Rausda6/TiRex-demo

TiRex – Zero‑Shot Time Series Forecasting App

A Gradio‑based interactive web app to perform zero‑shot time series forecasting using the TiRex model. Upload your own CSV/XLSX/Parquet files or choose from built‑in presets, filter series by name, and visualize quantile forecasts over your chosen horizon.

🔍 Features

• —Zero‑Shot Forecasting: Powered by the `NX-AI/TiRex` model.
• —Custom Data Upload: Accepts CSV, XLSX, and Parquet.
• —Preset Datasets: Includes loop.csv, air_passangers.csv, and ett2.csv for quick demos.
• —Interactive Filtering: Search, check/uncheck, and plot only the series you care about.
• —Quantile Forecasts: Displays historical data, median forecast line, and 10–90% quantile shading.
• —Configurable Horizon: Slider to set forecast length (1–512 steps).
• —Automatic Defaults: Detects best forecast‐length defaults for presets.

📊 Data Format

With Named Series

AAPL,120.5,121.0,119.8,122.1,123.5,...
AMZN,3300.0,3310.5,3295.2,3305.8,3315.1,...
GOOGL,2800.1,2795.3,2810.7,2805.2,2820.4,...

Without Named Series

120.5,121.0,119.8,122.1,123.5,...
3300.0,3310.5,3295.2,3305.8,3315.1,...
2800.1,2795.3,2810.7,2805.2,2820.4,...

Key Rules:

• —One row per time series
• —Consistent naming: Either all rows have names (first column) or none do
• —Numeric data: All values after the optional name column must be numeric
• —Minimum length: Time series must have at least forecast_length + 10 data points
• —Maximum constraints: Up to 30 time series and 2048 time steps per series

🔧 Configuration

Forecast Length

• —Default: 64 steps
• —Range: 1-512 steps
• —Auto-adjustment: Preset datasets have optimized forecast lengths:
• —loop.csv and ett2.csv: 256 steps
• —air_passangers.csv: 48 steps

Model Settings

• —Device: CUDA (T4 GPU)
• —Quantiles: 10%, 50% (median), 90% prediction intervals

📈 Output Features

• —Historical data: Blue line showing input time series
• —Median forecast: Orange line for point predictions
• —Uncertainty bands: Gray shaded area showing 10%-90%