webForecast

The from-scratch TimesFM 70M patched decoder, running in your browser via WebGPU. Instance-normalize → patch → forecast → denormalize, no server.

Sample data (click to load, ⭳ to download): electricity_hourly.csv⭳ solar_generation_hourly.csv⭳ network_traffic_5min.json⭳ web_visits.json⭳ temperature_daily.csv⭳ stock_price_daily.json⭳ retail_sales_weekly.csv⭳ co2_monthly.csv⭳ airline_passengers_monthly.csv⭳ sunspots_monthly.csv⭳ browse all on GitHub →
history forecast (q50) q10–q90

About the model

This runs the from-scratch TimesFM 70M model by Fareed Khan — the published PyTorch checkpoint, exported to ONNX and int8-quantized (~65 MB), executed client-side with onnxruntime-web on WebGPU. No server, no API.