A new preprint reports that a forecasting model called TransMod achieved the lowest listed errors when predicting urban mobility demand in New York City and Chicago. Its strongest test came when the model had limited historical ride-hailing data to learn from, using information from bike-sharing and metro systems instead.
The evaluation compares forecasts with three measures of error: mean absolute error, root mean square error, and mean absolute percentage error. The paper reports lower values for TransMod across the listed comparisons, but the supplied analysis does not report inferential uncertainty or statistical significance tests for that result.
Borrowing information across transport modes
TransMod is designed for a data mismatch between transport systems. Bike-sharing and metro are represented by stations, while the ride-hailing data are grouped into broader geographic zones. The model uses a spatially grounded soft assignment, a weighting approach that aggregates fine-grained station information into shared zone-level representations across mobility modes.
The paper’s main protocol withholds ride-hailing temporal demand during pre-training and transfer. Only a limited set of ride-hailing labels is used for supervised adaptation, while held-out records are reserved for evaluation. During that adaptation, the model’s memory pool and spatial mapping modules remain fixed, while its prompt network and prediction head are fine-tuned.
Before that adaptation, source training combines a forecasting loss with cross-modal alignment, geographic assignment regularization, and a regularizer intended to keep the model’s memory prototypes, or stored patterns, diverse. In plain terms, the system is trained both to forecast and to keep information from different modes and locations usable together.
The reported results
In the reported city comparisons, TransMod’s New York City scores were a mean absolute error of 10.78, a root mean square error of 13.16, and a mean absolute percentage error of 8.76%. In Chicago, the corresponding values were 8.30, 11.37, and 7.52%. The paper reports that these were the lowest MAE, RMSE, and MAPE values among the evaluated methods in both cities.
The cross-city test points in the same direction. When only 10% of Chicago data were used, the reported RMSE was 12.38 with transfer, compared with 16.21 without transfer. That is a result from this particular experiment, not proof that the same difference will appear in other cities or data settings.
A separate scarcity analysis reported comparable performance from TransMod using 30% of ride-hailing records, versus methods trained with 70% to 100% of the data. The study also reports that, for a longer forecast from time t to time t + 12, TransMod’s RMSE rose by 28.3%, compared with increases of 40% to 50% for most baselines.
Which parts of the model mattered
The paper’s component tests examined the shared spatial representation and memory-based transfer mechanism. Removing soft assignment increased MAE by 13.0% in New York City and 16.0% in Chicago. Replacing memory-based transfer with a GRU decoder increased MAE by 8.8% and 11.8%, respectively.
The model was also tested after all auxiliary features were removed. That change increased MAE by 10.8% in New York City and 11.8% in Chicago, a degradation the paper describes as non-catastrophic.
TransMod was reported to adapt faster than the listed competing transfer methods in the New York City test. Its transfer or adaptation time was 5.2 seconds per epoch, compared with 24.3 to 41.2 seconds per epoch for those competing methods.
A promising benchmark, with a narrow test
The evidence comes from a modeling study using shared-mobility records from New York City and Chicago. The modeled systems include station-based bike-sharing and metro, plus a zone-based ride-hailing system whose trips are aggregated over spatial zones. The paper specifically reports January 2018 bike-sharing data and does not report total observations, trip counts, zone counts, or the number of experimental repetitions.
The supplied analysis also notes that the meaning of the plus-or-minus values shown alongside some reported results is not defined. Statistical significance tests are not reported, and the limited-target-history protocol does not establish how the model would perform when substantial historical ride-hailing demand is already available.
The findings support a narrower conclusion: in the tested New York and Chicago benchmark settings, TransMod was reported to transfer information across mobility modes and to hold up under reduced data, spatial mismatch, and longer forecasting horizons. They do not establish gains in real-world transport operations or generalization to cities, years, mobility modes, or data distributions that were not evaluated.
The document is a version-one arXiv preprint dated 28 August 2026. Its results should therefore be read as reported findings from an early computational evaluation, with broader testing still needed before conclusions are extended beyond the settings studied.
Paper data and sources
Original title: Learning to Transfer Across Modes: Towards Unified Urban Mobility Forecasting
Authors: Yixuan Zhao, Man Luo
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text