A new traffic-forecasting model reported the lowest errors among the methods tested, even after researchers randomly removed 80% of the input entries. The model also had the shortest reported runtimes in the missing-data tests, according to an arXiv preprint dated 26 August 2026.
With complete observations, the model had the lowest reported mean absolute percentage error, or MAPE, and root mean square error, or RMSE, across both datasets and all three traffic measures. On the PeMS-D8 dataset, its volume MAPE was 11.20, compared with 21.65 for the best competing result. For occupancy, the figures were 13.19 and 25.77.
Built for several traffic signals at once
The model, called MVCTD, combines traffic measures in a coupled tensor decomposition. In practical terms, it is designed to learn shared patterns in where traffic is measured while allowing the different measures to retain their own time patterns. It also includes group-sparse regularization for correlated anomalies.
The system is intended for online forecasting, meaning it can update its current estimate as new traffic information arrives. Its online stage refines only the current latent tensor and updates the other model variables through closed-form steps that use summarized historical information. The initial model is fitted offline with a batch alternating direction method of multipliers, or ADMM, procedure.
The evaluation used two Caltrans Performance Measurement System datasets. PeMS-D4 covered 307 loop detectors over 59 days, while PeMS-D8 covered 170 sensors over 62 days. Both recorded volume, occupancy and speed at five-minute resolution.
A strong result under simulated sensor loss
Each rolling experiment used 15 days of historical data to make predictions over a five-day online horizon. To test missing observations, the researchers randomly masked 80% of the input entries, treating the gaps as a simulation of sensor failures.
Under that severe masking level, MVCTD ranked first on every reported metric. Relative to the fully observed case, its volume MAPE increased by 1.06 on PeMS-D4 and 0.95 on PeMS-D8. Occupancy MAPE rose by less than one percentage point on both datasets.
The reported runtimes were 101 seconds on PeMS-D4 and 49 seconds on PeMS-D8 with complete observations. With 80% of entries missing, they fell to 72 and 38 seconds, respectively. The paper reports that this was about 26 times faster than the fastest comparison on PeMS-D4 and more than 33 times faster on PeMS-D8.
The missing-data comparison has an important boundary. The paper says that circDMDsp, PatchTST, PreSTGNet and AmTGBiM could not directly process incomplete inputs, so they were used as reference baselines only in the fully observed tests.
Which parts of the design mattered most
An ablation test, which removes parts of a model to see how its reported results change, showed the largest difference when the anomaly component was removed. With 40% missing data on PeMS-D8, the no-anomaly version had a volume MAPE of 22.94, compared with 11.31 for the proposed model. Versions without global periodicity or local smoothness also showed weaker accuracy across views.
The results also favored modeling several traffic views together, although the gains were not identical everywhere. Against an independently trained single-view version, the coupled model had a speed MAPE of 4.49 versus 4.67 on PeMS-D8, and a volume MAPE of 15.48 versus 15.92 on PeMS-D4. On PeMS-D4, it had lower MAPE for all three views but slightly higher volume RMSE.
The choice of time lags mattered as well. Using only the immediate previous day was the weakest option, while a continuous window covering lags one through seven reduced errors. The proposed sparse lag structure performed better on most reported metrics than that continuous window, despite slightly higher volume RMSE.
In the PeMS-D8 convergence analysis, the solver's error fluctuated during its first 20 iterations before falling to the stopping criterion within 90 iterations across the missing-rate cases shown.
What the tests leave open
The study's tests covered two PeMS datasets, three traffic measures and a five-day forecasting horizon. The severe missingness scenario used random masking, and the incomplete-input comparison did not include every baseline. The paper is an arXiv version 1 preprint dated 26 August 2026.
The paper reports support from China's National Key R&D Program, the National Natural Science Foundation, the Hubei Provincial Natural Science Foundation and the Fundamental Research Funds for the Central Universities. The listed grant identifiers are 2021YFA1000300, 2021YFA1000301, 12171189, 12021001, 92473208, 2025AFB966 and XJ2026000601.
Paper data and sources
Original title: A Multi-View Coupled Tensor Decomposition for Lightweight Online Adaptive Traffic Prediction
Authors: Quan Yu, Jie Ni, Yu-Hong Dai, Xiongjun Zhang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text