acceptodds
Under review as a conference paper at ICLR 2027

STEPPY: Multi-scale, multi-species wildlife movement forecasting

Abstract

Reliable forecasts of wildlife movement would enable numerous advances in ecology and conservation, but forecasting wildlife movement is a challenging predictive task, requiring models to represent complex, heterogeneous, and multi-scale movement across species and individual animals. Recent empirical studies indicate that even state-of-the-art wildlife movement forecasting methods only modestly outperform simple statistical approaches, even when leveraging complex environmental predictors. Further, existing methods are typically specialized for individual species and for specific predefined forecasting horizons, though research in ecology indicates that movement patterns may be shared across taxa and across different scales of movement. We propose a new movement forecasting model, STEPPY, that produces movement forecasts for any species and any time horizon dynamically at prediction time, leveraging ecologically-inspired architectural designs that enable it to learn useful and shared patterns across species and spatiotemporal scales. STEPPY outperforms existing state-of-the-art forecasting methods significantly, improving probabilistic forecasting by 18% on future trajectories from individuals who appear in the training data and 20% on trajectories for held-out individuals, even without incorporating any explicit environmental predictors. We will make our code, models, and data public upon publication.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
probabilistic forecasting, trajectory analysis, movement ecology

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YUtvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces STEPPY, a new model for forecasting wildlife movement by leveraging movement patterns across spatiotemporal scales and species. Using a transformer, STEPPY encodes key forecasting information: the forecast horizon, taxonomic information, population-level spatial scale, and previous trajectory observations with day, time, and location. These observations are represented with relative spatiotemporal encoding and then combined with random Fourier basis features to predict a spatial probability distribution of where the animal will be. STEPPY outperforms several baselines on the MoveBench dataset across all setups, showing that cross-study training conditioned on long historical contexts strongly benefits wildlife movement modeling.

Critical strengths
  • The paper is well written, motivated, and clear, making it enjoyable to read
  • Proposing novel approaches to improve and scale wildlife movement forecasting is crucial to guide conservation policies better.
  • The temporal flexibility and good performance of STEPPY across setups, with clear improvements of each of its components shown through a well-executed ablation study.
Critical weaknesses

Generally, my main concern is that the paper lacks a bit of ambition, in my opinion:

  • The paper builds on top of MoveBench, an unpublished paper uploaded to arXiv 10 days before the ICLR deadline, following a similar structure and reusing some of its content, for example, the performance plot figure, which is the same in the two papers, except for the additional STEPPY curve (Figure 2 in STEPPY and Figure 3 in MoveBench). This makes the paper seem incremental compared with the dataset paper. I therefore expect an innovative methodology with in-depth analysis and/or very strong performance to distinguish this paper from the dataset paper. I concede here that the gains are quite important ~20%, but on the other hand, the authors also admit in the abstract that the state-of-the-art baselines “only modestly outperform simple statistical approaches”, which may question this statement of L419: “high predictive accuracy can be achieved without explicit access to environmental information”. In particular, the energy score is somewhat difficult to interpret, and there are only a few (potentially cherry-picked) qualitative examples in both the main text and Appendix. Adding more examples and/or experiments could help convince me that the performance gains are significant and translate into different conservation decisions.
  • To follow up on the previous point, the reason I think the performance could even be higher is the choice not to use environmental variables. I agree that STEPPY outperforming the baselines without explicit environmental information is remarkable, but I see a missed opportunity to improve performance further by integrating this valuable information, which is a key driver of wildlife movements. Additionally, wildlife movement modeling is also about trying to understand the relationships between movement and the environmental drivers. Relying only on past trajectories seems conceptually insufficient to model wildlife movement well. I believe Figure 4b suggests that combining MoveFormer's ability to incorporate environmental variables for more fine-grained predictions with STEPPY's ability to better leverage historical context would improve prediction. I therefore see no strong reason not to include this information, especially since it is readily available in MoveBench, and the transformer could be trained to be robust to missing environmental variables when they are unavailable or too heavy to use, as has been done recently in ecological modeling.
Questions for authors

I would like to hear the author's opinion on the concerns mentioned above, whether they consider them valid, and whether they will conduct additional experiments to address them. I believe the paper is valuable, and I am strongly willing to increase my rating if the authors address my concerns.

2 comments

Thank you very much for your thoughtful review and feedback. Part of W1 (Interpretability of energy score) is related to a request for a breakdown of the metric requested by Reviewer sb62, and W2 (Using environmental information) was also a concern raised by Reviewer xfdy. We have addressed both in the shared response above. Please feel free to raise any additional questions or points for clarification in either thread. To address your additional concerns:

W1 (Dataset vs. method): MoveBench is a prior work that proposes a benchmark for animal movement forecasting. Our work evaluates on that benchmark and explicitly references it as prior work. Steppy is a distinct and novel methodological contribution, it is the first method to enable multi-scale multi-species movement forecasting, and, as you pointed out, it achieves very strong empirical results when compared to the baselines proposed in MoveBench. We agree that calling this “high predictive accuracy” is imprecise and have updated the language to “higher predictive accuracy than previously demonstrated,” but we do stand by the observation that prior deep learning methods for animal movement forecasting “only modestly outperform simple statistical approaches,” as this does seem to be true empirically. As discussed in Section 4.1, the next-best method at every timescale (i.e. the best possible existing method other than Steppy, selected per-timescale) is only 6% better than a correlated random walk on the MoveBench temporal test set and 2% better than a correlated random walk on the MoveBench individual test set. We think that this speaks to the difficulty of the probabilistic wildlife forecasting task. Steppy proposes a new kind of approach that is empirically a big step forward and we hope can help unlock even more progress for the field.

W1 (Experiments and qualitative examples): Thank you for this suggestion. We have added new quantitative results breaking down prediction cases with potential relevance to ecology and conservation (Appendix Section D.2), as well as a significant number of new qualitative results (Appendix Section E). For the prediction breakdown, we have focused on five common movement scenarios that can be identified by heuristics applied to ground truth trajectories: revisitation, site arrival and departure, continuation, and residence (see Section D.2 for precise definitions). Breaking down test cases by these scenarios has provided additional insights, in particular that Steppy’s largest improvements over other methods are most pronounced in cases where an animal is resident (relatively stationary in a local region for several successive movements), returning to a previously-visited location, or maintaining its direction of travel with respect to recent movements. These complement our other results that indicate the utility of Steppy’s contextual processing, which takes recent observations into account at prediction time more effectively than prior work.

Then, we have added qualitative comparisons of every method, both on the ecologically-relevant scenarios described above (in Section E.2), as well as on randomly-selected examples across all of MoveBench (in Section E.1). For the latter, we randomly selected examples from both the temporal and individual test sets, grouped by the same taxonomic and range groupings in Figure 2, resulting in 24 pages of qualitative prediction comparisons. We have found these examples very helpful for building intuition about the behavior of the different methods on various forecasting cases, and thank you for your suggestion.

I appreciate the effort made by the authors to address my concerns, particularly the additional experiments in Appendices D and E, which demonstrate the relevance of Steppy to ecology and conservation. As the authors rightly point out, "Steppy’s largest improvements over other methods are most pronounced in cases where an animal is resident (relatively stationary in a local region for several successive movements), returning to a previously-visited location, or maintaining its direction of travel with respect to recent movements." This also suggests that Steppy could potentially benefit from incorporating environmental variables, particularly in situations where past movement patterns are less informative, or trajectories are less predictable. I am therefore somewhat disappointed that environmental variables have not been considered. To me, this represents a step backward compared to previous approaches, rather than merely a direction for future work, as suggested by the authors. I also remain unconvinced by the computational argument, as I believe there are relatively straightforward ways to address these challenges while also ensuring robustness to missing environmental data. Incorporating such variables would, in my view, make Steppy a more reliable tool for practitioners.

Despite these reservations, I still believe the paper makes a valuable contribution. My assessment would ideally fall between 2 and 3, but given the limited granularity of this year's rating scale, I have decided to round up and increase my rating from 2 to 3.

Reviewer_sb62OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

They study wildlife movement prediction in this paper. They aim at solving complex, heterogeneous and cross-species movement prediction. To realize this, they propose STEEPY which learns from multiple species and at multiple spatiotemporal scales. Specially, relative spatiotemporal positional encoder and random Fourier basis spatial decoder are adopted. Experiments on the benchmark show the effectiveness of STEEPY.

Critical strengths

-- Detailed ablation study is carried out on every component, namely inputs and history, space/time relative positional encoding, decoder and other training techniques.

--In Discussion section, advantages of representation learning in STEEPY, deep reasons for improvement and future opportunities are analyzed in detail.

