T-STAR: A Context-Aware Transformer Framework for Short-Term Probabilistic Demand Forecasting in Dock-Based Shared Micro-Mobility

Gon\c{c}alo Homem de Almeida Correia; Jingyi Cheng; Oded Cats; Shadi Sharif Azadeh

REVIEW 3 major objections 2 minor 1 cited by

T-STAR uses a two-stage transformer to forecast 15-minute bike-sharing demand by separating hourly patterns from localized fluctuations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-05-21 13:02 UTC pith:YR6BOQ2G

load-bearing objection T-STAR adds a practical two-stage transformer with metro context for 15-minute bike demand forecasts but the temporal splits need checking before the gains can be trusted. the 3 major comments →

arxiv 2602.06866 v2 pith:YR6BOQ2G submitted 2026-02-06 cs.LG

T-STAR: A Context-Aware Transformer Framework for Short-Term Probabilistic Demand Forecasting in Dock-Based Shared Micro-Mobility

Jingyi Cheng , Gon\c{c}alo Homem de Almeida Correia , Oded Cats , Shadi Sharif Azadeh This is my paper

classification cs.LG

keywords demand forecastingbike sharingtransformerprobabilistic forecastingmicro-mobilitytime seriescontext-aware

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents T-STAR as a transformer framework that predicts station-level demand for shared bikes every 15 minutes. It splits the task into a first stage that learns broad hourly patterns and a second stage that adds recent local changes plus real-time metro service data. Time series transformers in both stages produce forecasts that include uncertainty ranges. Experiments on Washington D.C. Capital Bikeshare records show gains in both point accuracy and probabilistic calibration, plus the ability to forecast in brand-new areas without retraining. Reliable short-term demand estimates of this kind could help operators reposition bikes faster and support integrated trip planning that combines bikes with transit.

Core claim

T-STAR is a context-aware transformer framework that forecasts station-level bike-sharing demand at 15-minute resolution through a hierarchical two-stage structure. The first stage captures coarse-grained hourly demand patterns; the second stage refines predictions by incorporating high-frequency localized inputs, including recent fluctuations and real-time demand variations in connected metro services, to account for temporal shifts. Time series transformer models generate probabilistic predictions in both stages.

What carries the argument

Hierarchical two-stage structure with time series transformers that first extracts coarse hourly demand patterns and then refines them using high-frequency localized inputs including metro variations.

Load-bearing premise

The two-stage structure is assumed to cleanly separate consistent demand from short-term fluctuations without bias from feature choices or data splits.

What would settle it

A drop in 15-minute forecast accuracy on the Capital Bikeshare dataset when the second-stage inputs or metro variations are removed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Outperforms prior methods on both deterministic point forecasts and probabilistic accuracy measures.
Maintains performance across many stations and different time periods.
Transfers to previously unseen service areas through zero-shot forecasting without retraining.
Supplies uncertainty-aware outputs that can feed directly into real-time fleet operations and multimodal trip planners.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same separation of scales might improve short-term forecasting for other dockless or dock-based vehicles such as e-scooters.
Operators could feed the probabilistic outputs into optimization routines that decide bike rebalancing every 15 minutes.
Adding weather or event data to the second stage could further reduce error on days with irregular demand.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

T-STAR adds a practical two-stage transformer with metro context for 15-minute bike demand forecasts but the temporal splits need checking before the gains can be trusted.

