SIMBAD tests when traffic forecasts should trust the past

When familiar patterns become misleading

Traffic forecasting relies partly on repetition: similar commuting schedules can produce similar flows from one weekday to the next. But historical patterns can also mislead a model when current conditions depart from that routine. The SIMBAD study asks how a forecasting system might retain useful history without letting it override more relevant recent observations.

Daniel Y. Lee and colleagues describe their approach in a paper accepted to NeurIPS 2026, with the arXiv version posted on November 6, 2026. Their results concern predictions from recorded traffic datasets, not a deployed traffic-management system or demonstrated reductions in congestion.

Comparing the present with the past

SIMBAD represents a traffic network as a graph. Sensors are its nodes, while connections describe relationships through which information can pass. The model combines observations from the most recent hour with historical windows from one day, one week and two weeks earlier to predict the next 12 five-minute intervals—one hour ahead.

Its central comparison is deliberately simple. It adds the absolute differences between corresponding observations in two traffic sequences, a calculation called L1 distance. A smaller distance means the sequences look more alike.

However, a recent sequence resembling last week’s does not guarantee that the next hour will do so. SIMBAD therefore learns to estimate both the future mismatch with each historical reference and uncertainty in that estimate. During training, the observed future supplies the comparison target; when making forecasts, the model estimates it from available inputs.

Those estimates help combine historical windows and construct a numerical mask that limits their contribution at each sensor. Where historical information is judged less reliable, the architecture gives recent observations additional emphasis. This uncertainty concerns the reliability of historical similarity; it is not presented as a calibrated confidence interval around every traffic forecast.

The model also adjusts relationships between sensors. One component learns connections informed by the relevance and reliability of historical patterns. Another reweights existing connections according to how similar sensors’ recent observations are. Attention mechanisms and graph and temporal convolutions then process these regulated signals to generate predictions.

What the benchmarks showed

The main experiments used four traffic-flow benchmarks, PEMS03, PEMS04, PEMS07 and PEMS08, containing between 170 and 883 sensors. The researchers allocated 60% of the data to training, 20% to validation and 20% to testing, and compared SIMBAD with 11 forecasting methods.

In the main comparison table, SIMBAD recorded the lowest errors across all four datasets on three measures: mean absolute error, root mean squared error and mean absolute percentage error. These measure average error size, place greater emphasis on large mistakes, and express errors relative to observed values, respectively.

On PEMS07, for example, SIMBAD’s mean absolute error was 17.86, compared with 19.20 for the strongest baseline on that measure. On PEMS08, it was 12.58, compared with 13.50. These are reductions in prediction error, not percentage-point improvements in an overall accuracy score.

To examine irregular conditions, the authors grouped predictions according to how far the actual future traffic departed from its corresponding historical pattern. They report that SIMBAD retained an advantage over selected comparison methods as this mismatch increased. That supports robustness to their measured form of irregularity, but does not establish that the system can anticipate particular accidents or other events before they affect observations.

Removing the historical mask, uncertainty mechanism or dynamic connection weights worsened the main benchmark results. Removing weekly historical inputs also hurt performance more than removing daily inputs: the aim is selective use of history, not its abandonment.

Where the approach falls short

The supplementary experiments provide an important counterweight to the headline results. SIMBAD performed well on PEMS-BAY and a subset of LargeST, but did not lead on METR-LA. The authors found that recent historical similarity was less informative about future historical similarity there, weakening a key assumption behind their method.

Benefits were also uneven when the temporal regulation module was added to other models. Across the appendix’s 120 dataset-and-metric comparisons, it improved 77 results, rather than helping universally.

Input access complicates comparisons because SIMBAD uses additional historical observations. The authors addressed this with a supplementary experiment extending baselines’ training-data coverage. SIMBAD remained competitive, although Graph WaveNet achieved lower mean absolute and root mean squared errors on PEMS03 in that comparison.

Finally, a cheap similarity calculation does not make the whole architecture the fastest: several baselines had shorter reported runtimes. The paper’s strongest contribution is therefore a bounded one—evidence that checking the relevance of history can improve traffic forecasts, provided present-day similarity is genuinely informative about what comes next.