Time Series Foundation Models as Strong Baselines in Transportation Forecasting: A Large-Scale Benchmark Analysis
| Authors | Javier Pulido & Filipe Rodrigues |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2602.24238 |
| Download | |
| Categories | cs.LG |
Abstract
Accurate forecasting of transportation dynamics is essential for urban mobility and infrastructure planning. Although recent work has achieved strong performance with deep learning models, these methods typically require dataset-specific training, architecture design and hyper-parameter tuning. This paper evaluates whether general-purpose time-series foundation models can serve as forecasters for transportation tasks by benchmarking the zero-shot performance of the state-of-the-art model, Chronos-2, across ten real-world datasets covering highway traffic volume and flow, urban traffic speed, bike-sharing demand, and electric vehicle charging station data. Under a consistent evaluation protocol, we find that, even without any task-specific fine-tuning, Chronos-2 delivers state-of-the-art or competitive accuracy across most datasets, frequently outperforming classical statistical baselines and specialized deep learning architectures, particularly at longer horizons. Beyond point forecasting, we evaluate its native probabilistic outputs using prediction-interval coverage and sharpness, demonstrating that Chronos-2 also provides useful uncertainty quantification without dataset-specific training. In general, this study supports the adoption of time-series foundation models as a key baseline for transportation forecasting research.
Engineering Breakdown
Plain English
This paper tests whether Chronos-2, a general-purpose time-series foundation model, can forecast transportation data without any task-specific training or fine-tuning. The authors benchmark Chronos-2 in zero-shot mode across ten real-world transportation datasets covering highway traffic, urban speeds, bike-sharing demand, and EV charging patterns. They find that Chronos-2 achieves state-of-the-art or competitive performance without modification, suggesting that foundation models trained on diverse time-series data can transfer effectively to new domains. This challenges the conventional wisdom that transportation forecasting requires custom architectures, careful hyperparameter tuning, and dataset-specific training.
Core Technical Contribution
The core contribution is demonstrating that foundation models pre-trained on diverse time-series data can serve as strong zero-shot baselines for specialized forecasting tasks, eliminating the need for task-specific architecture design and fine-tuning. Rather than inventing a new architecture or algorithm, the authors provide empirical evidence that Chronos-2—a model trained on broad time-series patterns—generalizes remarkably well to transportation forecasting across heterogeneous datasets. This reframes foundation models from domain-specific tools into universal time-series forecasters, similar to how large language models generalize across NLP tasks. The methodological contribution is establishing a rigorous benchmark protocol for evaluating foundation model zero-shot performance across transportation domains.
How It Works
Chronos-2 operates as a generalist time-series model that has been pre-trained on massive amounts of diverse temporal data, learning patterns of trends, seasonality, and autocorrelation that generalize across domains. At inference time, the model takes a historical time-series window (the lookback period) as input and produces probabilistic forecasts for the future horizon directly, without any adaptation or fine-tuning. The model likely encodes the input time series into learned representations and uses a transformer or attention-based decoder to generate forecasts, similar to sequence-to-sequence architectures. For each transportation dataset tested—traffic volume, speed, bike demand, charging patterns—the same pre-trained weights and hyperparameters are used; the model makes no assumptions about the domain or task structure. The evaluation compares Chronos-2's zero-shot performance against baselines that include both classical time-series methods (ARIMA, exponential smoothing) and task-specific deep learning models that were trained on each dataset.
Production Impact
This work dramatically simplifies deployment pipelines for transportation forecasting systems: instead of collecting labeled data, designing custom architectures, and running expensive hyperparameter search, engineers can deploy Chronos-2 as a ready-to-use baseline with competitive performance on day one. For companies building real-time traffic prediction, demand forecasting, or infrastructure planning systems, this means reducing the engineering effort from months of model development to hours of integration. The practical implication is that foundation models can serve as powerful drop-in replacements for domain-specific models when latency and accuracy requirements are moderate, freeing teams to focus on data quality, feature engineering, and business-specific logic rather than model tuning. However, production teams must still validate that zero-shot performance meets their SLA thresholds (e.g., MAE or RMSE targets), and they may need fine-tuning if task-specific accuracy is critical; the approach trades off potential peak performance for engineering velocity and time-to-value.
Limitations and When Not to Use This
The paper does not address whether Chronos-2 fine-tuning could exceed current state-of-the-art, so it remains unclear if the foundation model approach is ceiling-limited or can be improved with domain-specific adaptation. Zero-shot performance on transportation data may not generalize to highly irregular or out-of-distribution scenarios (e.g., unprecedented congestion events, infrastructure changes, or sensor failures), since the model was pre-trained on historical patterns it may not have seen. The benchmark focuses on aggregate-level forecasting (traffic flow, bike demand) and does not evaluate fine-grained spatial or multi-step ahead prediction on dense sensor networks, which have different characteristics and requirements. Real production systems require not just accurate point forecasts but calibrated uncertainty quantification, anomaly detection, and cold-start handling for new routes or stations—aspects not thoroughly discussed in the abstract. The paper also does not discuss computational costs, latency, or whether inference can be run on edge devices, which are critical for real-time transportation systems operating at scale.
Research Context
This work builds on the recent trend of foundation models in time-series, following earlier models like TimeGPT and others that demonstrated that learning general temporal patterns improves generalization. It directly advances the evaluation methodology for time-series foundation models by establishing a multi-domain benchmark protocol inspired by how NLP and vision community evaluate foundation models. The paper positions Chronos-2 as an alternative to the conventional deep learning pipeline (custom architectures like LSTM, GRU, Transformer variants), reducing the reliance on task-specific engineering in time-series forecasting. This opens a research direction toward developing better evaluation frameworks and understanding when foundation models succeed or fail in specialized domains, laying groundwork for future work on efficient fine-tuning, domain adaptation, and uncertainty quantification in time-series.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::