Critical weaknesses
  1. The novelty of the proposed method is relatively limited. Since all components adopted are existing modules, the technique contribution is not well demonstrated.
  2. More details about the comparison with baselines, as in Figure 2, are expected. For example, whether STEEPY’s performance is based on additional training data? Which baselines use environmental features? The results shown in Figure 2 indicate that some baselines achieve similar or even worse performance compared with the don’t move baseline. It is better to give some explanations.
  3. Energy Score consists of two components: Euclidean error of forecast and spread of forecast. It is preferable to report the variation in each component separately, so that the source of the overall improvement or change can be clearly identified.
  4. The model can learn cross-species motion features, but this is not clearly supported by experimental results. For example, in the caption of Figure 2, “produces 18% better forecasts than the next-best method for held-out future time points of individuals that occur in the training data, and 20% better forecasts for held-out individuals that do not”. It is unclear which subplot or which curve depicts learned species and which one corresponds to unknown species.
  5. Some typos and misleading sentences. Writing needs polish.
Questions for authors

Overall, the manuscript has a potentially interesting research topic, but the above concerns must be adequately addressed to establish the novelty and the claimed cross-species generalization. Further explanation of the evaluation is also expected.

2 comments

Thank you very much for your thoughtful review and considered feedback. W1 (Novelty), and W2 (Baseline details), and W3 (Energy score breakdown) were concerns shared by one additional reviewer, and we have addressed them in the shared response to all reviewers. Please feel free to raise any additional questions or points for clarification in either thread. To address your additional concerns:

W2 (Baselines, additional details): We have included significantly more details about the baselines at the beginning of Section 4 in the revised text; please see the shared response comment at the top of the page for more information. To respond to your additional points: 1. Steppy does benefit from additional training data, but this is not the source of most of its improvements over existing methods. We have ablated this in the final row of Table 1, and it can be seen that the extra training data improves energy scores by between 5–6%, whereas Steppy demonstrates 18–20% improvements in energy score over the next-best method overall. 2. We have added additional discussion about methods’ generalization performance, which is sometimes worse than the “don’t move” baseline, in Section 4.1 L355. Your observation is correct: sometimes existing methods generalize very poorly to new individuals. Our simple baselines (don’t move, and correlated random walk) serve as points of comparison for identifying these cases.

W4 and Q1 (Generalization): We first want to clarify that these generalization results are on the test sets in MoveBench, one of which (the “individual” test set) tests generalization to held-out individuals of observed species; there is no cross-species test set in MoveBench. We have added further details to the caption of Figure 2 to clarify this, and to clarify which subplots correspond to which test scenarios, thank you for the suggestion. Our discussion of learning cross-species features is separate from cross-species generalization—we do not claim to generalize to new species, but rather we show empirically that training a single model across all species (rather than a specialized per-species model) improves results overall. This is the observation supported by the “Cross-study training” row in the ablation, which shows that training across studies improves performance by 9.5–12.4% compared to training per-study (i.e. per-species) models.

W5 (Writing): We have taken another pass through the manuscript to address typos and clarify definitions and claims. In particular, with additional feedback from Reviewer xfdy, we have added mathematical definitions of common movement concepts and made references to these concepts more consistent. If there are any specific concerns with the writing that we have not addressed, please let us know and we will be happy to address them!

I appreciate the explanation about the baseline, sources of its improvements and additional discussion about the generalization and more detailed analysis, which address most of my concerns. Considering its methodological novelty, my rating falls between 2 and 3, finally rounded to 3.

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

Summary:

This paper introduces Steepy, the first multi-scale, multi-species wildlife movement forecasting framework. The forecasting framework is based on an autoregressive transformer architecture. The input is the locations and time of the animals, and the encoder is a transformer with causal self-attention on the observations. A relative spatiotemporal encoding is used to provide context about relative distance to other observations. The decoder maps the hidden representations to where the animals could locate. Multi-scale training is applied. The paper uses energy score, a proper scoring rule, to assess the prediction quality.

The paper experiments on MOVEBENCH, a benchmark for wildlife movement forecast. It shows that the proposed framework significantly outperforms prior frameworks in terms of prediction accuracy without using environment features. The paper also provides a detailed ablation study which explains how the proposed framework benefits from increasing context and better handles temporal gaps.

Critical strengths

[1] The paper presents solid progress towards wildlife tracking literature. The paper archives much better forecasting accuracy compared to prior baselines (see Figure 2 and 3), and it presents a very complete ablation study. The proposed framework can better leverage increasing context and better handles temporal gaps commonly seen in the data.

Critical weaknesses

[1] There is an insufficient introduction of baselines. For example, MoveFormer is the major baseline being discussed in the experiment section, but when it’s introduced the author only mentions it has environmental covariates. I think it’s important to give a more precise description of the baseline, so that the reader can better understand the limitations of prior methods and appreciate your framework. (P.S. I think there should also be more precise definitions of metrics in the paper. For example, I think step length is introduced below Figure 3. It’s not mathematically defined in the main text.)

