acceptodds
Withdrawn from ICLR 2027

Fracast-0: Fractal Weight Sharing for a Time Series Foundation Model with Only 85K Parameters

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time series foundation models must preserve multi-domain breadth, probabilistic output, and multiple temporal scales, but parameter count grows when each scale receives a separate representation. We introduce FracCast-0, a probabilistic forecasting foundation model that exploits temporal self-similarity to reuse one operator across scales. A parameter-free detector extracts significant seasonal structure. The encoder applies a shared local block along a geometric dilation ladder with scale conditioning, while the decoder combines context-gathered states with an explicit seasonal future state and reuses a second block along another ladder before emitting nine quantiles. Pretraining across six corpora preserves multi-domain breadth within 85,001 parameters. On 97 GIFT-Eval configurations without per-dataset fine-tuning, FracCast-0 is the smallest of 28 evaluated checkpoints and remains non-dominated in the aggregate parameter-accuracy plane with MASE and WQL . It uses 42.0% fewer parameters than TinyCast, whose MASE and WQL are 4.2% and 3.3% lower. These results support cross-scale weight reuse as a practical route to further time series foundation model compression. The official repository is released at [https://anonymous.4open.science/r/fracast-0](https://anonymous.4open.science/r/fracast-0).

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, foundation models, parameter sharing, fractal architecture

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

This paper introduces Fracast-0, a compact probabilistic time-series forecasting foundation model. With a parameter-free seasonal-structure detector, it exploits temporal self-similarity by reusing shared encoder and decoder operators across multiple temporal scales, so the parameter count does not grow with the number of scales. In particular, it uses 85K parameters across 97 GIFT-Eval configurations spanning 23 datasets. Compared with TinyCast, a light-weight foundation model baseline, it uses 42% fewer parameters. With this lightweight backbone, the paper also reports lower CPU latency and resident memory.

Critical strengths

Strength 1. The paper shows that sharing encoder and decoder operators across temporal scales, together with a parameter-free seasonal-structure detector, can substantially reduce model size while retaining competitive forecasting performance. This result demonstrates that cross-scale weight sharing is a viable design choice for compact time-series foundation models.

Strength 2. The paper also reports practical inference-efficiency gains with lower CPU latency and resident memory. This supports the potential usefulness of Fracast-0 in resource-constrained settings.

Critical weaknesses

Weakness 1. Limited methodological contribution. I understand the paper’s goal of building a highly parameter-efficient time-series foundation model. However, beyond sharing encoder and decoder blocks across scales, I find the architectural and training contributions limited. Moreover, while this design reduces parameter count, its forecasting performance trails stronger foundation model baselines. In particular, its performance gap relative to TinyCast widens on medium and long horizons (Section 4.1). The paper also lacks a controlled comparison that isolates the effect of weight sharing. These points make it difficult for me to assess the proposed design as a substantial advance, and they are the main reasons for my weak reject recommendation.

Weakness 2. Limited evidence for inference efficiency. In Figure 6, the reported latency and memory gains come from a single local, batch-1 CPU benchmark. Fracast-0 is faster than TinyCast in this setting, but it is not the fastest model. Without corresponding GPU measurements, it remains unclear how broadly the reported inference advantage applies.

Questions for authors

Question 1. Could you repeat the Figure 6 inference comparison on a GPU under matched settings, reporting latency and memory for Fracast-0 and the relevant baselines?

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

The paper propose a tiny foundation model for time series forecasting. It achieves that through reusing the network modules across different input scales. This means the model will have a smaller footprint in terms of the learnable parameters. The results show that this puts the model on the Pareto front, being smaller than other models and worse than many on the traditional benchmark.

Critical strengths

The paper is sound and I have no large complaints about the content.

The method is solid and makes sense.

The results are good and the model is positioned exactly as expected.

Critical weaknesses

The contribution is minimal.

The results are not that good, but fall exactly where we would expect.

The contribution of adding a point on the Pareto front is not very interesting and the method to do so is also not that novel or interesting.

This paper would be better positioned if they could show this approach can be used on SOTA models without dropping any performance.

Another thing missing is the isolating the weight sharing mechanism in this model only.

Finally, why not try larger versions and plot the whole pareto front and show that this particular architecture can dominate on much larger parameter sizes?

Questions for authors

See weaknesses.

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

This paper proposes Fracast-0, a time-series foundation model designed to reduce parameter count through weight sharing across different temporal scales. The architecture uses dilated convolutions, scale conditioning, and seasonality-based preprocessing to generate probabilistic forecasts. The model is pretrained on multiple datasets and evaluated on GIFT-Eval without dataset-specific fine-tuning. The paper examines the trade-off between model size, forecasting accuracy, and inference efficiency through comparisons with existing time-series foundation models. The results show reduced parameter count and inference latency, but lower forecasting accuracy compared to TinyCast, particularly for longer horizons.

Critical strengths
  • The paper explores cross-scale weight sharing to reduce model size, reusing encoder and decoder blocks across multiple dilations while keeping the parameter count low.
  • The model is evaluated across multiple domains and forecasting horizons on GIFT-Eval, providing a broad assessment of its zero-shot forecasting performance.
  • The paper also reports inference latency and memory usage, which are relevant when evaluating lightweight forecasting models.
Critical weaknesses
  • There are no ablations comparing shared and independent convolutional blocks under similar parameter budgets. The central contribution should have an ablation showing the tradeoff or validating the choice.
  • Several architectural choices, including scale conditioning, the periodic prior, and the seasonal decoder, are introduced without sufficient empirical justification. This makes it difficult to understand which components contribute to the final performance.
  • The impact of repeated autoregressive forecasting on this degradation is not sufficiently investigated. (Ref: Table 2/Figure 5 show worse performance on medium and long)
  • The experimental analysis lacks results across multiple independently trained checkpoints, making it difficult to assess the stability of the reported performance.
Questions for authors
  1. Can the authors compare shared and independent convolutional blocks, as well as a smaller TinyCast variant, under matched parameter budgets to establish the benefit of cross-scale weight sharing?
  2. What is the individual contribution of FiLM scale conditioning, the periodic prior, and the seasonal future-state decoder? Can the authors provide ablations for these components?
  3. Can the authors report results across multiple independent training seeds, including mean and standard deviation, to assess the stability of the reported performance?

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

Related papers

Discussion (0)

Sign in to comment.