acceptodds
Under review as a conference paper at ICLR 2027

What Drives Online Adaptation of Frozen Forecasters? A Controlled Study of Drought-Index Nowcasting

Abstract

Drought estimates depend on both recent weather and earlier dry or wet conditions. An online adapter can use these observations to improve a pretrained forecast, but that improvement does not reveal how much the forecast itself contributes. We study this question through retrospective drought-index nowcasting. Our evaluation compares three ways to use the same local observations: predict the index directly, correct a frozen forecast, or learn how much weight to give that forecast. A separate comparison holds the local features and update rules fixed, so forecast weighting can be examined without changing feature selection. Direct prediction explains much of the gain from fixed residual correction, while learning the forecast's weight improves further in both comparisons. We distinguish decisions made at each monthly step from decisions about where to apply adaptation. Historical errors help prioritize cases under a limited adaptation budget, but validation favors adapting all evaluated cases; reversion rarely changes predictions in the natural replay. These findings identify when a frozen forecast adds value beyond local learning in this retrospective setting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drought forecasting, Time series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gwpnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks what a frozen time-series foundation model contributes when an online adapter can also learn from the same local observations. The task considered is one-step monthy nowcasting of SPEI-GD drought index from different regions. From the results shown, the direct prediction without the forecast does as good as residual correction with the forecast weight fixed at once. A stack that learns this weight does best, also in a fixed design. Further analyses cover when to adapt and a reversion gate. Evaluation is a causal replay with region-clustered bootstrap intervals.

Critical strengths
  • The no-backbone control is the paper’s most useful idea. An online regression with the same inputs matches fixed residual correction (Table 16). With target history alone, RAIL and ORCA are indistinguishable (Table 13). Most of the gain therefore comes from the covariates, and other adapter papers should adopt this control

  • The protocol is careful. Forecasts precede target reveals, choices are frozen on FIT, and intervals are region-clustered

Critical weaknesses

The paper poses several major weaknesses. I list 2 (as required by conference) below

  • The paper does not isolate what the pretrained forecast adds. The stack nests direct prediction (α=0\alpha=0) and residual correction (α=1\alpha=1), so with this much pooled data its win is expected. From what i understand, this is standard regression-based combination. The learned α\alpha is never reported, and no univariate forecast such as AR or ETS is tried in place of the TSFM. The gain may have nothing to do with pretraining. Also, month-tt weather covers the target’s accumulation window, yet 11 of 12 backbones never see it. With that weather delayed by one month, direct prediction is no better than the bare average (see Table 23). The residual and stack forms are never rerun under that delay

  • The target seems to keep a strong annual cycle. The calendar-month mean gets MSE 0.48 at SPEI-30 versus 0.99 for persistence. In the Sahel at SPEI-180, seasonal naive beats persistence, which is odd for an index standardized per calendar period. Part of the benchmark may then be seasonal tracking, with drought thresholds crossed in ordinary dry seasons. The evidence is also narrow: one retrospective product, one horizon, and secondary results that add little. Gates change nothing, the threshold adapts everywhere, and the bound is about 10^4 MSE. This is too narrow for the general claim in the title.

Questions for authors
  • Is the annual cycle present in the released SPEI-GD product, or introduced by your sampling or standardization?

  • What values does the learned forecast weight take across backbones and scales?

  • Does Chronos-2r receive month-tt weather aligned with the target step, or only as a past covariate?

  • Could you rerun the fixed-design comparison (a) with an AR or ETS forecast in place of the TSFM, (b) on calendar anomalies, and (c) with delayed weather?

  • Figure 4c is hard to interpret. What are the (a) and (b) notations? What do they mean?

Reviewer_TZmmOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work introduces RAIL (Residual-Adaptive Intervention Layer), which is a pipeline for correcting time-series foundation models (TSFMs) for drought nowcasting (specifically, the authors use SPEI-GD, a drought index, as the prediction target).

RAIL incorporates a complex operational pipeline consisting of multiple selection and validation steps wrapping recursive least squares (RLS) updates of the linear adapter features:

  1. FIT-inner: A feature selection based on multiple techniques, generating \approx 50 experts.
  2. FIT-late: A grid search over aggregation (e.g., leader vs. top-three) and memory length.
  3. Reversion gate: A fallback mechanism to plain forecast if the adapter performs worse over a trailing window.

