acceptodds
Under review as a conference paper at ICLR 2027

FOMO: Forecasting using Other Markets Observations

Abstract

Time-series foundation models (TSFMs) increasingly accept multivariate inputs, yet adding related series as context can make their forecasts worse, a problem we call context degradation. In financial markets where contracts often have several related ones, this effect can be severe and unpredictable. The same kind of context that degrades one model by up to 15.6% can improve another by 6.8%. We find that the outcome depends on where series interact inside the network, and that early cross-series attention moves the target’s representation away from the state it has without context. We introduce FOMO (Forecasting with Other Markets’ Observations), a training-free algorithm that supplies related market histories as context and restricts cross-series attention to the later layers of a TSFM. Our experiments on four TSFMs and five datasets spanning prediction markets, equities, and crypto show that FOMO never falls below target-only forecasting in any setting where it changes the model, whereas native context mixing significantly degrades forecasts in half of the held-out settings. FOMO also reduces the worst-decile loss in every held-out setting, by a median of 5.3×.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Market Prediction

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

This paper studies context degradation in zero-shot probabilistic forecasting with multivariate time-series foundation models: supplying related time series can worsen a target series' forecast. The authors propose FOMO, a training-free method that retrieves other market histories and masks cross-series attention in the first half of a frozen forecaster, allowing interaction only in later layers. The evaluation covers four foundation models and five datasets from prediction markets, equities, and cryptocurrencies. Results suggest that delaying cross-series interaction can help in some settings, but the gains are uneven and FOMO does not consistently outperform native mixing. Depth ablations, representation analyses, and activation-patching experiments investigate why interaction depth affects forecasting performance.

Critical strengths
  1. Interesting finding. The paper shows that the depth at which cross-series information is introduced can materially affect pretrained time-series foundation models' forecasts. This suggests context degradation depends not only on which related series are supplied, but also on when the model combines them.
  2. Useful empirical analysis. The authors evaluate four foundation models across five datasets and provide depth ablations, representation analyses, and activation-patching experiments. These results offer empirical support for the importance of interaction depth.
Critical weaknesses
  1. The empirical gains are uneven and do not establish FOMO as a consistently better and generalized default. Although delayed interaction improves over target-only forecasting in many evaluated settings, the gains are sometimes small, and FOMO does not outperform other methods in every setting. The results therefore leave open when delayed interaction is beneficial and how reliably the observed gains transfer to other models, datasets, and forecasting tasks. The paper would be stronger with clearer characterization of the settings in which FOMO helps or underperforms.
  2. The contribution is not sufficiently positioned against alternative baselines and prior interaction strategies. The experiments compare target-only forecasting, native mixing, and different interaction-depth schedules, but do not provide a clear comparison with other relevant forecasting baselines or alternative methods for incorporating cross-series information. This makes it difficult to determine how much practical benefit comes specifically from FOMO, beyond the observation that interaction depth matters. The distinction from channel-independent forecasting and staged or adaptive cross-series interaction should be clarified.
Questions for authors

Q1 (Generalization and choice of interaction depth). FOMO uses a fixed half-depth rule, but its gains vary across models and datasets. What evidence supports using the same interaction boundary for unseen architectures or domains? Could a model- or dataset-adaptive choice of interaction depth improve consistency, and how could it be selected without tuning on the test set? Q2 (Comparison with other baselines and prior methods). Could the authors more clearly articulate FOMO's methodological novelty relative to existing multivariate forecasting baselines and alternative approaches to cross-series information integration, such as channel-independent and staged interaction methods? A clearer explanation of these distinctions, together with empirical comparisons against relevant baselines under comparable settings would help establish the practical benefit and novelty of FOMO. Q3 (Sensitivity to neighborhood size). The context-size ablation tests K=1, 3, and 7, with K=7 performing best among the tested choices. Could the authors test additional values, including values above 7 where feasible, to determine whether performance saturates or declines as more neighbors are added? It would also be useful to investigate whether different models benefit from different values of K, rather than using the same fixed neighborhood size across all models.

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

The paper shows 'context degradation' in multivariate time-series foundation models. The authors claim that additional features and cross-attention across all layers can make the forecasting worse. So they study that the outcome depends on where in the network cross-series interaction occurs.
To prove these points, they introduce FOMO, a training-free inference procedure, which is meta-data retrieval. They select related contracts with hand-craft ranking method. In addition, they only allocate cross-series attention on the late, which is a half of the total number of layers, layers.
They conducted experiments on four TSFMs datasets. FOMO improves on target-only forecasting in all 15 model–dataset settings where it changes the model. With the retrieved panel and weights fixed, early-half mixing degrades all three models while late-half mixing improves them.

