Apppublic

NicolaQua/timesfm-3-forecasting

sourceHugging Faceotherupdated 24d agoView on Hugging Face

0likes

App README

TimesFM 3.0 — zero-shot time-series forecasting

Status: awaiting GPU hardware. The Gradio app in this repo (app.py) is complete and targets ZeroGPU. The Space is currently published with sdk: static — a holding page — because running Gradio on Hugging Face requires PRO or a community GPU grant. Once hardware is granted, see `DEPLOY.md`: flipping four lines of frontmatter makes the live app run, with no code changes.

An interactive demo of `google/timesfm-3.0-pytorch`, Google Research's time-series foundation model. Hand it history, get back a distribution over the future — no training, no fitting, no per-dataset tuning.

What the demo does

Forecast a series — pick a bundled dataset or upload your own CSV, choose how much history to feed the model and how far ahead to predict. The plot shows the median forecast with the 80% (q0.1–q0.9) and 40% (q0.3–q0.7) quantile bands.

Leave backtest on and the last horizon observations are held out rather than shown to the model, so the forecast is scored against ground truth: MAE, RMSE, sMAPE, the ratio against a naive baseline, and how much of the truth actually landed inside the 80% interval. Turn it off to forecast past the end of the data.

Multivariate + covariates — TimesFM 3.0's headline feature. Select several correlated columns and they are forecast jointly, with attention across variates, optionally conditioned on past-only covariates (observed over history) and past-and-future covariates (known over the forecast window too). This tab always backtests, so the future covariates are real observations rather than guesses.

Every forecast is downloadable as a CSV with all nine quantiles.

Bundled datasets

Dataset	Shape
Airline passengers, monthly 1949–1960	multiplicative yearly seasonality on a trend
Melbourne minimum temperature, daily 1981–1990	smooth annual cycle, noisy day to day
Monthly sunspots, 1749–1983	the irregular ~11-year solar cycle
Female births in California, daily 1959	near-stationary counts, very little structure
ETTh1 electricity transformer, hourly	7 correlated variates, daily + weekly cycles

API and MCP

The Space is also callable as an API and as an MCP server (mcp_server=True), with /forecast and /forecast_multivariate endpoints returning JSON — history, median forecast, all nine quantiles, held-out actuals, and metrics.

python

from gradio_client import Client

client = Client("<namespace>/timesfm-3-forecasting")
print(client.view_api())

Licensing

The demo code is MIT. The model weights are distributed under the TimesFM Non-Commercial License v1.0 — non-commercial, non-production use only.

Bundled example datasets are long-standing public benchmark series (Box & Jenkins airline passengers, Melbourne temperatures, monthly sunspots, Newcastle daily female births, and ETDataset ETTh1).