A model designed to make probabilistic time-series forecasts has reported sizeable gains over established methods in benchmark tests. In short-term experiments, DecoVAE reported up to 14.96% lower CRPS and 23.30% lower NMAE. In a focused efficiency comparison against K2 VAE, it also reported 30–74% faster training per epoch and 10–35% faster inference. The evaluation covered seven benchmark datasets.
DecoVAE is a dual-stream variational autoencoder, or VAE, built to treat the two main patterns in a time series separately: its longer-term trend and its repeating seasonal behaviour. The study asks whether giving those components different modelling mechanisms can improve probabilistic forecast quality, make the components easier to examine and reduce the cost of running the model.
The scores come from a computational model comparison, not from people, animals or a live deployment. They show how the proposed system performed in the reported benchmark setting; they do not establish that the architecture itself caused the gains or that the same results will appear in every forecasting environment.
The document identifies itself as arXiv:2608.20052v1, dated 20 August 2026. It is therefore a preprint, and the supplied record reports no peer-reviewed journal status.
A forecast built from two signals
The process begins with a moving-average filter that decomposes the input series into trend and seasonal streams. Instead of sending the unchanged series through one general-purpose model, DecoVAE gives each stream its own VAE pathway.
The trend pathway works in the time domain and uses a differential latent regularizer to encourage smoothness. In practical terms, the model penalises abrupt changes in its hidden trend representation, using first- and second-difference terms rather than treating every movement as equally important.
The seasonal pathway works in the Fourier frequency domain, where repeating patterns can be represented through their amplitude and phase. It uses a complex Gaussian VAE, a version of the model designed to preserve those two properties of a frequency-based signal.
Forecasts from the two pathways are then combined. The design is intended to build different assumptions into different parts of the forecast: smooth behaviour for the trend and frequency-based structure for seasonality.
The test set included four ETT variants—ETTh1, ETTh2, ETTm1 and ETTm2—alongside Electricity, Weather and Traffic. The unit of analysis was a benchmark dataset rather than an individual observation or a person.
The data were divided chronologically, with 70% used for training, 10% for validation and 20% for testing. The main comparison results were reported as averages with standard errors from three independent runs that retrained and evaluated the models.
Where the reported gains appear
The paper uses two main measures. CRPS evaluates the quality of a distributional forecast—the model’s representation of possible outcomes—while NMAE measures point-prediction error. Lower values are preferred for both measures, but they capture different aspects of forecasting performance.
For short-term forecasting, DecoVAE had the best reported average ranks in the comparison table: 1.429 for CRPS and 1.143 for NMAE. These ranks summarise the model’s position across the reported benchmark comparisons rather than giving an error value in the original units of any one dataset.
The short-term accuracy figures are reported as maximum reductions: CRPS fell by as much as 14.96% and NMAE by as much as 23.30%. The wording matters. These are the strongest reported differences, not an average improvement claimed across all seven datasets.
The largest reported gaps came in the long-term ETT experiments. Against the second-best methods, the paper reports CRPS improvement of up to 52.68% and NMAE improvement of up to 26.51%.
Those long-term figures are also maxima and are specific to the ETT comparisons. They should not be read as evidence that DecoVAE was uniformly ahead on every dataset, metric or forecasting horizon included in the study.
A smaller footprint, according to a focused comparison
The paper also examines how much computing the model requires. In the focused comparison with K2 VAE, DecoVAE’s reported training time per epoch was 30–74% faster, while inference was 10–35% faster.
The reported resource differences extend beyond elapsed time. Test-memory use was 48–85% lower, and model size was 80–93% smaller than K2 VAE in the comparisons described by the paper.
One extreme example puts the contrast in concrete terms: a DecoVAE model was reported at 5MB, compared with 70MB for K2 VAE. That example illustrates the size gap in one comparison; it is not a universal model-size result for every dataset or baseline.
The efficiency evidence has a narrower base than the accuracy tables. The analysis focuses on selected datasets and baselines, and the supplied text does not provide the exact plotted times or hardware settings needed to judge how the ranges would translate to a hardware-matched training budget.
Ablations offer clues about what matters
The authors varied parts of the architecture to see how sensitive the results were to the model’s design choices. One ablation focused on the strength of the trend regularisation on ETTh1, comparing settings with and without the first-order penalty and also varying the second-order term.
With the first-order regularisation weight set to α1 = 0, the reported baseline scores were 0.33 for CRPS and 0.39 for NMAE. Across the reported moderate range of α1 from 0.5 to 7.0, the paper describes approximate improvements of 9% in CRPS and 8% in NMAE, with values settling near 0.30 and 0.36.
The second-order setting also mattered in that test: α2 values from 1.0 to 5.0 improved performance relative to the unregularised baseline. Because this was an ablation on one dataset and one horizon, it is best read as a sensitivity result rather than a general rule for choosing regularisation strengths.
For the proposed interpretation of the two streams, the paper uses principal component analysis, or PCA, to visualise their embeddings. It reports highly separable trend and seasonal groups: trend embeddings followed an almost distinct path, while seasonal embeddings appeared more clustered.
That separation is exploratory evidence about the model’s internal representations. It does not by itself show that independent users would find the trend and seasonal components understandable or useful, because no user-based interpretability assessment is reported in the supplied analysis.
What remains untested
The study’s evidence is limited to seven benchmark datasets, the stated short- and long-term forecasting settings and selected ablations. The paper does not report a basis for treating those datasets as representative of all time-series forecasting problems.
Three independent retraining and evaluation runs provide a repeatability check for the main tables, which report means and standard errors. The supplied analysis says that no formal significance tests are reported, so the numerical gaps should not be treated as a statistical verdict that one method is definitively superior.
The headline percentage gains are dataset-specific maxima, while the efficiency results come from a focused set of comparisons. Together, those limits mean the paper does not establish uniform accuracy, speed or memory advantages across all named baselines and possible workloads.
The study evaluates forecast quality with CRPS and NMAE, but the supplied analysis identifies calibration—the degree to which forecast ranges match actual uncertainty—as an open question beyond those reported scores. It likewise leaves independent assessment of interpretability unresolved.
Further tests would need to examine additional datasets, domains, horizons and distribution shifts, as well as reproduce the efficiency claims under matched hardware and training budgets. The current results are a benchmark report, not evidence of deployment-level utility.
The preprint states that its code is available through an anonymous GitHub repository. That offers a route for others to inspect or reproduce the reported comparisons, although broader independent evaluation is still needed to establish how widely the findings generalise.
DecoVAE’s central claim is therefore measured rather than sweeping: separating trend and seasonality, and giving each component a tailored latent representation, coincided with strong reported benchmark scores and a smaller computational footprint in the tests described. Whether those gains survive new data, different forecasting conditions and independent replication remains the question for work beyond this preprint.
Paper data and sources
Original title: DecoVAE: a Lightweight Interpretable Trend-Seasonal VAE Framework for Efficient Probabilistic Time Series Forecasting
Authors: Alexander Marusov, Dmitry Anikin, Alexey Zaytsev
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text