Critical strengths
  1. The authors provide a clear problem formulation. They clearly identify that adding related features can make forecasting worse. \
  2. They conduct very deep and well-controlled experiments. The retrieved panel and model weights remain unchanged, while only the interaction set SS is varied. This isolates the impact of cross-series interaction placement on predictive accuracy. By assigning the same number of interacting layers to the early-half and late-half schedules, they distinguish the effect of layer position from that of interaction depth. \
  3. They introduce a very simple method. FOMO requires no training or parameter updates, but it only adds a static attention mask and a graph lookup.
Critical weaknesses
  1. Many FOMO gains are not statistically significant. For instance, on Kalshi, only Chronos-2’s improvement over target-only excludes zero. At the per-target level, FOMO still makes 32–54% of Moirai-2’s targets worse in the held-out settings (Table 3). \
  2. Masking cross-series attention in the early layers is a simple inference-time modification. Encoding each series independently before letting series interact, is already built into Toto-4M’s architecture. The contribution is just an empirical observation rather than a new method. \
  3. The retrieval design is hand-crafted without sufficient justification. The retrieval relies on many manually chosen values with no stated rationale or sensitivity analysis. For instance, the Jaccard threshold is merely 0.1, and only 16 stored neighbors.
  4. No principled justification for the ⌈L/2⌉ boundary. The half-depth cut is a fixed heuristic applied to all models. The boundary therefore trades gain against tail risk, and the main text does not discuss how it should be chosen.
  5. Given the random-neighbor result, these choices appear arbitrary, and the complexity of the retrieval pipeline is not justified by its contribution.
Questions for authors
  1. Random admissible neighbors perform similarly to metadata-retrieved neighbors (Tables 2 and 22). What drives FOMO’s gains? Would they persist with unrelated real-world series, such as equity or electricity data for prediction-market targets? \
  2. Table 8 suggests that swapping only neighbor states recovers 77% of the native–FOMO gap in Moirai-2, compared with 64% from swapping target states. Doesn't this indicate that early mixing harms neighbor representations and their interaction with the target, rather than primarily displacing the target representation?
  3. Appendix E states that Figure 7 uses the last-token specification from Tables 9–11, yet the Chronos-2 panel is labeled “pooled” and reports different values (0.34–0.40 vs. 0.12–0.30 in Table 10). Moreover, Table 10 shows higher volatility R² for native mixing at layers 7, 9, and 10. How should these results be interpreted?
  4. Is there a principled way to select the boundary instead of using a fixed ⌈L/2⌉, particularly for new models without a development set?
  5. How were the retrieval hyperparameters selected, including Eq. 7 coefficients, the 0.2 Jaccard threshold, Appendix P.2 weights, and candidate/neighbor caps (512/16)?
Reviewer_UgBuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces "context deterioration" within time series (TS) domain: when target series is complemented with related series as context, the target series' performance becomes worse than its univariate performance. To mitigate the context degradation, a new method (FOMO) is proposed: to encode channels independently/univariately for the first half of the layers, and only then "mix" the channels with each other. Since all the tested TS foundation models are transformer-based, the desired effect is achieved via constraining the attention mask. The algorithmic contribution is demonstrated on the new 5 financial market datasets with 3 foundation models.

Critical strengths

A new approach how to mitigate "context degradation" without retraining or fine-tuning a TS foundation model

Critical weaknesses
  1. there's no comparsion with already existed related literature on "context degradation" and predicting financial markets (e.g. Abdelmalak et al. 2026, Mostafa et al. 2026);
  2. limited comparison with previously used datasets (only electricity in Appendix) and with well-established baselines (e.g. Zeng et al. 2022, Chen et al. 2023);
  3. no clear evidence that the new (I assume, since no literature references are provided) datasets have inter-channel dependency;
  4. the experiments' setup seems to be selected arbitarly (e.g. the set of datasets the conlusion are based on, hyperparameter optimization);
  5. vague claims (e.g. obs. 5 and 6)
  6. the paper is difficult to read (e.g. missing definitions, claim contradictions, unconventional naming);

Abdelmalak, Ibram, Kiran Madhusudhanan, Jungmin Choi, et al. 2026. “Channel Dependence, Limited Lookback Windows, and the Simplicity of Datasets: How Biased Is Time Series Forecasting?” In Advances in Knowledge Discovery and Data Mining - 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2026, Hong Kong, China, June 9-12, 2026, Proceedings, Part II, edited by Raymond Chi-Wing Wong, Hanghang Tong, Hua Lu, et al., vol. 16598. Lecture Notes in Computer Science. Springer. https://doi.org/10.1007/978-981-92-1462-4_46.

Chen, Si-An, Chun-Liang Li, Nate Yoder, Sercan O. Arik, and Tomas Pfister. 2023. “TSMixer: An All-MLP Architecture for Time Series Forecasting.” arXiv:2303.06053. Preprint, arXiv, September 11. https://doi.org/10.48550/arXiv.2303.06053.

Mostafa, Hamza, Om Shastri, and Dennis Lee. 2026. “TimeSeek: Temporal Reliability of Agentic Forecasters.” arXiv:2604.04220. Preprint, arXiv, April 5. https://doi.org/10.48550/arXiv.2604.04220.

Zeng, Ailing, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. “Are Transformers Effective for Time Series Forecasting?” AAAI.

