Zero-shot forecasting uses a pretrained model to predict a target time series without training or fine-tuning on that series. It transfers learned forecasting behavior directly to an unseen dataset, making inference faster than fitting a data-specific model. section 2.11 section 2.5
A practical zero-shot workflow is:
- Load a pretrained forecasting model and format the data with the model’s required identifiers, time column, and target column. section 2.5
- Reserve a forecast horizon when evaluating, then pass only the historical input to the prediction method. section 2.5
- Extract point forecasts or prediction intervals according to the model’s output format. section 6.3.1
- Compare predictions with known observations using metrics such as MAE and sMAPE, preferably through cross-validation. section 3.5.4 section 6.3.1
Pretrained N-BEATS on an unseen series
Load the pretrained model, read monthly Australian antidiabetic-drug prescriptions, reserve the final 12 observations, and call predict on the earlier observations. No model training occurs on this dataset. section 2.5
Probabilistic zero-shot forecasts
Moirai generates 100 samples for each future time step across four Walmart stores. Their medians become point forecasts, while percentiles form a 95% prediction interval; unseen future values still require cross-validation for evaluation. section 6.3.1
Zero-shot forecasting is a strong baseline and a rapid proof of concept, not a universal replacement for fine-tuning or data-specific methods. Performance varies by model, frequency, horizon, hardware, and data characteristics. