How many people come through your door next week?
Pick a location and read next week in one sentence. Then go back in time to see how right the forecast was, compare four methods, or try your own numbers. Everything is computed in your browser.
The demo needs JavaScript. The model and two years of counts are about 700 KB, downloaded once.
Real data: daily visitors counted by street sensors in Melbourne (City of Melbourne, CC BY 4.0), the last two years. Eight locations are shown here; the results below cover all 65.
11.1%Average forecast error, four weeks ahead, at 65 locations on 12 dates after the training data ended. Copying last week's numbers: 14.3%.
83%Of real days inside our model's likely range, which aims for 80%.
600 KBThe whole model. It runs in this page, and on any server it refreshes every location every morning. No GPU, no API bill.
How it works
DataTwo years of daily visitor counts at 65 locations. Days with missing hours are gaps, not guesses. Your drawn series gets the same treatment.
MethodsSame as last week, exponential smoothing and our gradient-boosting model run on your device; the foundation model was run once on a GPU. All four were tested the same way: trained on data before May 2026, then asked to forecast twelve dates they never saw.
ResultsExponential smoothing: 11.9% error. Chronos-Bolt, a large ready-made forecasting model run on a GPU: 10.8%, practically as accurate as ours from a model too big for a browser. A four-week forecast takes 2 ms here.
What you getA forecast with a range for every location, every morning, in the tool you already use, and an honest number for how wrong it will be. No GPU, no API bill.
Code, data pipeline and the full benchmark: github.com/balticml/demand-forecasting. Want this on your orders, calls or visitors? Get in touch