acceptodds
Under review as a conference paper at ICLR 2027

TempoDraft: Efficient Long-Horizon Rolling Forecasting

Abstract

Recent time-series foundation models (TSFMs) achieve strong long-horizon forecasting performance, but their deployment for rolling forecasting remains costly. Their autoregressive decoding invokes the model sequentially for each forecast token. In autoregressive LLM generation, new tokens extend an unchanged prefix, allowing its KV cache to be reused across decoding steps. Rolling forecasting does not preserve this structure. Each new observation advances the fixed-length input window by one time step. This changes the input patches and their causal context, preventing direct reuse of the previous KV cache. However, the previous forecast still covers most timestamps in the new horizon. Thus, we introduce TempoDraft, a training-free method that recasts rolling forecasting as a refinement of the previous forecast under new observations. This forecast provides a temporal self-draft for the next inference step. TempoDraft first rebuilds the current context state and computes the first forecast token exactly as in autoregressive decoding. It then shifts the previous forecast to form an aligned draft and refines the remaining forecast tokens in parallel. A rolling forecast ensemble combines current and earlier forecasts of the same timestamps to stabilize predictions and improve accuracy. On Timer, extending the forecast horizon on hourly ETTh1 from approximately one month to one quarter increases the decoder speedup over autoregressive decoding from 2.3× to 6.6×. At long horizons, TempoDraft reduces MSE by 3.3% on ETTh1 and 6.6% on Electricity relative to the autoregressive baselines. Further experiments across nine benchmarks and multiple TSFM backbones support the method's generalizability and broad applicability

Primary area:
learning on time series and dynamical systems
Keywords:
Long-horizon forecasting, rolling forecasting, time series foundation models, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5W4rOpenReview ↗
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

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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 HH 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 H=288H=288 (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 H=720H=720 and two with H=900H=900. 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 80.080.0%, 78.378.3%, and 80.080.0%, with corresponding skills of +6.43+6.43%, +1.50+1.50%, and +1.60+1.60%. All four backbones improve aggregate accuracy on all three metrics, with an overall 1.8×1.8\times 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.

ModelMSE win (%)MSE skill (%)MAE win (%)MAE skill (%)MASE win (%)MASE skill (%)Speedup
PatchFM73.3+17.6886.7+3.4493.3+3.793.4×
Time-MoE86.7+3.3980.0+0.7266.7+0.551.3×
Timer80.0+2.7173.3+1.5973.3+1.591.7×
TimesFM 2.580.0+0.9373.3+0.2386.7+0.411.4×
Overall80.0+6.4378.3+1.5080.0+1.601.8×
Reviewer_Ef5xOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Summary This paper addresses the challenge of efficient inference in basic time-series models and proposes a framework called TempoDraft to improve the inference efficiency and prediction accuracy of TSFM. It has two main features: first, it reuses historical predictions as input to increase parallelism; second, it improves accuracy by averaging multiple historical predictions. The framework has demonstrated effective performance across multiple benchmark tests.

Critical strengths

S1: The paper identifies an intriguing issue: the process of time series prediction is stream-based, yet due to the characteristics of time series data, existing models cannot reuse existing KV caches and predictions like textual LLMs can. The paper partially addresses this issue. S2: The method proposed in the paper, TempoDraft, is simple yet effective, and can be used out-of-the-box without altering the pre-trained weights. S3: TempoDraft significantly improves inference efficiency and prediction accuracy on multiple datasets and TSFM backbones Weaknesses

Critical weaknesses

W1: The improvement in accuracy and efficiency of the paper seems to be attributed to two different modules respectively. As can be seen from Table 2 and Figure 3, the improvement in prediction accuracy mainly stems from the smoothing effect brought by ensemble and Jacobi iteration. Meanwhile, the parallel modification based on the previous round of predictions as a draft is essentially an approximate accelerated decoder that maintains accuracy without compromising on performance. The paper should clarify this point more explicitly. W2: On datasets with a large number of channels, such as Traffic and PEMS, the speedup ratio of TempoDraft has significantly decreased. The computational bottleneck may have shifted to computing power rather than the number of serial calls. The paper lacks an analysis of the impact of the number of channels. W3: In Table 1, TempoDraft shows a decrease in accuracy on the weather dataset. The author suggests that this may be due to the influence of abnormal zero-value data in the weather dataset. I believe further analysis should be conducted on the method's sensitivity to anomalies such as zero values and abrupt changes. W4: The paper focuses on rolling prediction with a step size of 1, but in practical applications, the step size may not be 1, but rather a long prediction made every few time steps. How effective is TempoDraft for rolling prediction with other step sizes?.

