acceptodds
Under review as a conference paper at ICLR 2027

Time-Series Foundation Models for Short-Term Water Demand Forecasting

Abstract

Short-Term Water Demand Forecasting (STWDF) helps water-system operators schedule pumps around volatile electricity prices, but many existing forecasting methods require use-case-specific training data and maintenance effort that limit their practical deployment. We investigate whether pretrained Time-Series Foundation Models (TSFMs) can forecast short-term water demand without task-specific training. We evaluate ten openly available TSFMs in a zero-shot univariate setting on 26 real-world water demand datasets and compare against baselines. Seven of the ten TSFMs achieve lower mean Weighted Mean Absolute Percentage Error (wMAPE) than Seasonal Naive. Chronos-2 has a mean wMAPE of 17.3%, compared with 23.8% for Seasonal Naive, although several leading TSFMs are practically equivalent. Longer contexts and shorter horizons generally improve forecast quality. Weather and calendar covariates yield no consistent improvement in our setup, while fine-tuning can improve performance. In a comparison with 31 use-case-specific STWDF models, two TSFMs rank in the top three. Several pretrained TSFMs provide strong zero-shot baselines for STWDF in the evaluated setting. All code, data, and experimental results are publicly available on GitHub.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Short-Term Water Demand Forecasting, Time-Series Foundation Models, Deep Learning, Water Demand

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nCuLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper evaluates ten publicly available time-series foundation models on 26 water-demand series from five collections. The main setting uses univariate zero-shot forecasting with a 48-hour horizon and statistical baselines. Additional experiments vary context and horizon length, introduce weather/calendar covariates, fine-tune Chronos-2 globally or locally, and compare zero-shot models with published BWDF challenge entries. Several models outperform Seasonal Naïve on average; the analysis also reports practical equivalence among leading models, inference costs, and unresolved pretraining-overlap limitations.

Critical strengths
  1. A useful domain evaluation with an operationally relevant question. The paper evaluates whether pretrained models can provide effective forecasts without maintaining a separately fitted model for every local use case. The breadth of model families, forecasting settings, and water-demand series makes the results practically informative. The reduction in mean wMAPE from 23.8% for Seasonal Naïve to 17.3% for Chronos-2 is meaningful because the seasonal baseline is competitive in this application.

  2. The statistical interpretation and limitations are unusually careful. The analysis distinguishes statistical differences from a predefined region of practical equivalence, considers inference cost, and audits documented pretraining overlap. It explicitly avoids equating zero-shot evaluation with proven absence of pretraining exposure. It also acknowledges that the BWDF comparison uses heterogeneous challenge procedures. These qualifications improve the usefulness of the model-selection conclusions and should be retained.

Critical weaknesses
  1. Essential evaluation and metric details remain insufficiently specified in the manuscript. Section 3.4 states that forecast windows are shared, and fine-tuning uses held-out windows, but does not explicitly give the per-series training/evaluation cutoffs, forecast-origin stride, or window counts. These definitions are essential for interpreting global/local adaptation. Table 13 also needs a precise metric explanation. Appendix C.1 states that each series has a fixed wMAPE denominator, so MAE and wMAPE should scale proportionally within a series when evaluated on the same targets. This holds for the 15-minute SOWEKI, YW, and Mauth series, where the MAE/wMAPE ratio is nearly unchanged between zero-shot and fine-tuned Chronos-2. It does not hold for the hourly BWDF and ReCycle series: for BWDF 01, global fine-tuning changes MAE from 3.77 to 0.56 but wMAPE only from 12.64% to 9.03%, and the ratio differs by roughly 3–6× across these eleven series. A different unit, temporal resolution, or set of evaluation targets in the fine-tuning pipeline could explain this pattern. This is a request to establish the computation, not evidence of leakage or an arithmetic error, but it affects how the reported fine-tuning gains on these series should be read.

