multimodalart's picture
multimodalart HF Staff
Upload folder using huggingface_hub
4ea7a41 verified
|
Raw History Blame Contribute Delete
2.12 kB
metadata
title: FlexCast Forecasting
emoji: 📈
colorFrom: yellow
colorTo: indigo
sdk: gradio
sdk_version: 6.32.0
app_file: app.py
python_version: '3.12'
short_description: Flexible-context probabilistic time series forecasting

FlexCast: Flexible Context Learning for Time Series Foundation Models (paper · code · checkpoint)

Interactive demo of FlexCast's zero-shot forecasting: upload a CSV time series, choose the context length (any multiple of 16, up to 3072 steps) and the forecast horizon (96 / 192 / 336 / 720 steps). FlexCast treats forecasting as variable-length masked patch prediction, so any context length and any horizon are handled in a single forward pass — no autoregression, no retraining — producing a full 99-quantile predictive distribution for every channel, decoded monotonically and mapped back to the input scale with RevIN.

If the uploaded series contains at least context + horizon rows, the trailing rows are held out and the forecast is scored against them (MAE, RMSE, 90% interval coverage, mean interval width) so the calibration of the predictive band can be judged honestly.

The models/, layers/ and utils/ packages are the reference implementation from the official repository (MIT license); inference follows exp/exp_inference.py exactly (masked tail, NaN padding/missing masks, bf16 autocast, weighted-quantile point forecast).

Example data

The bundled examples use ETTh1 from ETDataset — the hourly electricity-transformer benchmark (7 channels + oil temperature) used in the FlexCast paper, originally introduced in Informer (AAAI 2021). Dataset license: CC-BY-ND 4.0; redistributed verbatim in examples/ with attribution.

Credits