Time Series

Time series inference generates forecasts for future time steps using a trained model. The model uses the learned seasonal patterns, trends, and lags to produce predictions for the specified forecast horizon.

Input required at inference time:

• Trained model checkpoint — exported from the training run
• Preprocessing config — normalization/scaling settings from training
• Training tail — the last N observations from training data (needed for lag-based features)

Available Models

• ARIMA – Autoregressive integrated moving average for stationary series
• SARIMA – ARIMA with seasonal components
• SARIMAX – SARIMA with external regressors
• Auto-ARIMA – Automatically selects ARIMA order parameters
• Prophet – Facebook's decomposable time series model with holiday support
• Vector Autoregression – Multivariate autoregressive model
• TBATS – Handles multiple seasonal periods with trigonometric seasonality
• Theta Method – Decomposition-based model competitive on many benchmarks
• Exponential Smoothing – Weighted averaging of past observations
• XGBoost Time Series – XGBoost trained on lag features
• LightGBM Time Series – LightGBM trained on lag features
• CatBoost Time Series – CatBoost trained on lag and calendar features