[2] This is a nice application paper, but the technical novelty/innovation is moderate. It mostly applies existing tools into the new context of wildlife movement forecast.

Questions for authors

[1] Maybe I have missed this. The paper stresses that STEEPY achieves better prediction accuracy even without environmental features. Would the performance be better if the environmental features are included?

[2] At line 300, the author mentions MoveFormer truncates the context. What does truncate the context mean (does it simply mean throwing away data when large temporal gaps occur?) Since MoveFormer performance doesn't grow with context (Figure 4a), having the context truncated is not the main reason it underperforms. Is this understanding correct?

2 comments

Thank you very much for your thoughtful review. The first part of W1 (Baseline details), as well as Q1 (Environmental features) were highlighted by one other reviewer as well, and we have addressed them in the shared response above. Please feel free to raise any additional questions or points for clarification in either thread. To address your additional concerns:

W1 (Definitions): We have added additional definitions in Section 2 and more consistent references to terminology throughout. In particular, we added mathematical definitions of common movement concepts: step, step length,bearing, turn angle, and burst. Thank you for this suggestion which has helped clarify many concepts throughout the paper. We have also clarified the explanation of the step length results in Figure 3 in Section 4.1 (paragraph starting at L360).

Q2 (Truncating context): Yes, “truncating” means discarding older observations. MoveFormer does this when large temporal gaps occur, discarding observations before the gap. To make this more clear, we have added a definition of “bursts” (contiguous trajectory subsequences) in Section 2, clarified the explanation text you referenced (L366), and changed our use of the word “truncate” to be more consistent. We have also added clarification to Figure 4. In response to your question, the x axis of Figure 4a represents the maximum amount of context available to both methods; they differ in how they construct the actual context window. This is one reason why MoveFormer’s performance does not improve with more context—it does not fully utilize it because of its truncation behavior. To clarify and illustrate this, we have added an additional subplot (Figure 4b) to illustrate the discrepancy between the maximum allowed context length and the actual length of the context constructed by each method. Note that this is not the only reason it underperforms, as Fig 4c illustrates a case where both methods have 64 observations in context at prediction time, but MoveFormer still underperforms compared to Steppy.

Reviewer_xfdyOpenReview ↗

Questions Addressed

Thank you for the rebuttal! I think my question has been addressed. I believe the presentation has been improved with clearer definition of baselines and terminologies. I plan to keep my original score (weak accept).

Comments on the paper

3 comments
AuthorsOpenReview ↗

Author response to reviewers

We would like to thank all reviewers for their constructive feedback. We were encouraged that reviewers consistently noted the strengths of our proposed wildlife forecasting method, Steppy, which significantly outperforms prior work across a range of species and spatiotemporal scales, even when compared to species-specific and scale-specific models, with additional benefits such as flexibility to accommodate common real-world wildlife data issues such as temporal gaps (Reviewers xfdy and YUtv). All three reviewers also noted the strength of our ablation study, which was described as “very complete” (xfdy), “detailed” (sb62), and “well-executed … with clear improvements of each of its components shown” (YUtv).

Reviewers had insightful feedback that we believe has strengthened the work, and also raised a few shared concerns which we have carefully addressed. We have posted a revision of the manuscript with our improvements, and we add additional details on the updates here and in responses to individual reviewers. To avoid repetition, we have addressed any concerns raised by multiple reviewers in this central comment thread. We address any additional reviewer-specific concerns in comments to the individual reviewers below. We thank you all again for your thoughtful reviews, which have helped improve the manuscript, and we look forward to further discussion.

AuthorsOpenReview ↗

Shared concerns 1 and 2

More baseline details (Reviewers xfdy and sb62): Reviewers felt that more implementation details on baselines would be beneficial for interpreting the results and to clarify methodological differences. We agree and have added an additional section to the beginning of Section 4 that includes: (1) a more significant introduction to each baseline, and (2) an additional paragraph explicitly comparing and contrasting our method with MoveFormer (the previous state-of-the-art method, and the one that is conceptually most similar to ours) to make our methodological contributions more clear.

