Case · Transportation
Pick a moment and a road sensor in real benchmark data — every PEMS-BAY sensor in San Jose / South Bay (PeMS District 4), every EXPY-TKY link on the Tokyo expressways, or the 207 METR-LA sensors on Los Angeles County freeways — forecast ahead, then compare with what actually happened.
Interactive demo · backtest
This is a backtest on real data: choose a dataset, a date, a time and a sensor; the model sees only the real data from the two days before the chosen moment and forecasts what comes next. Then reveal the actual speeds and, for comparison, the same time a week earlier (or the day before when last week has no data). Values are vehicle speeds, not traffic volumes.
This is a backtest on classic academic benchmark datasets (PEMS-BAY, EXPY-TKY, METR-LA), recorded years ago; the live mobility board further down this page shows current data.
Wed 3 May 2017, sensor 400665 (I-680 S): speed fell from 66 to 6 mph within the next 60 minutes.
Roads are drawn from the dataset’s own sensor positions, not from a map; each sensor colours its own stretch, one side per direction. Scroll or use +/− to zoom, drag to pan, click a stretch to forecast that sensor.
Sensor
Requesting the model forecast…
Data: Caltrans PeMS, as compiled in the PEMS-BAY benchmark (Li et al., ICLR 2018); 325 sensors, San Jose / South Bay (PeMS District 4), California, USA, January–June 2017, 5-minute speeds in mph. Forecasts are generated by the SequenxAI demo model and are not travel advice.
About this demo: times are local Pacific time (PST before 12 March 2017, PDT after). Each sensor and moment is forecast live by the SequenxAI demo model on first request and cached; if inference is unavailable no forecast is shown. A speed of 0 is treated as a missing reading. SequenxAI Traffic, our transportation model, is still in research preview and is not used here.
Live
Beta · Technical demoBeyond the backtest, SequenxAI issues forecasts automatically on live public mobility streams (Hong Kong SAR, China journey times and passenger traffic; Singapore taxis and car parks) and scores them once the actuals arrive. Technical demo, not travel advice.
Technical demo, not advice for any decision; see the Terms of Service.
Loading the board…
Transportation series are driven by holidays, weather, events, and incidents, with sharp regime shifts and localized congestion spillover. Classical statistical models and single thresholds often fail at peaks and anomalies, and rarely hand usable uncertainty to operations desks.
Co-creation & evaluation
Try the backtest demo; evaluation scope and delivery are set by written agreement.