Forecasting concepts
Understand targets, horizons, zero-shot transfer, covariates, and uncertainty before comparing models.
History, cutoff, cadence, and horizon
The history ends at the forecast origin. The horizon is the number of future steps, and cadence defines how far apart those steps are. A 24-step horizon can mean one day of hourly data or 24 days of daily data. Keep timestamps and units explicit even when a model accepts only a numerical array.
Zero-shot and adaptation
Zero-shot forecasting applies pretrained weights to a new series without fitting those weights on that series. Fine-tuning changes model parameters using task data. Covariates add related information and must respect the issue-time boundary; known-future covariates differ from future values that are only available after the event.
Point forecasts and uncertainty
A mean and a median answer different questions. Quantiles describe ordered thresholds of a predictive distribution, but their nominal labels do not establish empirical calibration. Evaluate coverage and quantile loss across a relevant held-out sample, especially when decisions depend on tail risk.