Including environmental variables (Reviewers xfdy and YUtv): Steppy demonstrates that it is possible to predict wildlife movement far more accurately than was previously thought using trajectory information alone. This is counter to prevailing ideas in the field and we believe opens up many new exciting possibilities. As emphasized in our paper, Steppy is significantly more predictive than the previous state-of-the-art, improving next-location predictions by 18–20% over the next-best method on average. The fact that it is possible to do so with only trajectory inputs, when the baselines frequently have access to additional environmental covariates, indicates that it is possible to learn much more from recent movement context than has been previously demonstrated, and our design allows us to learn and utilize this information much more effectively than prior work. We hypothesize that this is because there is actually a lot of environmental context implicit in the movement trajectories themselves, i.e., recent movement itself says a lot about the environmental conditions that the animal is moving within. For instance, animals may move more slowly and tortuously in dense forested environments compared to open desert landscapes, and may revisit a resource-abundant location or nest that they have been to recently. We agree that incorporating explicit environmental information is a very interesting future direction, and may help improve forecasting performance in situations where explicit environment variables provide additional information that is not already implicitly captured by recent movement history. However there are many open research and engineering challenges in figuring out what kinds of environmental context is beneficial (i.e. what variables and at what spatiotemporal scales), and how to incorporate this context in a useful way—for instance, one engineering challenge is the scale of environmental raster storage, particularly when using long context windows, which introduces significant data loading and GPU memory challenges. We hope that Steppy can provide a foundation for future methods to explore these questions in depth.

AuthorsOpenReview ↗

Shared concerns 3 and 4

Novelty (Reviewers xfdy and sb62): Our contribution is a new model architecture designed specifically to address the unique challenges of probabilistic wildlife movement forecasting, and demonstrating very strong empirical performance compared to existing approaches. While several of the individual components used in our method were previously established—for example transformers, rotary position encodings, and coordinate-based MLPs—there is strong precedent for both the novelty and impact of adapting valuable methodological components to a new problem. Notable examples include ViTs, which adapt transformers for images (Dosovitskiy et al., ICLR 2021); DETR, which adapts ViTs for object detection (Carion et al., ECCV 2020); VideoRoPE, which adapts rotary positional encodings to video (Wei et al., ICML 2025); and NeRF, which adapts coordinate-based MLPs for 3D scene representation (Mildenhall et al., ECCV 2020).

Steppy contains a number of domain-specific innovations to address the distinctive structure and requirements of wildlife trajectory data, which consists of irregularly spaced observations that vary greatly in spatiotemporal scale. Existing approaches to address these challenges have required subsampling and truncating trajectories to make them more homogeneous before inputting them to models, which discards significant information, as well as training specialized models that are limited to one forecast horizon. We show that it is possible to both utilize the raw, unfiltered observations to improve performance, and to train a single unified model that can generate better forecasts than scale-specific models. One way we enable this is with a novel adaptation of rotary positional encodings. Prior work represents the spatial location of each animal observation as input features, and uses standard positional encodings to represent the index of each observation in the input sequence. Rotary positional encodings instead enable positional encodings to be based on inputs’ relative positions to each other, rather than their absolute position in the sequence, by rotating query and key vectors inside the attention mechanism. We identify this as an opportunity to more flexibly handle observational data with irregular sampling rates by using time instead of index as the variable that determines the rotation; further, we identify the opportunity to use relative positional encodings to also incorporate relative spatial locations of observations—in world coordinates—to more easily support learning representations that depend upon where the animal is in relation to where it has been before, an idea that is supported by ecological work on spatial memory. To the best of our knowledge we are the first to propose such a mechanism for forecasting, and we show that this leads to significantly improved performance over alternative designs (see Table 1, Space/time RPE rows). We further show that it is possible to train a single model that can forecast any horizon dynamically—a capability not exhibited by prior work—through a multi-scale training strategy that scales the spatiotemporal positional encodings according to the movement scale and requested forecast horizon. We have reworked Section 3 to highlight the novel dimensions of our approach more clearly. Thank you for your feedback.

Interpreting the energy score (Reviewers sb62 and YUtv): Reviewer YUtv noted that the energy score, the main metric used by MoveBench for evaluating probabilistic forecasts, can be difficult to interpret, and Reviewer sb62 proposed addressing this by breaking down the metric into its constituent components. We think this is a great suggestion and have added a breakdown of the energy scores in Appendix C.1 with a more in-depth discussion of the sources of overall improvement or change. The results indicate that Steppy improves its energy scores compared to other methods by improving the expected Euclidean error of its forecasts while simultaneously making them tighter (lower spread). The results from this analysis also corroborate results from Kay et al. (2026) demonstrating why a proper scoring rule is necessary for evaluating probabilistic predictions, as simply measuring expected Euclidean error rewards methods that hedge toward the geometric median of the distribution rather than methods that predict the true distribution (more details on this as well in Appendix C.1).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.