- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper addresses the problem of long-horizon forecasting using Jacobi-style parallel refinement. Instead of vanilla autopregressive forecast, the method treats previous forecast as a base for the next forecast, allowing for the efficient use of KV cache to make the long-horizon forecast faster and more accurate. The method works on frozen TSFMs and does not require additional training.
Critical strengths
-
The 4 TSFM architectures are representative of transformer TSFMs
-
Theoretical results and derivations seem to be correct
-
The approach is training free and is compatible with pretrained transformer-based TSFMs
Critical weaknesses
-
The method is only tested on transformer-based architectures.
-
Limitations of the method are not discussed
-
The benchmarks used for validation are not representative of the foundation model testing accepted in the literature. Please add results at list on GiftEval and fev-bench
-
Rolling ensemble is not novel
-
The Jacobi refinement computation seems to have a precedent - novelty concern
Questions for authors
-
Does the method work for xLSTM or state-space models? If so, could you please present results on Tirex or S4D? If not, this is a clear limitation of the method that should be discussed.
-
Authors position rolling ensemble as a novel contribution: "Ensembling is a common approach to improving forecasting accuracy, typically by combining predictions from different models (...). In contrast, we combine forecasts produced by the same TSFM at different inference steps, averaging the current forecast with earlier predictions aligned to the same future timestamps.". First of all, this is a very common industrial practice, specifically to address forecast revision. Additionally, there are published works discussing the ensembling of forecasts from the same model created at different lags - see [1,2]. Therefore, I do not believe this is novel, and certainly the positioning with respect to the noted references is necessary in the literature review.
-
It looks like the use of Jacobi or Gauss-Seidel fixed point iteration method has already been proposed in [3] to accelerate the auto-regressive prediction problem using parallel computation. I do not see how proposed approach is fundamentally different or novel compared to [3].
-
The method claims training free compatibility with TSFMs, however, only transformer based models are tested. Is this method applicable to other architectures, for example, xLSTM/S4D? If so, can you provide experimental evidence to support this?
-
Many important industrial applications require probabilistic quantile forecasts. Decision making is based on rare events and tail probabilities, more often than not. A practical concern with block-autoregressive forecasting is uncertainty propagation when predicted median values are fed back as inputs. For example, predictive intervals may widen over the first 128-step block, then abruptly narrow at step 129 if the preceding median predictions are treated as observed values rather than uncertain forecasts. This is very suboptimal and practically unusable behavior. How does the proposed method handle this? Can you share anecdotes quantifying this behaviour for a few architectures and proposed vs. vanilla auto-regressive forecasting?
[1] DelSole, T., Trenary, L., & Tippett, M. K. (2017). The weighted-average lagged ensemble. Journal of Advances in Modeling Earth Systems, 9, 2739–2752. https://doi.org/10.1002/2017MS001128
[2] W. Potosnak, M. Wolff, M. Cao, R. Ma, T. Konstantinova, D. Efimov, M.W. Mahoney, B.N. Oreshkin, K.G. Olivares, Forking-Sequences: Statistically and Computationally Efficient Multi-Horizon Forecasting with Reduced Volatility, TMLR, 2026, https://openreview.net/pdf?id=dXdycy7WCX
[3] Y. Song, C. Meng, R. Liao, S. Ermon, Accelerating Feedforward Computation via Parallel Nonlinear Equation Solving, ICML 2021, https://arxiv.org/pdf/2002.03629
6 comments
Response 2
The second form of forecast reuse, rolling ensemble (W4/Q2), builds on the established statistical idea of combining overlapping predictions across successive forecast origins. DelSole et al. (2017) study lagged ensembles and forecast weighting, while Forking-Sequences (Potosnak et al., 2026) shares context encoding across forecast origins and aggregates overlapping predictions. Within TempoDraft's sequential rolling workflow, overlapping forecasts from earlier origins have already been generated during previous requests. TempoDraft aggregates these already available, target-aligned predictions at output time with negligible additional compute and no additional model call, turning cross-request forecast reuse into a direct accuracy gain integrated with accelerated rolling inference.
We agree that the novelty statements in our initial submission were too broad and needed qualification, particularly where they could be read as claiming rolling ensembling itself as new. We have revised several such statements in the manuscript, including the sentence highlighted by the reviewer in Section 4.3, and added the relevant citations in Section 2.3 and Sections 4.2 and 4.3 to distinguish established components from TempoDraft's contribution as an integrated framework for cross-request reuse.