  2. The breadth of independent evidence and the resulting general insight are limited. The 26 series come from five collections, with BWDF and SOWEKI contributing 21 of them, so a series-level macro-average gives those two collections most of the weight. The BWDF ranking is a useful case study but cannot replace a controlled comparison with locally trained alternatives, as the paper itself recognizes. The covariate result is also configuration-dependent, and its mechanism is not isolated. Consequently, the current evidence supports strong starting baselines for the sampled setting more clearly than a broader understanding of when foundation models transfer successfully to water demand.

Questions for authors
  1. Protocol clarification: Please give the chronological cutoffs, forecast-origin stride/counts, and exact configuration or code locations defining the zero-shot and fine-tuning windows. How are all local and global adaptation examples separated from evaluation targets?

  2. Metric clarification: What are the exact formulas, aggregation order, units, and temporal resolution used for MAE and wMAPE in Table 13? Why does the MAE/wMAPE ratio change between zero-shot and fine-tuned configurations for the hourly BWDF and ReCycle series but not for the 15-minute series, and do all configurations use identical evaluation targets?

  3. Collection-level robustness: Averaging the collection-level wMAPE values in Table 3 with equal weight gives Timer-S1 22.8%, Chronos-2 23.0%, and Seasonal Naïve 33.6%, so the advantage over the seasonal baseline appears robust while the top ordering changes. Do the Friedman/Conover and Bayesian signed-rank conclusions also hold when collections, rather than individual series, are treated as the units of analysis?

  4. Presentation of the main results: The main text contains no figures; all four figures are in the appendix, and the central evidence is conveyed only through dense tables. In particular, Table 3 squeezes 13 methods × 2 metrics into the page width at a font size far below the body text, making the primary result difficult to read. For an empirical evaluation whose conclusions concern accuracy–cost trade-offs, practical equivalence, and sensitivity to context and horizon, this substantially weakens accessibility. Could the authors restructure the main results, for example, by splitting or transposing Table 3, adding an accuracy-versus-inference-time plot, and summarizing the Friedman/Conover results with a critical-difference diagram in the main text?

2 comments

Thank you very much for your feedback and questions. We aim to keep our responses concise and refer to the revised manuscript for further details. Major changes are marked blue in the text.

CW 1

Your observation about the MAE/wMAPE ratio prompted us to rerun the Chronos-2 comparison. Zero-shot, global, and local forecasts now use identical held-out targets, resolution, and scale. We independently recomputed the errors and confirmed that the ratio agrees across configurations within each series. Section 4.3 and Table 6 report the five-seed results, and Appendix F (Tables 15–16) gives the per-series values.

CW 2

We screened water-demand datasets from the literature for public access, temporal resolution, record length, and data quality. dataset_selection.md documents inclusion and exclusion decisions. BWDF and SOWEKI supply 21 of the 26 series across our five collections (Section 3.3, Appendix A). QfA 3 reports how the results change when each collection receives equal weight.

We chose BWDF because its participants developed 31 methods specifically for the same forecasting task. Against these specialist methods, zero-shot TSFMs were competitive without local training (Section 4.4). This comparison addresses practical performance on BWDF. A systematic comparison of conventional architectures and tuning choices would require a separate study.

Prior water-demand studies guided our choice of weather and calendar inputs. On the two collections with weather data, adding these inputs increased inference cost without consistent accuracy gains. Thus, inputs used by conventional water-demand models did not reliably improve zero-shot forecasts here. Section 5 discusses whether demand history already captures some of this information and identifies this explanation as a hypothesis (see also Section 4.2).

QfA 1

Each series was split chronologically into approximately 70% training, 15% validation, and 15% test data, with cutoffs rounded to local midnight in Europe/Berlin. Local adapters used only their own training partition. The global adapter pooled those partitions, each with its own cutoff. Validation loss determined checkpoint selection. Rolling forecasts could use earlier test observations as context. Test targets were excluded from fitting and checkpoint selection.

With two weeks of context and a 48-hour horizon and stride, all variants shared 1,397 complete windows and 268,224 targets (Section 3.4.1, Appendix F). Under results/chronos2_matched_comparison/, series_protocol.csv records cutoffs and window counts per series, forecast_windows.csv lists the windows, and each seed's paired_metrics.csv records the errors. The shared configuration is defined in build_steps() in src/tsfm_water_demand_forecasting/run_chronos2_matched_comparison.py and documented in chronos2_matched_comparison.md.