Questions for authors

See Weaknesses

2 comments

Thank you for the close reading of Table 2 and Figure 3. We have clarified the role of each component in the manuscript and added the three requested empirical studies.

W1. Accuracy and efficiency come from different modules, and the paper should state this explicitly.

Thank you for pointing out the need to clarify the roles of these two components. Our Jacobi-style refinement converts serial prediction into parallel prediction with negligible loss of forecasting accuracy, as shown by the close agreement between unaveraged TempoDraft and standard AR, and can mitigate autoregressive error accumulation under distribution shift in some TSFMs. Rolling ensembling improves forecasting accuracy for both TempoDraft and AR with negligible additional inference overhead by reusing existing forecasts. We have added this description and analysis to Section 5.3 (Ablation Studies), alongside Table 2.

W2. Speedup drops on many-channel datasets as the bottleneck shifts toward compute, and a channel-count analysis is missing.

We agree that channel count affects speedup. Larger channel batches increase the computation in each parallel refinement pass, reducing the relative savings from fewer sequential model calls. On Traffic, PatchFM's speedup on an RTX 3090 Ti decreases from 3.1×3.1\times at 7 channels to 1.2×1.2\times at 862 channels. At 862 channels, the RTX PRO 6000 Blackwell achieves 1.4×1.4\times under the same refinement setting. These results show that useful acceleration remains achievable for many-channel streams by matching GPU capability to the workload. We have added the channel-scaling results and a short latency analysis to Appendix D.5 (Table 10).

PatchFM speedup over AR on Traffic (H=720H=720) at different channel counts DD. All results use two refinement rounds and eight-forecast averaging. Each ratio uses AR on the same GPU.

GPUD=7D=32D=128D=256D=512D=862
RTX 3090 Ti3.1×1.6×1.3×1.2×1.2×1.2×
RTX PRO 60004.2×2.3×1.5×1.4×1.4×1.4×

W3. Accuracy drops on Weather, and sensitivity to zero values and abrupt changes should be analyzed.

Thank you for suggesting this analysis. Weather contains negative missing-value codes in the carbon dioxide channel and numerical jumps caused by wind-direction wraparound. For example, 359∘359^\circ and 0∘0^\circ are physically only 1∘1^\circ apart, although their scalar representations differ sharply. These irregularities and abrupt changes in observations make it harder to adapt reused forecasts to the latest context.

The differing Weather results are consistent with weaker adaptation of reused forecasts by Timer and Time-MoE than by TimesFM 2.5. The default budget of J=2J=2 refinement rounds can be insufficient to fully adjust historical trajectories to abrupt changes, allowing earlier trends to persist in the output. This helps explain the small accuracy losses on Weather despite the gains on most other datasets.

To assess robustness to common transient perturbations, we evaluate Timer and TimesFM 2.5 on ETTh1, ETTh2, ETTm2, and Weather with single-step spikes, short level shifts, and zero replacement. Only the observed context is perturbed, and future evaluation targets remain clean. The table below gives representative results as the percentage change in MSE of TempoDraft (J=2,M=8J=2, M=8) relative to unaveraged AR. Negative values indicate lower error, and σ\sigma denotes each channel's training standard deviation. Appendix E.5 (Table 24) gives the full results.

ModelDatasetClean3σ, 1 step6σ, 1 step3σ, 8 stepsZero, 8 steps
TimerETTh1−1.5%−1.7%−2.0%−2.2%−1.7%
TimerWeather−0.3%−0.3%−0.7%−0.5%−18.9%
TimesFMETTm2−2.5%−3.0%−3.4%−6.7%−4.1%
TimesFMWeather+0.5%−0.8%−0.9%−4.8%−42.3%

TempoDraft achieves lower MSE than unaveraged AR in all 32 perturbed settings, with larger relative reductions than under the corresponding clean controls. These results support its robustness to common transient perturbations. Rolling ensembling buffers their impact by combining the current forecast with historical forecasts generated before the perturbation.

AuthorsOpenReview ↗

Response 2

W4. How effective is TempoDraft at rolling strides other than 1?

