, a branch of science dedicated to guessing what happens next in a sequence of data, like tomorrow's weather, next month's stock prices, or the traffic flow on your commute. For a long time, scientists have used a clever trick called diffusion models to do this. Think of a diffusion model like a sculptor working with a block of marble that is initially covered in thick, chaotic fog. The sculptor's job is to slowly wipe away the fog, step by step, revealing the perfect statue hidden underneath. In the computer's world, the "fog" is random noise, and the "statue" is the future data pattern. The standard belief was that the more carefully and slowly the sculptor wiped away the fog, the better the final statue would look.

But what if the sculptor keeps wiping the stone even after the statue is already clear? What if, in their eagerness to polish the surface, they start to accidentally chip off the very details they were trying to save? This is the surprising question a team of researchers from FPT Software and Aalborg University decided to investigate. They looked closely at the final steps of this "fog-wiping" process and discovered something counterintuitive: sometimes, doing less work actually gives you a better result. They found that while the early steps of cleaning up the noise are essential for finding the big picture, the very last steps can actually mess things up, causing the prediction to drift away from reality.

The "Over-Cleaning" Problem

The researchers started by watching how these AI models behaved as they tried to predict time series data. They noticed a strange pattern. In the beginning, when the data is very noisy and blurry, the model works wonders. It quickly figures out the big trends, like the general shape of a wave or the seasonal rise and fall of temperatures. This is the "structure recovery" phase. However, as the model gets closer to the end of its job—when the noise is almost gone and the picture is supposed to be crystal clear—it starts to stumble.

The authors suggest that in these final, low-noise stages, the model gets confused. Instead of refining the clear image, it starts trying to predict tiny, random fluctuations that are essentially just static or measurement errors. It's like a musician who has already played the main melody perfectly but keeps trying to add tiny, random notes at the end, which only makes the song sound worse. The researchers call this "statistical drift." They found that by continuing to "denoise" (clean up) the data all the way to the very end, the model often introduces new errors, making the final forecast less accurate than if it had stopped a bit earlier.

The "Stop Early" Solution

To fix this, the team proposed a simple but powerful idea: stop the process early. Instead of forcing the model to go through all 1,000 steps of cleaning (a common setting), they developed a way to detect exactly when the model has done enough. They call this a "label-free global stopping criterion."

Here is how it works without needing to know the answer in advance: The researchers watch the model's predictions as it cleans the data. They look for the moment when the predictions stop getting better and start wobbling or drifting away. Once they spot this "tipping point" on a few practice runs, they set a rule to stop all future predictions right there. It's like a chef tasting a soup and deciding, "Okay, it's perfect now; if I keep cooking it, it will burn." By stopping early, they found they could speed up the process by 20% to 50% while actually getting more accurate results. In their tests across eight real-world datasets, this method reduced the error (MSE) by about 4.3% to 16.4% compared to other top methods, proving that sometimes, less really is more.

Training the Model to Focus on the Right Stuff

The researchers also realized that because the model spends so much time trying to clean up the "fog" in the beginning, it needs to be trained differently. Usually, models are taught to practice cleaning at every single level of noise equally. But since the final, low-noise steps are actually harmful, the team changed the training schedule. They introduced a Bernoulli timestep sampler, which is a fancy way of saying they made the model practice more on the "foggy" parts and less on the "clear" parts.

Imagine a student studying for a test. If they keep reviewing the easy questions they already know perfectly, they waste time. But if they focus their energy on the difficult, confusing parts where they actually need help, they get much better results. The team's new method forces the model to spend more of its training time on the high-noise, high-impact steps where it learns the most important patterns. They kept a little bit of practice on the easy steps just to be safe, but the heavy lifting happened where it mattered most.

The Verdict

The paper concludes that the old idea—that every single step of the diffusion process makes the forecast better—is wrong for time series. In fact, the very last steps can be detrimental. By identifying the exact moment to stop and retraining the model to focus on the most important parts of the process, the researchers created a system that is faster, cheaper to run, and more accurate. Their experiments showed that this approach works consistently across different types of data, from electricity usage to traffic patterns, suggesting that the key to better AI forecasting isn't just doing more work, but knowing exactly when to stop.

Technical Summary: When Denoising Hurts: Rethinking the Terminal Step of Diffusion Time Series Forecasters

1. Problem Statement

Diffusion models have emerged as a powerful tool for time series forecasting, offering a natural framework for modeling uncertainty through iterative denoising. However, the prevailing assumption in existing literature is that the reverse diffusion process is a uniformly beneficial refinement procedure where every step monotonically improves forecast quality.

This paper challenges that assumption. The authors observe that while general temporal structures (trends, seasonality) are recovered during the early-to-mid stages of reverse diffusion (high-to-intermediate noise levels), continued refinement in the low-noise regime often leads to statistical drift. In this late stage, the denoising network focuses on predicting inherent, unpredictable noise rather than meaningful signal, causing residual errors to accumulate and degrading the final forecast. Furthermore, existing methods often allocate substantial model capacity to these low-noise regions during training, despite their limited or even harmful contribution during inference.