read the letter

Hi, the main thing here is that T-STAR uses a hierarchical transformer to first capture hourly demand patterns and then refine them with 15-minute inputs that include recent fluctuations and real-time metro variations, producing probabilistic station-level forecasts. They test it on Capital Bikeshare data from Washington D.C. and report better deterministic and probabilistic accuracy than baselines, along with spatial-temporal robustness and zero-shot transfer to new areas without retraining. The architecture is a reasonable response to the mix of regular and sudden shifts in micro-mobility demand, and pulling in metro data as an external signal is a straightforward way to add context that operators would actually care about. The experiments on real public data plus the robustness and transfer tests give the work some practical grounding that pure synthetic studies lack. The soft spot is the data partitioning. In 15-minute time-series forecasting, especially with a coarse stage feeding a fine stage, any non-chronological split or station grouping that lets future patterns leak into training can inflate the reported margins and make the zero-shot results look stronger than they are. The abstract describes the two-stage design but does not detail the exact train-test division or how overlap was prevented, so the outperformance claims rest on evidence that is not yet fully verifiable. This paper is aimed at people who build or use operational forecasting tools for shared mobility and multimodal planning. A reader who needs concrete methods for high-resolution probabilistic demand with uncertainty estimates would find the framework and the D.C. case study useful. It shows clear problem framing and engages with existing transformer time-series work, so it deserves a serious referee even if the validation details require more attention. I would send it for peer review but ask reviewers to focus first on the temporal integrity of the splits.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces T-STAR, a two-stage transformer-based probabilistic framework for station-level bike-sharing demand forecasting at 15-minute resolution. The first stage captures coarse hourly demand patterns while the second stage incorporates high-frequency localized inputs such as recent fluctuations and real-time metro demand variations. Experiments on Washington D.C. Capital Bikeshare data are reported to show outperformance over existing methods in both deterministic and probabilistic accuracy, along with spatial-temporal robustness and effective zero-shot transfer to unseen service areas without retraining.

Significance. If the empirical results hold under rigorous temporal partitioning, the work would provide a practical hierarchical transformer approach for uncertainty-aware short-term forecasting in micro-mobility, supporting real-time operations and multimodal integration. The zero-shot transfer experiment and use of real public data with robustness checks across stations and periods represent concrete strengths.

major comments (3)

[Experimental Setup] The experimental setup section does not provide a clear description of the chronological train-validation-test split or any safeguards against temporal leakage. Given that the second stage uses recent fluctuations and metro variations at 15-minute resolution, any non-strict temporal partitioning (or shared time windows between coarse and fine stages) risks future information leaking into predictions, which would undermine the outperformance and zero-shot transfer claims.
[Results] The results section reports outperformance and robustness but the abstract and main text provide no specific quantitative metrics (e.g., MAE, RMSE, CRPS, or interval scores), baseline implementations, or statistical significance tests. This makes it impossible to evaluate the magnitude of improvements or whether they exceed what could arise from leakage or hyperparameter tuning.
[Framework Design] The framework design section assumes the hierarchical two-stage structure cleanly disentangles consistent demand from short-term fluctuations without bias from feature selection or data partitioning. No ablation studies on the metro variation inputs or sensitivity analyses to alternative splits are described, leaving the central premise unverified.

minor comments (2)

[Abstract] The abstract would be strengthened by including at least one key quantitative result (e.g., percentage improvement in CRPS) to support the outperformance claim.
[Methods] Notation for probabilistic outputs (e.g., quantiles or distribution parameters) should be defined consistently in the methods section and reused in figures and tables.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough and constructive review of our manuscript. We have addressed each major comment point by point below, and we will incorporate revisions to enhance the manuscript's clarity and completeness.

read point-by-point responses

Referee: [Experimental Setup] The experimental setup section does not provide a clear description of the chronological train-validation-test split or any safeguards against temporal leakage. Given that the second stage uses recent fluctuations and metro variations at 15-minute resolution, any non-strict temporal partitioning (or shared time windows between coarse and fine stages) risks future information leaking into predictions, which would undermine the outperformance and zero-shot transfer claims.

Authors: We appreciate the referee's concern regarding potential temporal leakage, which is critical for time series forecasting validity. We confirm that the experiments employed a strict chronological split to ensure causality. However, we acknowledge that the description in the manuscript was insufficiently detailed. In the revised version, we will add a dedicated paragraph in the Experimental Setup section that specifies the exact time ranges for training, validation, and test sets, and explicitly states that all inputs, including recent fluctuations and metro variations, are based solely on data available prior to the forecast time. We will also include safeguards such as ensuring no overlap in time windows between stages. revision: yes
Referee: [Results] The results section reports outperformance and robustness but the abstract and main text provide no specific quantitative metrics (e.g., MAE, RMSE, CRPS, or interval scores), baseline implementations, or statistical significance tests. This makes it impossible to evaluate the magnitude of improvements or whether they exceed what could arise from leakage or hyperparameter tuning.