Thank you for suggesting this interesting setting of rolling forecasting with step sizes other than one. At a rolling request stride of Δ\Delta steps, TempoDraft shifts the previous forecast by Δ\Delta steps and refines it in parallel under the updated context. In Appendix E.1 (Table 19 and Figure 14), we evaluate Δ∈{1,2,4,6,8}\Delta \in \{1, 2, 4, 6, 8\} on ETTh2, PEMS04, Electricity, and Traffic using Timer with context length C=672C=672, horizon H=720H=720, J=2J=2 refinement rounds, refresh period R=96R=96, and M=8M=8 lagged averaging against matched AR (M=8M=8) over all valid test requests. Because consecutive contexts shift further apart as Δ\Delta grows, both TempoDraft and matched AR rescale cached forecasts to the mean and standard deviation of the current context before averaging. That rescaling step was omitted from the stride-1 pipeline in the main text, which accounts for the slight numerical difference at Δ=1\Delta=1 compared with Table 2.

DatasetMetricMethodΔ=1Δ=2Δ=4Δ=6Δ=8
ETTh2MSEAR0.3870.3870.3890.3900.392
MSETempoDraft0.3870.3870.3890.3920.393
MAEAR0.4230.4230.4230.4240.425
MAETempoDraft0.4220.4220.4230.4250.426
SpeedupTempoDraft vs. AR2.1×2.1×2.1×2.1×2.1×
PEMS04MSEAR0.1910.1900.1890.1890.189
MSETempoDraft0.1900.1890.1890.1890.190
MAEAR0.2810.2800.2800.2820.284
MAETempoDraft0.2800.2800.2800.2820.284
SpeedupTempoDraft vs. AR1.4×1.4×1.4×1.4×1.4×
ElectricityMSEAR0.2130.2120.2130.2130.215
MSETempoDraft0.2140.2130.2140.2140.216
MAEAR0.2990.2980.2990.2990.301
MAETempoDraft0.3000.2980.2990.3000.302
SpeedupTempoDraft vs. AR1.4×1.4×1.4×1.4×1.4×
TrafficMSEAR0.4440.4420.4430.4440.444
MSETempoDraft0.4450.4430.4440.4440.445
MAEAR0.2950.2930.2930.2930.295
MAETempoDraft0.2950.2930.2930.2940.295
SpeedupTempoDraft vs. AR1.4×1.4×1.4×1.4×1.4×

TempoDraft stays within 0.5% of matched AR at every stride across all four datasets, with maximum absolute differences of 0.0017 in MSE and 0.0008 in MAE, while maintaining speedups of 1.4× to 2.1×. The small accuracy gap relative to matched AR increases slightly as Δ\Delta grows, most noticeably on ETTh2 and Electricity, because more newly observed steps enter the context per request and make the shifted draft slightly less aligned with the updated context.

Reviewer_9ihfOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how to reduce the inference cost of long-horizon rolling forecasting with autoregressive time series foundation models. The consecutive forecast horizons have large overlap, yet conventional inference repeatedly regenerates predictions for the same timestamps. The key insight is to reuse the previous forecast as a temporal draft and refine it under the updated observations. This paper thus proposed a training-free method TempoDraft to recompute the current context and first forecast token, update later tokens through parallel Jacobi-style refinement, and average timestamp-aligned forecasts from successive windows. Experiments demonstrate consistent inference acceleration and mostly preserved or improved forecasting accuracy without additional training.

Critical strengths
  1. This paper aims to addresses a practical source of repeated computation and proposed a simple, training-free inference procedure. The insight is interesting, to reuse previous forecasts avoids a separate draft model and allow existing pretrained backbones to benefit without changing their weights, making the contribution practically useful.
  2. The evaluation is comprehensive, covers nine datasets with different backbones. And the evaluation results show good efficiency improvement.
Critical weaknesses
  1. Reliability across consecutive approximate updates is not sufficiently characterized. The refresh-period comparison reports similar average MSE on one ETTh1 validation stream, but does not show whether MSE/MAE deteriorate as the next refresh approaches.
  2. The evaluation does not sufficiently establish TempoDraft’s accuracy–latency tradeoff relative to competing acceleration strategies. Table 2 compares forecasting accuracy with STRIDE and shift-and-paste but does not report their latency, while the main speedup comparisons use full AR.
Questions for authors
  1. How does TempoDraft compare with STRIDE and shift-and-paste in measured latency as well as forecasting accuracy? Table 2 reports their accuracy, but a joint comparison under the same execution setting would help establish whether TempoDraft provides a preferable accuracy–latency tradeoff.
  2. Does forecasting error increase with the number of rolling updates since the last AR refresh? Appendix E.1 compares average MSE across refresh intervals, but MSE/MAE grouped by position within the refresh cycle would help determine whether aggregate results hide deterioration before a refresh.
1 comment

Thank you for your constructive suggestions. We have added an analysis of whether forecast errors accumulate within a refresh cycle, together with a joint comparison of forecasting accuracy and latency against baseline methods.

