acceptodds
Under review as a conference paper at ICLR 2027

Traffic24: A 24-Year Traffic Dataset and Benchmark for Long-Term Forecasting

Abstract

Understanding how traffic forecasting models benefit from extended data histories and generalize across changing conditions requires reusable longitudinal datasets. We introduce Traffic24, a curated traffic dataset covering 24 years of observations in San Diego, from September 2001 to September 2025. Traffic24 contains 505.6 million station–time records at five-minute resolution, including traffic flow, speed, occupancy, truck-related measurements, and lane-level observations. Six nested historical subsets, spanning one to 24 years, support the construction of forecasting tasks with different amounts of historical data. A weather-augmented subset adds regional meteorological context aligned with traffic observations through a shared UTC time axis. Standardized schemas, explicit field definitions, and retained quality and processing indicators support consistent data access and interpretation. To establish an initial forecasting benchmark, we evaluate multiple models on selected subsets using the open-source TimeCopilot and report their predictive performance. These evaluations provide reference results for subsequent research using Traffic24. By bringing together long observational histories, structured temporal subsets, and regional weather context, Traffic24 supports research into how much historical data forecasting models can effectively use, how they generalize over time, and how environmental information can inform traffic prediction.

Primary area:
datasets and benchmarks
Keywords:
Datasets and Benchmarks, Longitudinal Data, Long-Term Forecasting, Spatiotemporal Traffic-Weather Data, Traffic Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_fxthOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Traffic24, a San Diego traffic dataset spanning September 2001 to September 2025. It contains approximately 505.6 million station–time records from 261 traffic nodes at five-minute resolution, combining traffic measurements, data-quality indicators, and regional weather data aggregated from 25 weather stations. The authors provide a unified data schema and preprocessing pipeline and benchmark 29 models for one-step-ahead speed prediction using the preceding hour of speed observations. XGBoost achieves the lowest reported MAE and RMSE among the evaluated models.

Critical strengths

Traffic24 integrates 24 years of traffic observations with temporally aligned regional weather data, providing a resource for studying long-term changes in traffic patterns, cross-year generalization, and the impact of weather conditions on traffic prediction performance.

Critical weaknesses

1. More direct empirical evidence would help substantiate the dataset’s distinctive value. Traffic24 primarily curates and integrates two existing data sources—PeMS traffic records and MesoWest weather observations—through temporal alignment and regional weather aggregation. Its extended temporal coverage and inclusion of weather information are potentially useful. However, the reported benchmark focuses on five-minute-ahead speed prediction using only historical speed observations. It does not compare different historical training spans, evaluate the contribution of weather information, or explicitly examine temporal distribution shift. The current experiments therefore provide limited evidence of the benefits of the dataset’s distinguishing features. Focused comparisons addressing these aspects would help clarify its contribution relative to existing resources.

2. Some passages leave the completeness of the submission and the reproducibility of its analyses unclear. For example, Section 6.2 states:

“These plots are descriptive: the figure does not specify the station-grouping rule or the treatment of missing and flagged values, so the apparent between-group gradient should not be interpreted as a validated population-level finding.”

This caution is reasonable, but the grouping criteria and preprocessing procedures used to generate the plots should also be documented so that readers can assess the analysis. Similarly, the Reproducibility Statement reads:

“All benchmark models are run with the open-source TimeCopilot framework. The dataset, anonymous download link, and exact evaluation code/configuration are not included in the source package and must be supplied or described before submission.”

It is unclear whether this statement refers to the current submission or an earlier stage of preparation. There is also a discrepancy between the claim that the full evaluation protocol is provided and the acknowledgment in Section 6.1 that split boundaries and missing-window handling are unspecified. Clarifying these statements and providing the corresponding materials and methodological details would help readers assess the submission and reproduce its analyses.

Questions for authors
  1. Could the dataset, along with the code used for its construction and evaluation, be made available as supplementary material or through an anonymous repository?