Furthermore, our results extend the understanding of rolling ensembling in two respects. In the Selected variant in Appendix E.3, we explored adaptive Mixture-of-Experts (MoE) routing based on input statistics and learned input representations. This variant provides further gains and shows that the framework supports richer aggregation, while simple averaging already captures most of the benefit. We also show that rolling ensembling remains effective when its members are produced by approximate refinement. TempoDraft achieves nearly the same forecasting accuracy as full autoregressive decoding with rolling ensembling at a lower decoding cost, showing that the benefit of ensembling can be retained under a reduced inference budget.
W1, Q1, and Q4. Only Transformers are tested. Does the method work for xLSTM or state-space models (TiRex and S4D)?
TempoDraft also applies to state-space and recurrent models because it reuses and refines forecast values rather than architecture-specific hidden states. We have added results for S4D [(Gu et al., 2022)] and the xLSTM model TiRex-1 [(Auer et al., 2025)] in Appendix D.7 (Tables 14–16).
For S4D, we train a model for each dataset and keep its weights fixed for both decoders. Across nine full-test datasets, TempoDraft lowers MSE on eight and MAE on all nine, while achieving 1.2× to 1.6× speedups over native AR. It also achieves 1.2× to 1.5× speedups on four GIFT tasks. These comparisons show that forecast reuse remains effective with lightweight recurrent updates.
For TiRex-1, TempoDraft directly uses the original pretrained weights without additional training. It lowers MSE on 11 of the 13 tasks and MAE on nine, reaching up to 4.4× speedup over serial AR. This extends the evidence for TempoDraft beyond Transformer backbones to both state-space and xLSTM architectures.
W2. Limitations are not discussed.
We have revised Appendix G.3 as Limitations and Implications, focusing on three directions motivated by forecast reuse.
TempoDraft reuses the previous forecast but still rebuilds the context state at each request. Sliding the input window changes both patch contents and their causal context, so our current training-free framework does not achieve near-lossless KV-cache reuse across requests. This remains an important opportunity as TSFM contexts grow: TimesFM 2.5supports 16,384 time steps, corresponding to 512 input patches of length 32. Training objectives or architectures designed for incremental context updates could enable reliable cache reuse and substantially reduce prefill time.
TempoDraft performs approximate decoding with a limited number of refinement rounds, rather than reproducing the full autoregressive forecast exactly. A further direction is to design a direct prediction head that combines retained state with a few newly arrived observations to produce exact forecast updates. Such a design would need to preserve equivalence to full decoding while avoiding repeated autoregressive rollout.
Forecast reuse also requires overlapping target windows across adjacent requests. The native GIFT-Eval [(Aksu et al., 2024)] and fev-bench [(Shchur et al., 2025)] protocols use non-overlapping target windows, with some tasks containing only one test window per series; TempoDraft therefore falls back to native decoding in these settings (Appendix D.6). Dense rolling evaluation captures a complementary deployment setting in which long-horizon forecasts are repeatedly updated as new observations arrive. This motivates evaluating both forecast quality and update latency under such continuous use.
Response 3
W3. Benchmarks are not representative, and evaluations on GiftEval and fev-bench should be added.
Thank you for recommending these broader benchmarks. We have added dense rolling evaluations on GIFT-Eval (Aksu et al., 2024) in Appendix D.6. Its native protocol uses non-overlapping forecast windows, so it does not provide the overlap between successive requests that TempoDraft is designed to exploit. In applications that update forecasts whenever new observations arrive, adjacent requests instead share most of their context and target timestamps. We therefore retain the benchmark data and forecast horizons while evaluating consecutive rolling requests with stride one.
This setting also addresses the computational constraints behind sparse benchmark evaluation. fev-bench (Section 3.1) limits the number of rolling windows to balance computation and statistical reliability, while KAIROS (Appendix C.2.2) (Feng et al., 2025) uses stride 96 because of Chronos latency. Dense evaluation can expose behavior missed by sparse sampling. In our GIFT-Eval comparison, changing the stride from to 1 changes the lowest-MASE model on hourly Solar from Time-MoE to TimesFM and reverses the relative ranking of PatchFM and Timer. Reducing rolling inference costs therefore supports both frequent forecast updates and more comprehensive evaluation.
The maximum forecast horizon in fev-bench is (Shchur et al., 2025, Appendix A). We use GIFT-Eval to evaluate longer-horizon rolling forecasts. GIFT-Eval contains 21 long-horizon tasks. Six are drawn from the ETT and Weather datasets already represented in our main experiments, so we evaluate the remaining 15 tasks with PatchFM, Time-MoE, Timer, and TimesFM 2.5. These comprise 13 tasks with and two with . The task names and sampling frequencies are listed in Table 12 (Appendix D.6).
