Time Series Foundation Models as Strong Baselines in Transportation Forecasting: A Large-Scale Benchmark Analysis

Javier Yanes-Pulido; Filipe Rodrigues

cs.LGFeb 27, 2026

Time Series Foundation Models as Strong Baselines in Transportation Forecasting: A Large-Scale Benchmark Analysis

Authors: Javier Yanes-Pulido, Filipe Rodrigues

Organizations: Technical University of Denmark, Kgs. Lyngby, Denmark

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.

Figures & tables

Dataset	Type	Range [unit]	Time Span	Granularity	Train/Val/Test	Horizons Evaluated	Source
PeMSD7(M)	Traffic Speeds	3–82.6 mph	01/04/2016 – 30/06/2016	5 mins	34/5/5 days	15, 30, 45 mins	[ 2 ]
Urban1	Traffic Speeds	0.5–110 km/h	01/04/2018 – 30/04/2018	5 mins	70/10/20 %	30, 45, 60 mins	[ 18 ]
NYC Citi Bike	Pickups/Dropoffs	0–108 trips	01/04/2016 – 30/06/2016	30 mins	63/14/14 days	1 to 12 steps	[ 19 ]
PeMSD4	Traffic Volumes	0–919 veh.	01/01/2018 – 28/02/2018	5 mins	60/20/20 %	1 to 12 steps	[ 20 ]
SZ-taxi	Traffic Speeds	0–86.4 km/h	01/01/2015 – 31/01/2015	15 mins	80/-/20 %	15, 30, 45, 60 mins	[ 4 ]
METR-LA	Traffic Speeds	0–70 mph	01/03/2012 – 30/06/2012	5 mins	70/10/20 %	15, 30, 60 mins	[ 3 ]

TABLE I: Summary of Transportation Benchmark Datasets and Evaluated Horizons.

	MAE	MAPE	RMSE
Model	15/ 30/ 45 min	15/ 30/ 45 min	15/ 30/ 45 min
LSVR*	2.50/ 3.63/ 4.54	5.81/ 8.88/ 11.50	4.55/ 6.67/ 8.28
FNN*	2.74/ 4.02/ 5.04	6.38/ 9.72/ 12.38	4.75/ 6.98/ 8.58
FC-LSTM*	3.57/ 3.94/ 4.16	8.60/ 9.55/ 10.10	6.20/ 7.03/ 7.51
GCGRU*	2.37/ 3.31/ 4.01	5.54/ 8.06/ 9.99	4.21/ 5.96/ 7.13
STGCN(Cheb)	2.25 / 3.03/ 3.57	5.26/ 7.33/ 8.69	4.04/ 5.70/ 6.77

TABLE II: RESULTS FOR PEMSD7(M) - TRAFFIC SPEEDS IN CALIFORNIA.

	MAE	MAPE	RMSE
Model	30/ 45/ 60 min	30/ 45/ 60 min	30/ 45/ 60 min
VAR*	5.06/ 4.99/ 4.97	23.10/ 22.82/ 22.73	7.04/ 6.92/ 6.88
LSVR*	3.82/ 3.89/ 3.93	15.35/ 17.99/ 17.39	5.64/ 5.74/ 5.84
ARIMA*	3.49/ 3.79/ 4.04	15.40/ 16.85/ 18.09	5.28/ 5.65/ 5.94
FC-LSTM*	3.91/ 3.92/ 3.92	17.29/ 17.32/ 17.31	6.38/ 6.39/ 6.39
DCRNN*	3.17/ 3.46/ 3.73	13.52/ 14.83/ 15.95	4.94/ 5.30/ 5.61

TABLE III: RESULTS FOR URBAN1 - TRAFFIC SPEEDS IN SOUTH KOREA.

Model	MAE	RMSE
XGBoost*	2.469	4.050
FC-LSTM*	2.303	3.814
DCRNN*	1.895	3.209
STGCN*	2.761	3.604
STG2Seq*	2.498	3.984
Graph WaveNet*	1.991	3.294

TABLE IV: RESULTS FOR NYC CITI BIKE - PICKUPS AND DROPOFFS IN NYC.

Model	MAE	MAPE	RMSE
ARIMA*	33.73	24.18	48.80
VAR*	24.54	17.24	38.61
FC-LSTM*	26.77	18.23	40.65
TCN*	23.22	15.59	37.26
GRU-ED*	23.68	16.44	39.27
DSANet*	22.79	16.03	35.77

TABLE V: RESULTS FOR PEMSD4 - TRAFFIC VOLUMES IN CALIFORNIA.

