walidhadri/timeseries-forecasting-playground

Time Series Forecasting Playground

Gradio-based Hugging Face Space for uploading time series data, generating forecasts with foundation models, comparing outputs, and evaluating optional ground truth.

Current scope

• —Phase 1 scaffold implemented
• —Toto integration (Datadog/Toto-Open-Base-1.0)
• —Chronos-2 integration (amazon/chronos-2, optional dependency)
• —Statistical baselines: Naive, Seasonal Naive, Mean, Drift, Moving Average
• —Classical models via statsforecast: AutoARIMA, AutoETS, Theta, Holt-Winters
• —Additional foundation model wrappers: TiRex, Moirai-2.0, Sundial-Base
• —Model selection grouped in UI: Foundation Models and Statistical Models
• —CSV/Excel upload, auto timestamp + target detection
• —Quick test datasets in UI: Airline Passengers, ETTh1, and sample datasets
• —Forecast plotting + model comparison plot
• —Optional evaluation metrics and residual analysis

Local run

pip install -r requirements.txt
python app.py

Notes

• —Toto model loading requires toto-ts and adequate compute.
• —Statistical models require statsforecast.
• —TiRex requires tirex-ts.
• —Some foundation models may require additional dependencies/checkpoint downloads at runtime.
• —If GPU is available, Toto and Chronos-2 use CUDA automatically.