  2. Which historical data configuration and station set were used in the reported experiments? What were the exact dates of the training, validation, and test periods, and how many valid samples were included in the evaluation?

  3. Were the models trained independently for each station or jointly across stations? If the training setup differed across models, please specify the setting used for each.

  4. Could you compare shorter and longer training histories while keeping the test period, station set, and sample-construction rules fixed to assess the value of the extended historical coverage? Similarly, a paired comparison of traffic-only and traffic-plus-weather inputs under the same evaluation conditions would help establish the contribution of weather information.

Reviewer_Ynj8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Traffic24 that compiles 24 years of traffic observations from San Diego, with regional weather information and six nested historical subsets. It standardizes data schemas and reports results from 29 forecasting models on a five-minute speed prediction task.

Critical strengths
  1. The long observation period offers potential value for studying temporal distribution shifts and the usefulness of historical data.
  2. Standardized traffic and weather fields could reduce data preparation effort for subsequent research.
Critical weaknesses
  1. The construction of six nested datasets is useful, but it does not itself demonstrate their research value. Without experiments demonstrating the effect of historical coverage, the manuscript does not validate the central motivation behind Traffic24. A necessary experiment would fix the same target evaluation period, sensor set, prediction horizon, and preprocessing procedure while progressively increasing the amount of training history.

  2. The paper provides no experimental evidence that the added meteorological variables are useful for forecasting. It repeatedly emphasizes the potential for weather-aware forecasting, analysis of adverse weather, and environmental-context modeling. However, it evaluates all 29 models using traffic speed alone. No model is evaluated using the traffic-plus-weather configuration.

3.The dataset lacks quantitative quality audits. No dataset access or evaluation code is provided, and exact temporal splits and missing-window handling remain unspecified.

  1. Insufficient comparison with recent traffic datasets and benchmarks. The related-work discussion overlooks several recent studies that are directly relevant to the central contribution of Traffic24, such as XXLTraffic (ACM SIGSPATIAL 2025) and its subsequent extension, EvoXXLTraffic (2026).
Questions for authors

See the Critical Weaknesses.

Reviewer_y764OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Traffic24 assembles 24 years of 5-minute PeMS measurements from 261 traffic nodes in the San Diego region, aligns them with a regional MesoWest weather series, and organises the resulting 505.6 million station–time rows into six nested spans with traffic-only and weather-augmented versions. Its initial benchmark evaluates 29 models through TimeCopilot on next-step speed prediction from one hour of speed history.

Critical strengths

The paper targets a real and practically relevant need in traffic forecasting: long-span observations with consistent semantics for studying temporal change, the value of data history and weather effects.

Critical weaknesses

Regarding dataset: The submission didn't provide data access, which make the paper's central resource difficult to verify. On the other hand, the positioning relative to existing long-span PeMS resources remains unclear (e.g., XXLTraffic [a] and LargeST [b]) .

Regarding experiments: The current benchmark provides limited evidence for the dataset's intended uses, i.e., only one-step speed prediction from one hour of speed history, using no weather and data whose historical span and split boundaries are unspecified. It remains unclear what the benchmark establishes beyond an initial next-step reference.

Questions for authors
  1. The title frames Traffic24 as a benchmark "for Long-Term Forecasting", a term that in the time series literature denotes long forecast horizons (e.g., 96–720 steps in [c]), while the only task is a one-step 5-minute forecast?
  2. Why does the benchmark not include standard spatio-temporal graph baselines such as DCRNN, STGCN, Graph WaveNet, AGCRN and PDFormer, which are commonly evaluated on the traffic datasets?

References

[a] XXLTraffic: Expanding and Extremely Long Traffic Forecasting beyond Test Adaptation. SIGSPATIAL 2025.

[b] LargeST: A Benchmark Dataset for Large-Scale Traffic Forecasting. NeurIPS 2023.

[c] Are Transformers Effective for Time Series Forecasting? AAAI 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.