	MAE	RMSE
Model	15/ 30/ 45/ 60 min	15/ 30/ 45/ 60 min
T-GCN	4.517/ 4.572/ 4.621/ 4.671	5.997/ 6.034/ 6.064/ 6.100
HA	4.630	6.463
HA+LR	3.464/ 3.507/ 3.534/ 3.554	4.998/ 5.057/ 5.091/ 5.115
Chronos-2	2.463 / 2.468 / 2.470 / 2.472	3.912 / 3.919 / 3.925 / 3.929

TABLE VI: RESULTS FOR SZ-TAXI - TRAFFIC SPEEDS IN SHENZHEN FROM TAXI TRAJECTORIES.

	MAE	MAPE	RMSE
Model	15/ 30/ 60 min	15/ 30/ 60 min	15/ 30/ 60 min
ARIMA*	3.99/ 5.15/ 6.90	9.6/ 12.7/ 17.4	8.21/ 10.45/ 13.23
VAR*	4.42/ 5.41/ 6.52	10.2/ 12.7/ 15.8	7.89/ 9.13/ 10.11
SVR*	3.99/ 5.05/ 6.72	9.3/ 12.1/ 16.7	8.45/ 10.87/ 13.76
FNN*	3.99/ 4.23/ 4.49	9.9/ 12.9/ 14.0	7.94/ 8.17/ 8.69
FC-LSTM*	3.44/ 3.77/ 4.37	9.6/ 10.9/ 13.2	6.30/ 7.23/ 8.69

TABLE VII: RESULTS FOR METR-LA - TRAFFIC SPEEDS IN LA.

	MAE	MAPE	RMSE
Model	15/ 30/ 60 min	15/ 30/ 60 min	15/ 30/ 60 min
ARIMA*	1.62/ 2.33/ 3.38	3.5/ 5.4/ 8.3	3.30/ 4.76/ 6.50
VAR*	1.74/ 2.32/ 2.93	3.6/ 5.0/ 6.5	3.16/ 4.25/ 5.44
SVR*	1.85/ 2.48/ 3.28	3.8/ 5.5/ 8.0	3.59/ 5.18/ 7.08
FNN*	2.20/ 2.30/ 2.46	5.2/ 5.4/ 5.9	4.42/ 4.63/ 4.98
FC-LSTM*	2.05/ 2.20/ 2.37	4.8/ 5.2/ 5.7	4.19/ 4.55/ 4.96

TABLE VIII: RESULTS FOR PEMS-BAY - TRAFFIC SPEEDS IN CALIFORNIA.

	MAE	RMSE
Model	1h/ 2h/ 3h	1h/ 2h/ 3h
ARIMA*	10.41/ 11.84/ 13.00	19.14/ 21.76/ 23.90
STGCN*	6.49/ 7.06/ 7.94	11.73/ 12.93/ 15.37
DCRNN*	5.88/ 6.19/ 7.72	9.85/ 10.39/ 12.37
STGNN*	5.79/ 6.00/ 7.56	9.80/ 9.98/ 11.91
MVGCN*	5.65/ 7.72/ 8.00	9.64/ 13.53/ 13.93

TABLE IX: RESULTS FOR NYC BIKE IN- AND OUT-FLOWS.

	MAPE	RMSE
BTMF*	7.70	4.59
TRMF-GRMF*	8.20	4.83
TRMF-ALS*	8.36	4.96
Linear-LSTM-ReMF*	8.01	4.59
BiLSTM-GL-ReMF	7.83	4.50
LSTM-ReMF	7.64	4.42

TABLE X: RESULTS FOR SEATTLE TRAFFIC SPEED DATA.

	RMSE ( $\times 10^{-2}$ )	MAPE (%)	RAE ( $\times 10^{-2}$ )	MAE ( $\times 10^{-2}$ )
Model	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)
LO	8.32/ 10.6/ 12.38/ 13.06 (11.09)	24.83/ 37.55/ 48.57/ 53.89 (41.21)	33.31/ 49.02/ 61.69/ 67.25 (52.82)	4.04/ 5.97/ 7.54/ 8.22 (6.44)
AR	10.39/ 10.73/ 11.0/ 11.04 (10.79)	57.64/ 58.85/ 59.72/ 59.91 (59.03)	59.08/ 60.78/ 61.9/ 62.02 (60.94)	7.18/ 7.4/ 7.57/ 7.58 (7.43)
ARIMA	10.32/ 10.94/ 11.48/ 11.67 (11.1)	52.18/ 54.72/ 56.97/ 57.88 (55.44)	56.55/ 60.09/ 62.98/ 63.97 (60.89)	6.86/ 7.31/ 7.69/ 7.81 (7.42)
FCNN	7.86/ 8.95/ 9.16/ 8.23 (8.55)	27.36/ 37.08/ 39.46/ 35.21 (34.78)	35.02/ 45.0/ 46.93/ 41.07 (42.01)	4.25/ 5.49/ 5.74/ 5.03 (5.13)
LSTM	8.15/ 9.43/ 9.58/ 8.42 (8.9)	28.03/ 39.45/ 40.71/ 35.69 (35.97)	36.9/ 48.51/ 49.85/ 42.36 (44.4)	4.48/ 5.93/ 6.11/ 5.19 (5.43)