QfA 2

For each series, let E be the sum of absolute errors over all scored targets, N their count, and D total observed demand. MAE = E/N in the series' original unit, and wMAPE = 100 × E/D in percent. Thus, identical targets give the same MAE/wMAPE ratio across configurations.

The earlier fine-tuning pipeline divided hourly BWDF and ReCycle values by four when converting to 15-minute steps. This changed their MAE scale. The zero-shot reference also covered a different evaluation period. In the revised comparison, all variants use the same 15-minute targets and repeat each hourly value unchanged across four steps.

Tables 15–16 give the mean and sample standard deviation across seeds for each series and fine-tuning variant. Table 6 first takes the median across series per seed, then reports its mean and sample standard deviation across seeds. Zero-shot is computed once. Section 3.4.2 distinguishes these summaries from the observation-weighted main benchmark.

QfA 3

With equal collection weights, Timer-S1 has the lowest mean wMAPE (22.84%), followed by Chronos-2 (23.02%). Seven TSFMs still outperform Seasonal Naïve (33.59%), and Chronos-2 retains the best mean rank. With the five collections as blocks, the permutation-calibrated Friedman test still indicates differences among the ten TSFMs (p < 10⁻⁵). However, 13 of 45 Conover/Holm decisions and 21 of 45 Bayesian classifications change. All four series-level equivalence findings involving Chronos-2 become inconclusive at the 95% posterior threshold (Section 4.1, Appendix C.2).

QfA 4

We split the original Table 3 into separate MAE and wMAPE tables (Tables 3 and 4), increased the font size, and kept both on the same page. Section 4.1 discusses accuracy, inference time, and practical equivalence. Appendices C.3 and D provide the detailed runtime and context/horizon figures.

Thank you for the response and the revised manuscript.

My protocol and metric questions have been addressed. The matched evaluation clarifies the fine-tuning comparison, and the revised aggregation description is consistent with the reported results. I also appreciate the collection-level analysis, the qualifications added to the main claims, and the improved readability of the split tables.

The corrected fine-tuning comparison reports median-wMAPE reductions of approximately 0.44 pp and 0.62 pp for global and local adaptation. These gains are useful but modest, and do not substantially strengthen the contribution beyond the zero-shot benchmark. Presentation has improved, and I have raised my presentation score accordingly, although the main text still lacks a visual summary of the accuracy–cost trade-off.

However, my main concern about the breadth of evidence and the resulting insight remains. Twenty-one of the 26 series come from two collections, so the evidence for broader generalization is limited. The BWDF comparison is useful, but its heterogeneous procedures do not provide a controlled comparison with locally trained alternatives. Overall, the paper establishes that pretrained TSFMs are useful baselines in the sampled setting, but provides limited additional understanding of when they transfer successfully to water demand.

I therefore maintain my original rating. The revision resolves my evaluation concerns, but in my view the paper would benefit from further development before publication.

Reviewer_4SsaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a systematic empirical evaluation of pretrained Time-Series Foundation Models (TSFMs) for short-term water demand forecasting (STWDF). The authors evaluate ten publicly available TSFMs on 26 real-world water demand time series from five data collections, primarily focusing on zero-shot univariate forecasting. The evaluation compares TSFMs against statistical baselines and further investigates the effects of context length, forecasting horizon, exogenous covariates, fine-tuning, and inference efficiency.

Critical strengths
  1. Comprehensive evaluation of recent TSFMs on a practically relevant domain.

The paper investigates ten pretrained TSFMs with different architectures, model sizes, and capabilities across 26 real-world datasets. The evaluation spans both hourly and 15-minute forecasting scenarios, making it relevant to practical water demand forecasting applications. The inclusion of recent models provides a useful comparison of current TSFM capabilities.

  1. Multi-dimensional experimental analysis beyond forecasting accuracy.

In addition to the primary zero-shot comparison, the authors investigate context length, forecasting horizon, covariate usage, fine-tuning, and inference efficiency. These experiments provide useful evidence about the trade-offs involved in deploying TSFMs

Critical weaknesses
  1. Limited methodological novelty and insufficient insight into the underlying mechanisms.

