acceptodds
Under review as a conference paper at ICLR 2027

Structure at the Top, Noise at the Leaves: Hierarchy-Aware Residual Adaptation of Frozen Time-Series Foundation Models

Abstract

Time-series foundation models (TSFMs) forecast each series on its own and ignore the aggregation structure that ties many real systems together. We ask whether an explicit hierarchy can improve a frozen TSFM without touching its weights. This work continues the hierarchical residential load forecasting study of Sucaldito & Ostia (2025), which found that per-series AutoARIMA and LightGBM models forecast aggregates well but degrade sharply on zero-inflated appliance series, and which called for hybrid, hierarchy-aware models. We introduce such a hybrid, the Hierarchy-Aware Residual Adapter (HAA), a 19k-parameter MLP that sits on top of frozen TimesFM 2.5 forecasts. It predicts leaf-level residuals from structural context: bottom-up category, household and system aggregates, the leaf's share of each, recent temporal statistics, and category identity. Bottom-up aggregation of the corrected leaves keeps every level exactly coherent. On a four-level residential energy hierarchy (Total → 25 households → 105 categories → 218 sub-metered appliances, Pecan Street New York, 15-minute resolution), we evaluate a 24-hour forecast against naïve, seasonal-naïve, Chronos and TimesFM baselines. HAA lowers pooled RMSE relative to frozen TimesFM at every level: by 2.0% (appliance), 1.1% (category), 5.1% (household) and 20.2% (system total; Diebold–Mariano ). A closer analysis gives a more cautious picture. HAA beats TimesFM on only 24%, 30% and 36% of individual appliance, category and household series. Series-bootstrap confidence intervals for the lower-level gains include zero. Almost all of the gain at the total level comes from removing a systematic under-forecast (the bias share of TimesFM's total-level MSE falls from 73% to 47%), while the error variance actually grows. The adapter also produces negative loads in 31% of leaf predictions and worsens MAE and MAPE at the leaves. Projecting leaf predictions onto the non-negative orthant before aggregation raises the total-level RMSE gain to 29.4%. We argue that structural context is most useful as an aggregation-amplified bias corrector. We also release a protocol (rolling origins, household hold-out, hierarchy permutation, cross-city transfer) for testing whether such adapters show real structural generalization.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series foundation models, hierarchical forecasting, forecast reconciliation, parameter-efficient adaptation, residual adapter, frozen backbone, structural generalization, residential energy forecasting, load forecasting, TimesFM, Chronos

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_FNWsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper tackles the problem of turning leaf-level forecasts of a time-series foundation model into aggregated and hierarchy-aware forecasts. In this work the leaves can be appliances, categories or households, and their forecasts are turned into system-level forecasts using a Hierarchy-Aware Residual Adapter (HAA). This is an MLP designed to predict corrections. The main contribution is a parameter-efficient, backbone-agnostic adapter that keeps the forecasts coherent across the levels of the hierarchy.

Critical strengths

None.

Critical weaknesses
  1. The data and model selection is questionable. First the dataset contains too few points for a meaningful evaluation in this context. In addition, the evaluation uses only one day for calibration and one day for testing. This is far too little data. Second, both models used rank low in current benchmarks (TimesFM-2.5 ranks 8th and Chronos 16th on fev-bench). Chronos is outdated and has a direct successor, Chronos-2 (published in 2025). TimesFM-2.5 is outdated and has a direct successor, TimesFM-3 (2026). State-of-the-art TSFMs such as Chronos-2 or TimesFM-3 could capture the underlying structure better without needing any adapter.
  2. Methodological this paper does not meet the requirements of ICLR. It is weak in his mathematically description, weak in the description of the experimental setup and weak in form (Figures formatting and overall structure). As it is this paper is not understandable in its current form. Further, major concern are for the results. Exemplarily, the seasonal naive clearly beats the method overall (see Table 1). The visualization of the forecasted days suggests that the model is not applied correctly, as it does not capture the daily pattern at all. At the very least, I would expect a midday valley and morning and evening load peaks.

The first point alone leads me to recommend rejection, as such a selection allows no conclusions about generalization and therefore shows no benefit of this work. But with the second it is a clear rejection from my side.

Questions for authors
  1. Is the HAA learned level by level, i.e., from appliance to category, then from category to household, and from household to total? Is the forecast form the applienc level (FM+ HAA) afterwards used for the category level.
  2. Experimental Setup: How the appliance level forecast propagate into the next level and is this next level again propagated?
  3. Is the context length sufficient? Current literature argues that more context is needed.
  4. Figure 3: The data shown here seem unrealistic. Seasonal naive with a lag of 96 (one day) at least reproduces the daily pattern. In a typical application, a time-series foundation model should easily beat it, or at least perform comparably. Why does the TimesFM forecast show no daily pattern, and why does the Chronos forecast show none either? This does not seem consistent.
  5. Do you think your findings generalize, given that the evaluation uses a single day, a special one at that (31 October, Halloween), and a single data setting? Evaluate it consecutively maybe in form of a time series cross-fold, trading of context but having a stronger evaluation.
Reviewer_g7caOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Hierarchy-Aware Residual Adapter (HAA), a lightweight adapter to improve forecasts from frozen time-series foundation models by using hierarchical information present in time series. The novelty of HAA compared with the state of the art is that it uses hierarchy as an input to correct leaf-level forecasts, instead of using it only for post-hoc reconciliation. To achieve on an power consumption forecast use case, HAA aggregates TimesFM forecasts at the category, household, and system levels, combines them with temporal and identity features, and uses a small MLP to predict leaf-level residual corrections. Evaluation on one residential electricity dataset with 25 households demonstrates small and uncertain improvements at lower levels and a larger system-level RMSE reduction mainly caused by bias correction. However, HAA often loses to the original models on individual series and does not consistently outperform simple baselines.

Critical strengths
  • The paper is well written and clearly presents its method and limitations.
  • HAA is lightweight and preserves hierarchical coherence by construction.
  • The detailed analysis explains that the main aggregate improvement comes from bias correction.
Critical weaknesses
  • The practical relevance of the problem in the selected use case is not sufficiently established.
  • The study is limited to electricity-consumption forecasting.
  • The improvements are modest and inconsistent. Simple/vanilla models sometimes perform better.
  • The evaluation uses only one dataset, one city, one season, and one test day.
Questions for authors

I would be good to motivate the relevance of the problem in the studied use case. Specifically, power grid operators typically do not care about the leaf nodes at low voltage but are rather interested in the impact of power consumption on the grid itself (ie cables, substations, etc.). Could the authors provide a concrete example of why improving leaf-node prediction is important in energy use cases?

How general the problem is (i.e. in what other fields the proposed method would bring practical benefits)?

Reviewer_wnxFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents the numerical performance of adding an adapter to TimesFM. The adapter is developed to leverage the hierarchical aggregations of residuals based on the data structure specifically considered in this paper. The numerical results are presented by comparing the proposed adapter to the baseline foundation models, including TimeFM and Chronos. No clear improvement is evident.

Critical strengths

The main strength of this work might be the design of new adapter on top of the existing foundation models by leveraging the geographic structure arising from the energy consumption data.

Critical weaknesses

The weaknesses are clearly described in the first paragraph of Section 8. In addition to that, the proposed adapter does not show significant improvement in error metrics or intuitions leading to more research.

Questions for authors
  1. How could the proposed approach be extended to other time series data?
  2. What is the intuition of the proposed adapter design?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.