Probabilistic Forecasting

Uncertainty quantification in Twiga covers parametric distribution heads, quantile regression, conformal prediction intervals and forecast uncertainty visualisation.

Neural models follow a strict backbone and head contract: every architecture exposes an encode() method producing a latent vector consumed by interchangeable distribution heads — Normal, Laplace, LogNormal, Gamma, Beta, StudentT, QR and FPQR. Any backbone pairs with any head without modifying backbone code, giving you a full matrix of probabilistic models from a small set of building blocks. Adding a new architecture or distribution head is a focused, isolated change rather than a fork of an entire model class.

• Distribution Families
• Quantile Regression
• Conformal Prediction & Uncertainty Quantification