The table below reports the full TempoDraft method with rolling ensembling. Across the 60 model--task pairs, the MSE, MAE, and MASE win rates against native AR are , , and , with corresponding skills of , , and . All four backbones improve aggregate accuracy on all three metrics, with an overall speedup. Win rates give half credit to ties, and positive skill indicates lower error. Appendix D.6 includes the individual results for all 15 tasks.
| Model | MSE win (%) | MSE skill (%) | MAE win (%) | MAE skill (%) | MASE win (%) | MASE skill (%) | Speedup |
|---|---|---|---|---|---|---|---|
| PatchFM | 73.3 | +17.68 | 86.7 | +3.44 | 93.3 | +3.79 | 3.4× |
| Time-MoE | 86.7 | +3.39 | 80.0 | +0.72 | 66.7 | +0.55 | 1.3× |
| Timer | 80.0 | +2.71 | 73.3 | +1.59 | 73.3 | +1.59 | 1.7× |
| TimesFM 2.5 | 80.0 | +0.93 | 73.3 | +0.23 | 86.7 | +0.41 | 1.4× |
| Overall | 80.0 | +6.43 | 78.3 | +1.50 | 80.0 | +1.60 | 1.8× |
Response
We thank the reviewer for a careful, constructive review, for pointing out three relevant references, and for suggesting additional experimental settings to assess generality. We have added these three references to the revised manuscript and bibliography, and clarified their relationship to TempoDraft in Section 2 and Sections 4.2 and 4.3. We have also expanded our evaluation across the suggested models and benchmarks, and hope the following clarification addresses your concerns about novelty and generality.
NOVELTY
TempoDraft's central novelty and contribution lie in identifying and analyzing an underexamined problem that limits inference efficiency in time series foundation models (TSFMs), and in proposing and validating a general way to address it. Novelty resides jointly in this problem insight and in the complete framework developed to resolve it.
Inference cost already constrains both benchmark evaluation and practical use of TSFMs. For instance, fev-bench Section 3.1 limits the number of rolling evaluation windows to balance computational cost and statistical reliability, with some datasets evaluated using only one window, while KAIROS (Feng et al., 2025) Appendix C.2.2 uses a stride of 96 because of Chronos latency. In practical deployments, applications in domains such as finance and energy often need to update forecasts as new observations arrive. Across consecutive rolling requests, adjacent windows may differ by only one observation and their prediction targets overlap heavily, yet window-relative patching shifts patch boundaries and changes the model inputs, defeating ordinary exact-prefix KV-cache reuse. Repeated full computation across rolling windows therefore wastes this substantial overlap.
To address this inefficiency, we investigate reusing information from the previous window, examining both its KV cache and its forecasts. Designing an architecture or training framework explicitly for cross-window reuse is a promising longer-term approach. Because training TSFMs can require substantial computational resources, our training-free approach makes cross-request reuse immediately applicable to existing frozen models. Within this general direction, more elaborate refinement and ensemble designs are certainly possible, yet simple mechanisms already perform well and retain broad applicability across diverse TSFMs.
We first examine cross-request KV-cache reuse as a candidate route. Section 5.5 and Appendix F.2 show why an exact linear patch-embedding update under a shifted window does not directly extend through attention, layer normalization, and nonlinear transformations into exact deep-layer KV updates. In our empirical tests on frozen models, approximate cross-request KV reuse leads to accumulating forecasting errors across rolling steps. These analytical and empirical findings motivate reusing forecasts in our training-free framework, while leaving efficient KV reuse through suitable architectures or training strategies as a promising direction for future work.
We therefore turn to forecast reuse in two complementary forms that accelerate decoding and improve forecast quality across rolling windows. The first form, Jacobi refinement (W5/Q3), initializes parallel iterative decoding from a prior forecast, which also clarifies our relation to Song et al. (2021). Their work studies general feedforward acceleration through parallel nonlinear equation solving, including autoregressive image generation, under a formulation that allows general initialization. We found that their public autoregressive generation implementations use zero initialization. In our rolling forecasting experiments, reaching a normalized suffix-MSE tolerance of 0.01 relative to AR requires 4.87 to 6.76 refinement rounds on average with zero initialization, compared with 1.59 to 2.36 with temporal self-drafts. Zero initialization also produces substantially larger deviations from AR at the same iteration budget. An effective temporal draft therefore requires careful design for reuse across windows. TempoDraft exploits the link between successive forecasting problems by using the previous request's time-aligned forecast as a temporal self-draft. Specifically, we align the stored unaveraged forecast in original units, adapt it to the current window's normalization and patching, and refine it in parallel under the new context. The temporal self-draft therefore provides an informed starting point for adapting previous predictions to the new observations. Supporting this temporal initialization, Appendix F.1 Table 25 shows that with the same two rounds, temporal drafts reduce the suffix discrepancy from full autoregressive decoding by 90.9% to 96.8% versus zero initialization on four datasets.