2. Methodology

The proposed framework addresses these issues through two primary mechanisms: a label-free early-stopping criterion and a specialized training strategy.

A. Analysis of Reverse Diffusion Regimes
The authors identify two distinct regimes in the reverse diffusion trajectory:

Structure-Recovery Regime: At high noise levels, the model rapidly reconstructs dominant temporal patterns.
Drift Regime: At low noise levels, the intermediate predictions diverge from the target, introducing instability and error accumulation.

B. Global Label-Free Stopping Criterion
To prevent the model from entering the detrimental drift regime, the authors propose an early-stopping mechanism that terminates reverse diffusion before statistical drift emerges.

Mechanism: Instead of relying on ground-truth future observations (which are unavailable at inference), the method estimates a universal stopping point ( $t_\delta$ ) using a small set of pilot trajectories.
Process: It computes the expected clean prediction across multiple sampled trajectories and applies piecewise-linear regression to detect the "breakpoint" where the error trajectory transitions from stable recovery to divergence.
Outcome: Inference is truncated at this estimated $t_\delta$ , returning the corresponding clean prediction. This reduces sampling costs by 20–50% without requiring access to future data.

C. Drift-Aware Bernoulli Timestep Sampling
Recognizing that high-noise regions are more critical for forecasting accuracy, the authors introduce a new training sampling strategy.

Mechanism: A Bernoulli sampler is used to reallocate training probability. With probability $p$ , timesteps are sampled from the high-noise region ( $[t_\delta, T]$ ); with probability $1-p$ , they are sampled from the low-noise region ( $[1, t_\delta)$ ).
Benefit: This emphasizes learning on heavily corrupted inputs (where denoising has the greatest impact) while retaining coverage of the full diffusion trajectory to ensure the model does not lose the ability to handle low-noise inputs entirely.

D. Architecture
The experiments utilize a simple, long-diffusion schedule ( $T=1000$ ) combined with a Transformer-based network (Patch-based Denoising Network). The authors argue this design reduces sensitivity to hyperparameters and complex guidance mechanisms.

3. Key Contributions

Identification of Drift: The paper reveals that low-noise refinement in time series diffusion models can introduce statistical drift, degrading forecast quality rather than improving it.
Label-Free Early Stopping: A global stopping criterion is proposed that detects the optimal termination point based on generated-sample statistics, eliminating the need for ground-truth future data during inference.
Bernoulli Timestep Sampling: A training strategy that prioritizes the high-noise regime while maintaining full trajectory coverage, improving the model's robustness to corruption.
State-of-the-Art Performance: The combined approach achieves superior probabilistic and deterministic forecasting performance across diverse benchmarks while significantly reducing inference time.

4. Experimental Results

The method was evaluated on eight real-world datasets (including ETT, Weather, Electricity, Exchange, and Traffic) against strong baselines such as DLinear, PatchTST, iTransformer, and various diffusion-based forecasters (CSDI, TMDM, CNDiff).

Deterministic Forecasting: The proposed method achieved the best overall Mean Squared Error (MSE) and Mean Absolute Error (MAE) across all benchmarks. On average, it improved MSE by 4.3% and MAE by 5.8% over the strongest baseline (TimeMixer). Notably, it reduced average MSE by 16.4% compared to CNDiff.
Probabilistic Forecasting: While full reverse diffusion sometimes yielded lower Continuous Ranked Probability Score (CRPS) due to increased sample dispersion (drift), the early-stopped method suppressed this drift. It achieved competitive or superior CRPS scores while providing more accurate point forecasts, indicating a better balance between accuracy and distributional calibration.
Efficiency: Early stopping reduced inference time by 13.6% to 56.6% depending on the sampling budget, with negligible overhead for breakpoint estimation.
Ablation Studies:
- Removing early stopping degraded point forecasting performance, confirming the harm of late-stage denoising.
- Bernoulli sampling improved probabilistic metrics (CRPS) by focusing training on informative high-noise regions.
- The method remained robust across different noise schedules (Linear, Quadratic, Cosine, Sigmoid), whereas full sampling exhibited significant instability.

5. Significance and Claims

The paper claims that the effectiveness of diffusion-based time series forecasting does not depend on completing every denoising step. Instead, performance is maximized by allocating computational and learning resources to the regions that contribute most to the forecast structure (high-noise regimes) and avoiding the detrimental refinement of unpredictable noise in the low-noise regime.

The authors emphasize that their findings explain why prior methods often favored narrow diffusion architectures or specific schedules: these were implicit attempts to mitigate the low-noise drift problem. By explicitly addressing this through early stopping and targeted sampling, the proposed method achieves state-of-the-art results with a simpler architecture and significantly lower computational cost.