The paper mainly applies existing pretrained models and standard evaluation procedures to a specific application domain. While the empirical results are useful, the work does not introduce a new methodology or provide substantial explanations for the observed model behaviors.

Questions for authors

One question in my mind is that what is the actual benefit of TSFM pretraining compared with training a lightweight model locally for this specific application. I noticed that the authors didn't include some traditional baselines such as lightgbm. Although that we do observe that pretrained TSFMs generally provide better performance and generalization ability across different domains, but I am still curious what would be the case for this specific application.

1 comment

Thank you very much for your feedback and questions. We aim to keep our responses concise and refer to the revised manuscript for further details. Major changes are marked blue in the text.

CW 1

Our contribution is systematic empirical evidence about the transfer of general-purpose TSFMs to water-demand forecasting. Across ten model families and 26 real-world series, we examine accuracy, context and horizon sensitivity, covariates, adaptation, and inference cost. These findings inform model selection and deployment without local training, while addressing broader questions about how TSFM capabilities translate into useful forecasting performance.

The covariate results illustrate this contribution. Weather and calendar inputs increase inference cost without consistent accuracy gains in our configurations, despite their reported value in task-specific water-demand models. This shows that supporting covariates does not by itself establish their benefit for a particular forecasting task. Our findings motivate examining when these inputs provide information beyond the historical context and when their added complexity is justified. Sections 4.2 and 5 discuss this contrast and distinguish the observed results from possible explanations, including redundancy with patterns already captured in the context.

The adaptation study addresses a second question relevant beyond water-demand forecasting: how much of the benefit of task-specific adaptation can be shared across related time series? Under the matched protocol, one global adapter recovers a substantial portion of the improvement achieved by separate local adapters, consistently across five seeds. This provides evidence for an intermediate approach between unchanged zero-shot deployment and individual adaptation for every series. Section 4.3 quantifies this trade-off under the evaluated configuration, providing a basis for investigating when shared adaptation is sufficient and when local specialization adds value.

We recognize that these experiments do not fully explain the underlying mechanisms. The breadth of the evaluation and the page limit constrain the depth of individual mechanism studies. Nevertheless, we believe that identifying where additional inputs provide limited benefit and where adaptation gains can be shared contributes useful knowledge about general-purpose TSFMs and motivates further investigation beyond this application.

QfA

The practical benefit we investigate is obtaining competitive forecasts without establishing a local training and tuning workflow. This study was conducted within a research project involving water utilities. Participating utilities reported that establishing and maintaining a conventional machine-learning pipeline would be difficult for them to implement. They therefore regarded models requiring no local training and little task-specific configuration as offering a major practical advantage. Pretrained TSFMs address this need by using local historical measurements as forecast context without requiring local parameter training.

For comparison, we considered the established BWDF methods a particularly relevant reference. Their authors developed and refined forecasting approaches specifically for the water-demand task that we address with TSFMs. We therefore chose to build on this task-specific expertise and the published comparison of 31 methods, rather than construct a small selection of conventional models whose results would depend on our particular feature, architecture, and tuning choices. A comprehensive evaluation of conventional forecasting architectures and configurations for water demand would constitute a substantial separate comparative study. Our focus is on assessing the alternative offered by pretrained TSFMs and placing it alongside established approaches.

The BWDF results support the practical value of this alternative: zero-shot Chronos-2 and Time-MoE rank second and third for one-day forecasts (Section 4.4, Table 7). Exploratory internal comparisons with simple configurations of locally trained baselines, including LightGBM and XGBoost, also favored out-of-the-box TSFMs in forecasting accuracy. These preliminary observations informed our study design, while the reported quantitative comparison rests on the statistical baselines and the BWDF results.

Section 5 explains this comparison strategy. Because BWDF submissions use heterogeneous training and preprocessing procedures, the comparison does not isolate the causal effect of pretraining. It nevertheless provides relevant evidence of competitiveness against methods specifically developed for this application.

