TBATS

TBATS (Trigonometric seasonality, Box-Cox transformation, ARMA errors, Trend, Seasonal) is designed for time series with multiple overlapping seasonal periods (e.g., daily, weekly, and yearly seasonality simultaneously).

When to use:

• Complex seasonality that SARIMA cannot handle (e.g., hourly data with daily + weekly + annual patterns)
• Retail, energy, or transportation data with multiple periodic patterns
• When seasonality periods are non-integer or very long

Input:

• Trained model checkpoint — exported TBATS fit
• Preprocessing config — transformation settings
• Training tail — last N observations
• Steps — forecast horizon

Output: Forecasted values for the specified horizon

Model Settings (set during training, used at inference)

TBATS automatically selects its own seasonal periods and model structure during training. Key parameters set during training:

• Seasonal periods — detected or specified periods (e.g., [7, 365.25])
• Box-Cox transformation — whether variance stabilization was applied
• Trend damping — whether trend is damped to prevent long-term drift

Inference Settings

No dedicated inference-time settings.