W1 / Q2. Does error increase with the number of updates since the last AR refresh?

We conducted the analysis you suggested on the ETTh1 test set using Timer with the parameter settings from the main text (C=672C=672, H=720H=720, J=2J=2, M=8M=8), grouping MSE and MAE by the number of consecutive rolling updates since the last refresh. The table below (Table 18 on page 27 of the revised manuscript) reports the overall averages of TempoDraft and AR under this setting, together with their late-cycle averages. For example, at R=96R=96, the late-cycle averages cover the last 32 updates of each cycle. AR is run independently on the same forecasting tasks as a reference.

Overall MSEOverall MAELate-cycle MSELate-cycle MAE
RTempoDraftARTempoDraftARTempoDraftARTempoDraftAR
40.42640.42570.43870.43820.42650.42570.43880.4383
160.42650.42570.43880.43820.42700.42610.43920.4385
960.42660.42570.43880.43820.42930.42810.43820.4374

Compared with AR, TempoDraft exhibits no noticeable upward trend in forecasting loss. We further visualize the changes in forecasting loss at each update step across the three refresh periods (Figure 13 on page 27 of the revised manuscript). After consecutive updates, the prediction discrepancy of TempoDraft relative to AR does not noticeably increase, and fluctuations in MSE mainly depend on the forecasting difficulty of the corresponding data rather than error accumulation.

These results show no systematic deterioration in forecasting accuracy over consecutive rolling updates in this evaluation.

W2 and Q1. Latency of STRIDE and shift-and-paste alongside accuracy, under the same execution setting.

Comparing latency with AR and shift-and-paste is important for assessing TempoDraft's balance between forecasting accuracy and inference speed. We therefore added the table below to Appendix D.4 (Table 9 on page 22 of the revised manuscript) to compare TempoDraft with AR, STRIDE, and shift-and-paste. Compared with shift-and-paste, TempoDraft lowers MSE on all eight datasets and MAE on seven. Shift-and-paste is faster on some datasets with many channels, but does not adapt quickly to distribution shifts. Compared with STRIDE, TempoDraft requires no separately trained draft model and achieves lower MSE and MAE on seven of eight datasets. On RTX 4090, TempoDraft delivers speedups of 1.1× to 2.5×, compared with 0.8× to 1.5× for STRIDE. STRIDE is even slower than AR on ETTm1, PEMS04, and PEMS08 because, when draft acceptance is low, draft generation and verification cost more than the target-model computation they save. In these cases, STRIDE still requires about 4 to 5 target-model calls and 12 to 13 draft-model calls per forecast. AR requires 8 target-model calls, whereas TempoDraft and shift-and-paste each average about 3. On datasets with many channels, relaxing STRIDE's acceptance criterion can match or exceed TempoDraft's speed, but with substantially higher forecasting errors. Overall, TempoDraft offers a better balance between forecasting accuracy and inference speed than either baseline.

Forecasting accuracy and inference speedup of Timer at H=720. Speedups are relative to AR measured on RTX 4090. Best values are bold.

MethodMetricETTh1ETTh2ETTm1ETTm2ElectricityTrafficPEMS04PEMS08
ARMSE0.4340.3890.5210.3980.2210.4580.1930.309
MAE0.4430.4240.4640.4100.3070.3100.2820.281
Speedup1.0×1.0×1.0×1.0×1.0×1.0×1.0×1.0×
TempoDraftMSE0.4270.3870.5120.3970.2140.4450.1900.307
MAE0.4390.4220.4600.4100.3000.2950.2800.279
Speedup2.5×2.5×2.5×2.5×1.6×1.1×1.2×1.3×
Shift-and-pasteMSE0.4370.3970.5210.4000.2550.4870.2080.317
MAE0.4470.4250.4540.4110.3250.3280.2900.287
Speedup2.5×2.5×2.5×2.5×2.0×2.0×1.9×2.1×
STRIDEMSE0.4550.4170.4960.4070.2210.4580.1930.309
MAE0.4660.4470.4570.4120.3080.3100.2830.281
Speedup1.2×1.3×0.9×1.5×1.4×1.1×0.9×0.8×

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.