Authors: We agree that including specific quantitative results would allow readers to better assess the improvements. Although the manuscript states that T-STAR outperforms existing methods, we will revise the Results section to include a comprehensive table reporting exact values for key metrics such as MAE, RMSE, CRPS, and interval scores for T-STAR and all compared baselines. We will also detail the baseline implementations and add statistical significance tests (e.g., using paired statistical tests) to demonstrate that the observed improvements are significant and not due to random variation or tuning. revision: yes
Referee: [Framework Design] The framework design section assumes the hierarchical two-stage structure cleanly disentangles consistent demand from short-term fluctuations without bias from feature selection or data partitioning. No ablation studies on the metro variation inputs or sensitivity analyses to alternative splits are described, leaving the central premise unverified.

Authors: We thank the referee for this suggestion to further validate the framework design. To address this, we will conduct and report additional ablation studies in the revised manuscript. These will include variants of the model with and without the metro variation inputs, as well as comparisons of the two-stage approach versus a single-stage model. Furthermore, we will perform sensitivity analyses using alternative temporal splits to verify the robustness of the results. These new experiments and analyses will be added to the Experiments section to substantiate the central premise of the hierarchical structure. revision: yes

Circularity Check

0 steps flagged

No circularity: T-STAR predictions arise from trained transformer on external data with no self-referential reductions

full rationale

The paper introduces a hierarchical two-stage transformer architecture that first models coarse hourly demand patterns and then refines them with high-frequency localized inputs including recent fluctuations and metro variations. Probabilistic forecasts are generated by training these models on observed Capital Bikeshare data and evaluating deterministic and probabilistic accuracy plus zero-shot transfer. No equations, ansatzes, or uniqueness theorems are presented that reduce outputs to fitted parameters by construction, nor does any load-bearing step rely on self-citation chains or rename known results. The derivation chain is therefore self-contained and externally grounded in independent training data and standard transformer training procedures.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard transformer time-series modeling assumptions and the premise that demand exhibits separable coarse and fine temporal scales; no new physical entities or ad-hoc constants are introduced beyond typical neural network hyperparameters.

free parameters (1)

transformer hyperparameters (layers, heads, embedding size)
Chosen during model development and training on the target dataset; typical for deep learning architectures.

axioms (1)

domain assumption Time series demand data contains hierarchical temporal structure that can be disentangled into coarse hourly patterns and short-term fluctuations.
Invoked to justify the two-stage design described in the abstract.

pith-pipeline@v0.9.0 · 5786 in / 1185 out tokens · 69824 ms · 2026-05-21T13:02:02.215386+00:00 · methodology

0 comments

read the original abstract

Reliable short-term demand forecasting is essential for managing shared micro-mobility services and ensuring responsive, user-centered operations. This study introduces T-STAR (Two-stage Spatial and Temporal Adaptive contextual Representation), a novel transformer-based probabilistic framework designed to forecast station-level bike-sharing demand at a 15-minute resolution. T-STAR addresses key challenges in high-resolution forecasting by disentangling consistent demand patterns from short-term fluctuations through a hierarchical two-stage structure. The first stage captures coarse-grained hourly demand patterns, while the second stage improves prediction accuracy by incorporating high-frequency, localized inputs, including recent fluctuations and real-time demand variations in connected metro services, to account for temporal shifts in short-term demand. Time series transformer models are employed in both stages to generate probabilistic predictions. Extensive experiments using Washington D.C.'s Capital Bikeshare data demonstrate that T-STAR outperforms existing methods in both deterministic and probabilistic accuracy. The model exhibits strong spatial and temporal robustness across stations and time periods. A zero-shot forecasting experiment further highlights T-STAR's ability to transfer to previously unseen service areas without retraining. These results underscore the framework's potential to deliver granular, reliable, and uncertainty-aware short-term demand forecasts, which enable seamless integration to support multimodal trip planning for travelers and enhance real-time operations in shared micro-mobility services.

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

T-STAR addresses key challenges in high-resolution forecasting by disentangling consistent demand patterns from short-term fluctuations through a hierarchical two-stage structure. The first stage captures coarse-grained hourly demand patterns, while the second stage improves prediction accuracy by incorporating high-frequency, localized inputs, including recent fluctuations and real-time demand variations in connected metro services
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat recovery unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The Transformer architecture offers several key advantages for demand forecasting, including the ability to model complex temporal dependencies without recurrence

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting
cs.LG 2026-07 conditional novelty 4.0

STAGformer forecasts bike-station demand with linear-complexity spatial-temporal agent attention and reports lower RMSE/MAE than listed baselines on NYC and Chicago data.