The mechanism is exhaustively evaluated over a large number of cases (12 regions x 3 accumulation scales x 12 TSFMs = 432 cases) and with a sensible sequential testing protocol to prevent data leakage.

For the ablation, if the AI backbone is necessary, the authors run different strategies for including a local observer, which uses only the direct observer (features of the experts), combines them in a fixed way with the frozen forecast (fixed), or also allows the weight of the frozen forecast to vary (stacked). These are sensible setups to investigate whether the local features alone can yield the forecast or whether the frozen TSFM drives the performance.

Critical strengths
  1. The authors built a highly robust testing framework with a prequential testing strategy that prevents data leakage. The sheer scale of the evaluation (number of locations and data) gives their results credibility. This includes the performance of the models that they are comparing, as well as the ablations that they run to single out the effect of the TSFMs and the adapters.
  2. The authors engineered a complex pipeline (expert pools, reversion gates, FIT-late grid search) that effectively works as a harness for the TSFMs. It seems to successfully hedge against outlier failures of online linear models. Some of the safety guards, such as the reversion gate, turned out to not be that important, which was highlighted as a valuable negative finding. This could save future researchers engineering effort.
Critical weaknesses

My major concerns are with the scientific and/or engineering contribution of this work.

  1. Regarding the scientific contribution, even the paper title leads with the question, "What Drives Online Adaptation of Frozen Forecasters?" The authors investigate this with the drought nowcast, which, given its deeply physical and spatiotemporal complexity, hardly seems to be a task I would want to tackle with TSFMs. With this in mind, the finding that most of the performance can be achieved by the adapter alone and that combining both in a flexible way is still the best choice does not seem surprising. But the surprising-vs-unsurprising question aside, is the attribution finding something that carries meaning for TSFMs in general?

  2. The additional question is whether answering this scientific question requires this large machinery of selection, validation, and fitting steps, and whether it is really just required as a harness to make TSFMs work robustly in this context. With the scientific contribution being thin, it brings me to the engineering contribution. The RAIL framework seems to be a comprehensive mechanism to find robust adaptation experts and mechanisms for TSFMs, but the work also does not lead with the contribution of this being a way to build performative TSFMs for droughts. It kind of connects back to what I wrote under point 1. It feels like it is a complex patch for an underlying model choice (univariate TSFMs) that is simply not the right tool for this task. This is supported by the table that compares it to the other locally trained baselines, which are already competitive.

Questions for authors

I have two questions regarding my concerns about the broader applicability of the scientific finding or engineering contribution.

  • Can you explain to me why the RAIL pipeline is necessary to answer the central question, "What Drives Online Adaptation of Frozen Forecasters," or whether the findings somehow extend beyond the specific use case of drought forecasting?

  • Can you convince me that this pipeline could somehow lead to an actually competitive drought nowcasting model that can compete with SoTA?

Reviewer_eMa4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework called RAIL (Residual-Adaptive Intervention Layer) and compares three strategies: direct prediction, fixed residual correction, and learning-based prediction weighting. The study investigates the additional predictive value offered by pre-trained forecasting models—beyond what is captured through local online learning—and explores how these predictions can be effectively integrated into the final estimate. Experiments are conducted based on 36 nowcasting scenarios for drought indices, covering 12 regions and three accumulation scales of the SPEI: 30, 90, and 180 days.

Critical strengths
  1. This paper designs a comprehensive set of model variants and systematically evaluates the impact of individual modules on predictive performance.
  2. A stress test—specifically a "degradation" test—is conducted on the "Reversion Gate," with the resulting residual feedback stress summarized and reported.
Critical weaknesses
  1. I think that the selected baselines are not sufficiently recent. There is a lack of representative time-series forecasting methods proposed for 2025–2026.
  2. The code lacks readability, as it contains a significant amount of extraneous material unrelated to the model itself, which hinders the clarity of the model implementation.
  3. There are issues with the writing and formatting of the paper—for instance, there is excessive white space in the layout. Furthermore, the model diagrams do not meet professional standards and appear as though they were generated directly by an AI tool.
Questions for authors
  1. Why does HGB achieve such outstanding predictive performance? Did all the models in the comparison use the same input features and preprocessing procedures?
  2. Compared with DLinear, PatchTST, iTransformer and TimesNet, does HGB use more exogenous meteorological features?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.