TABLE XI: RESULTS FOR URBANEV - EV STATION OCCUPANCY IN SHENZHEN.

	RMSE	MAPE (%)	RAE	MAE
Model	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)
LO	5.80 / 8.53/ 10.66/ 11.45 (9.11)	1.24/ 2.23/ 3.20/ 3.78 (2.61)	0.25/ 0.39/ 0.51/ 0.56 (0.43)	2.63/ 4.22/ 5.57/ 6.12 (4.64)
TimeXer	4.58/ 5.23/ 5.75/ 4.94 (5.13)	1.47/ 2.10/ 2.62/ 2.40 (2.15)	0.21/ 0.27/ 0.31/ 0.27 (0.26)	2.25/ 2.82/ 3.18/ 2.81 (2.76)
Chronos-2	1.55 / 1.55 / 1.55 / 1.55 ( 1.55 )	0.45 / 0.45 / 0.45 / 0.45 ( 0.45 )	0.08 / 0.08 / 0.08 / 0.08 ( 0.08 )	0.82 / 0.82 / 0.82 / 0.81 ( 0.82 )

TABLE XII: RESULTS FOR URBANEV - EV STATION CHARGING SESSION DURATION IN SHENZHEN.

	RMSE	MAPE (%)	RAE	MAE
Model	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)
LO	489.15/ 556.71/ 665.07/ 646.01 (589.24)	18.95/ 29.77/ 46.44/ 54.76 (37.48)	0.32/ 0.39/ 0.48/ 0.48 (0.42)	102.00/ 127.18/ 157.37/ 157.59 (136.04)
TimeXer	353.63/ 376.01/ 370.44/ 299.58 (349.92)	25.91/ 32.50/ 42.66/ 35.53 (34.15)	0.25/ 0.28/ 0.29/ 0.24 (0.27)	80.96/ 89.81/ 93.86/ 78.10 (85.68)
Chronos-2	140.85 / 140.95 / 140.77 / 140.87 ( 140.86 )	8.87 / 8.96 / 9.07 / 9.20 ( 9.03 )	0.10 / 0.10 / 0.10 / 0.10 ( 0.10 )	31.65 / 31.66 / 31.62 / 31.60 ( 31.63 )

TABLE XIII: RESULTS FOR URBANEV - EV STATION CHARGING VOLUME IN SHENZHEN.

Dataset	Rank	Improvement vs. HA (%)			Improvement vs. Best DL (%)
Dataset	Rank	MAE	MAPE	RMSE	MAE	MAPE	RMSE
PeMSD7(M)	1	+25.6	+27.9	+24.3	+18.8	+15.9	+12.7
Urban1	1	+11.0	+10.4	+12.9	+5.4	+5.9	+6.7
NYC Citi Bike	1	+8.8	-	+22.0	+9.5	-	+21.1
PeMSD4	1	+28.3	+20.5	+34.0	+2.0	-6.3	+9.0
SZ-taxi	1	+46.6	-	+39.2	+47.1	-	+35.6

TABLE XIV: Summary of Chronos-2 performance across benchmarks. Values denote the average percentage reduction in forecasting error relative to the statistical baseline (HA) and the strongest deep learning (DL) competitor. When multiple horizons are reported without averaging, the longest one is used. Rank denotes the rounded average rank across available metrics.