Questions for authors

dear authors,

could you please clarify the following:

  1. what does the symbol ⊬\nvdash in the line 130 and in the equation (7) mean? Is it intended to be \mathbb{1}?
  2. what is the reasoning behind the coefficients' values for each of the terms in equation (7)?
  3. what are the target and features spaces in the datasets?
  4. what is the target: "contract prices" (line 245) or its "volatility" (line 376)? how are these connected?
  5. is there a particular reason for excluding Polymarket datasets (development and later-period) from Table 2?
  6. why is "later-development" missing in the table 3?
  7. Is there a particular reason why table 6 was excluded/separated from the Figure 2?
  8. Figure 2 demostrates the results only on the "development" dataset. I might have missed it but could you also provide/point to the detailed/aggregated results across all datasets? The detailed results I found is in Appendix C and it's only for the "development" dataset.
  9. is there a particular reason for selecting seeds 7, 13 and 29 (line 728)?
  10. the proposed metrics "Harmed" and "Worse-10%" use the "0.1%" and "10%" thresholds respectivelly (lines 173-176). Do the conclusions hold under diffferent thresholding?
  11. Is there a particular reason for K=7 in the line 147?
  12. what does the "target" indexing mean in equation (5)?
  13. what does "target-only quantiles rescaled around the median by one factor" mean in the lines 274-275?
  14. there's an experiment involving ridge regression (line 369). Is there a description how/whether it was tuned? I found only "run_probe_alpha_sweep.py" in the code. Is there a particular reason to set the search grid of weight decay to [10.0, 50.0, 100.0, 500.0, 1000.0]?
  15. what do the three numbers divided by a slash in the "Drift" column in Table 2 mean?
  16. why were the windows subsampled for Table 2? Is there a particular reason to set it to 600 windows per cell for Crypto and Equities while setting it to 555 for Kalshi?

there might be a contradiction in the following:

  1. "Unrestricted early cross-market attention degrades forecasting performance. " (line 311). "Unrestristed early cross-market attention", I assume, is interchanably used for "native mixing". If it's true, then table 1 shows that native mixing also improves the forecasting performance (models Chronos-2 and TimesFM-3, datasets Later-period, Kalshi and Crypto).

  2. "FOMO never falls below target-only forecasting." (line 260). However, there's the claim "On 321 electricity load series (GluonTS), with neighbors chosen by past-change similarity and no shared metadata, native mixing degrades Moirai-2 by 26.3% and FOMO by 18.0%, while the intervals for Chronos-2 and TimesFM-3 include zero under both allocations." (lines 1075-1079). Is this degradation relative to the target-only forecasting?

  3. "Confining cross-series attention to later layers converts degradation into positive gains" (lines 317-318). However, if we compare "middle layer only" vs "first layer only" in the table 6 for Chronos-2, the middle layer only has a negative gain (i.e. performance deterioration) and while the first layer only has a positive gain.

  4. Experiment setup defines the context/look-back window as

    1. T=256 hourly values (line 123)
    2. T=256 days (line 250)
    3. another experiment set it to 72 hours (line 370)
  5. it's claimed that "we build one graph of related markets ahead of time and look up neighbors in it at forecast time." in the line 186. A few lines below (lines 202-205) it's written "When ranking candidates for target i, we only consider contracts first observed no later than i itself".

  6. According to the lines 221-223, the FOMO forecasting is a TS foundation model forecasting where "up to ⌈L/2⌉\lceil L/2 \rceil ...processed form its own history". So, a ceiled half of the layers. It's also claimed that "Toto-4M only mixes series in its last block (layer 4 of 4)" (line 226). Table 1 claims that Toto-4M native mixing is FOMO (line 280).

could you please define:

  1. what does it mean to "share a field value" in the line 190?
  2. "the target’s volatility is as decodable" (line 346)
  3. "ridge probes" (line 369). Is it linear probing with ridge regression?
  4. "Native mixing is unpredictable, and FOMO is safe" (line 267)
  5. "Intermediate target representations largely carry the depth allocation effect." (line 372)
  6. "Early mixing displaces the target’s representation in every domain" (line 376)

miscellaneous/typos

  1. two captions of table 6
  2. line 135, CRPS has been proposed for the first time not by Gneiting and Raftery, 2007.
  3. could you please make the datasets distinguishable in figure 3?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Retrieval pipeline is NOT what our paper focuses on and same-domain context (market prediction) is EXACTLY what our paper is about

Thank you for taking the time, we'd likely to emphasize 2 key points on the focus/main contribution of this paper:

The retrieval pipeline was never the focus of the paper, the idea of the paper was to mask certain parts of where context is mixed in the model to minimize context degradation. The idea of the paper was also exactly what they say "it rests on", it is about the late-layer / altered layer interaction as well as using related market which is same-domain context to help with forecasting. The entire field of market forecasting rests on this premise, and so we would respectfully challenge this comment as we believe both points do not apply to our main contribution in this paper!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.