Reviewer_r5o8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The main goal of this paper is to answer whether pretrained Time-Series Foundation Models (TSFMs) can forecast short-term water demand without task-specific training. To do this investigation, the authors selected 10 open TSFMs to evaluate on 26 curated datasets in zero-shot univariate mode. Finally, the evaluation results demonstrated that pretrained TSFMs provide strong zero-shot baselines for STWDF in the evaluated setting.

Critical strengths
  • The authors choose the latest public model checkpoint with explicit exclusion rules and collect 26 datasets for evaluation. Furthermore, the code and data are available on GitHub. Both operations enhance the reproduction of this paper.

  • This paper provided many ablation studies about context/horizon sensitivity, covariate-aware forecasting, and a fine-tuning study to further investigate the efficiency of TSFMs on water demand forecasting.

Critical weaknesses
  • The evaluation metrics is unstable. The arithmetic mean wMAPE and especially MAE across datasets are dominated by particular series: mean MAE values of ~61 with std of ~225 (Table 1 ARIMA) are driven almost entirely by ReCycle (MAE up to 1180).

  • In 4.3, One model (Chronos-2), one random seed, one fixed LoRA configuration (rank 8, lr 10⁻⁵, 2,500 steps), no hyperparameter sensitivity. The claim that "one global adapter achieves 91% of the median-error reduction of 26 dataset-specific adapters" could easily be seed- or config-dependent. The authors acknowledge this, but the result is still presented as a substantive finding.

  • Lack of temporal robustness or concept-drift analysis, despite non-stationarity being named a key deployment concern.

Questions for authors

Table 3 is visibly garbled in the PDF (misaligned columns, e.g., merged cells like "1180.1 8.5”)

1 comment

Thank you very much for your feedback and questions. We aim to keep our responses concise and refer to the revised manuscript for further details. Major changes are marked blue in the text.

CW 1

We acknowledge that ReCycle's large absolute errors strongly influence the aggregated MAE across methods, and we now explicitly discuss this effect in Section 4.1. Following your suggestion, we evaluated the aggregates with and without ReCycle and report both in Table 12 (Appendix C.1). Section 3.4.2 also clarifies the weighting and standard deviations. Excluding ReCycle leaves the weighted mean wMAPE ranking of all 13 methods unchanged, with seven TSFMs still outperforming Seasonal Naïve. We also checked the remaining accuracy analyses for their dependence on ReCycle. These checks support the paper's original central conclusions, particularly the strong zero-shot performance of several TSFMs. The sensitivity analysis clarifies the influence of this series without fundamentally changing our interpretation of the results.

CW 2

Thank you for this helpful suggestion. We have now evaluated both fine-tuning strategies with five seeds. The average median errors across these seeds remain close to the single-seed results, supporting the consistency of the fine-tuning findings under the chosen configuration. The revised global share of the local median-error reduction primarily reflects the matched zero-shot reference, now evaluated over the same held-out test period, rather than substantial variation between seeds. Section 4.3 and Table 6 report the updated results. We distinguish this evidence of seed stability from robustness to other hyperparameter configurations, which these repetitions do not assess.

CW 3

Non-stationarity and concept drift are discussed in prior forecasting research, both in water-demand studies and in other application domains. This literature motivated us to consider changing operating conditions when discussing the potential and limitations of TSFMs. However, we could not reliably identify and validate specific distribution shifts in our datasets that would support a targeted before-and-after evaluation. We therefore could not directly assess whether pretraining or adaptation improves robustness to such changes. Section 5 explicitly acknowledges this limitation and identifies temporal robustness as an important direction for future work in water-demand forecasting and other domains.

QfA

We recognize that the compact layout made the table difficult to read. Our inspection found minor alignment differences, but no merged numerical entries. We have replaced the original Table 3 with two separately typeset tables for MAE and wMAPE, now Tables 3 and 4, using larger text while preserving the values.

Comments on the paper

1 comment
Area_Chair_zLheOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers for your valuable feedback on this submission. As the discussion deadline is approaching, I would kindly encourage all reviewers to review the authors’ responses and share your follow-up comments, particularly on whether your concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions.

Thank you all for your time and contributions to a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.