A probabilistic model represents uncertainty by assigning a conditional distribution to possible outcomes. In forecasting, it maps an input sequence to a distribution of future values rather than one fixed prediction. section 4.1 section 6.1
The distribution supports prediction intervals: choose a desired interval after obtaining possible future values. A deterministic model instead returns a single value for each forecast step. section 4.1
Forecasting models condition their output on past observations and may use exogenous features. Those features must be known, or separately forecast, across the future horizon. section 3.2
(3.1)
Use the conditional-probability formulation to trace how past values and future exogenous features determine a future sequence distribution. section 3.2
Implementations vary. Lag-Llama assumes a Student’s t-distribution and generates intervals from sampled predictions. Chronos scales and quantizes continuous series into tokens, applies a language model, then reverses the transformation. section 4.1.2 section 5.2
Figure 5.2 Overview of the Chronos framework, which adapts a language model for time-series forecasting. First, the time series is tokenized, which involves mean scaling and quantization. Then the output can be used to train the language model.
Follow the tokenization pipeline from mean scaling and quantization to language-model prediction and the reversal that recovers forecast values. section 5.2
Reducing a distribution to a point forecast
Moirai produces forecast samples for each time step. When a scenario requires point forecasts, stack the samples and use their median; retain the distribution when uncertainty matters. section 10.2.6