	Coverage (% , target 80%)	IQR ( $\times 10^{-2}$ )
Model	3/ 6/ 9/ 12 h (Avg)	3/ 6/ 9/ 12 h (Avg)
AR	84.84/ 84.76/ 84.82/ 84.89 (84.83)	26.29/ 27.15/ 27.84/ 27.96 (27.31)
ARIMA	86.11/ 85.90/ 85.65/ 85.60 (85.82)	26.22/ 27.83/ 29.21/ 29.69 (28.24)
FCNN	84.20/ 84.60/ 84.49/ 85.04 (84.58)	18.48/ 23.21/ 24.69/ 22.78 (22.29)
LSTM	83.18/ 83.75/ 84.85/ 84.72 (84.12)	16.15/ 21.30/ 22.76/ 19.48 (19.92)
GCN	74.07/ 72.85/ 72.75/ 75.61 (73.82)	12.12/ 14.16/ 15.98/ 16.19 (14.61)

TABLE XV: RESULTS FOR URBANEV (PROBABILISTIC METRICS) - EV STATION OCCUPANCY IN SHENZHEN.

Dataset	Coverage (%)	IQR
PeMSD7(M)	77.69	7.54
Urban1	79.79	8.65
NYC Citi Bike	63.87	4.90
PeMSD4	77.12	57.93
SZ-taxi	63.42	7.41
METR-LA	71.46	12.26

TABLE XVI: PROBABILISTIC EVALUATION OF CHRONOS-2: CALIBRATION (EMPIRICAL COVERAGE) AND SHARPNESS (IQR), AVERAGED ACROSS HORIZONS FOR AN 80% NOMINAL INTERVAL.

Time-Series Foundation Models

Time-series foundation models are pretrained sequence models designed to transfer across forecasting, classification, anomaly detection, and other temporal tasks.

See all 287 papers

Do Time Series Foundation Model Benchmarks Hide Regime-Dependent Failures? Evidence from Traffic Speed Forecasting
Jun 16, 2026Yingshuo Wang, Xian Sun, Lingdong Kong +4Time Series Forecasting Forecasting Benchmarks
University of California, Berkeley, CA, USA · Duke University, Durham, NC, USA · National University of Singapore, Singapore +3
Foundation Models and Fine-Tuning: Toward a New Generation of Models for Time Series Forecasting
Jul 25, 2026Morad Laglil, Bertrand Pracca, Emilie Devijver +1Zero-Shot Time Series Forecasting Time-Series Foundation Models
Savoye, 18 bd des Gorgets, 21000, Dijon, France · Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, Grenoble, 38000, France
How Good Are Time-Series Foundation Models for Pedestrian Crowd Count Forecasting? A Cross-Dataset Comparative Study
Sep 14, 2026Theivaprakasham Hari, Ziteng Li, Yanan Xin +2Crowd Counting Time Series Forecasting
Department of Transport & Planning, Faculty of Civil Engineering and Geosciences, Delft University of Technology, 2628 CN Delft, The Netherlands

Jun 16, 2026J/K move · Enter open · S save

Do Time Series Foundation Model Benchmarks Hide Regime-Dependent Failures? Evidence from Traffic Speed Forecasting

Yingshuo Wang, Xian Sun, Lingdong Kong, Wei Gao+3

University of California, Berkeley, CA, USA · Duke University, Durham, NC, USA · National University of Singapore, Singapore+3

Time Series Forecasting Forecasting Benchmarks Traffic Forecasting Time-Series Foundation Models Probabilistic Time Series Forecasting Regression Prediction Intervals

Open paper AskPDF arXiv

Standard benchmarks evaluate time series foundation models (TSFMs) using aggregate metrics, but these can mask severe failures in critical operating regimes. We introduce regime-stratified evaluation and apply it to three TSFMs on two standard traffic speed benchmarks. Traffic exhibits abrupt regime switching between free-flow and congested states, producing bimodal speed distributions during transitions. When we stratify by traffic regime, both accuracy and prediction-interval coverage degrade sharply during transitions: transition-regime MAE reaches 11 mph (versus 3 mph overall), and empirical coverage of 90% prediction intervals drops as low as 55%. These failures are invisible in aggregate metrics because free-flow observations dominate the sample. A simple historical conditional baseline (sampling from per-sensor training distributions) achieves better transition coverage than any TSFM, but has far worse overall accuracy. We propose bimodal mixture augmentation (BMA), a post-hoc method that combines TSFM forecasts with historical distributional knowledge, approaching the historical baseline's transition coverage while preserving the TSFM's accuracy. Our results suggest that TSFM benchmarks should incorporate regime-aware evaluation to surface failures that aggregate metrics hide.

Abstract

Time-Series Foundation Models

Explore similar work

Do Time Series Foundation Model Benchmarks Hide Regime-Dependent Failures? Evidence from Traffic Speed Forecasting

Foundation Models and Fine-Tuning: Toward a New Generation of Models for Time Series Forecasting

How Good Are Time-Series Foundation Models for Pedestrian Crowd Count Forecasting? A Cross-Dataset Comparative Study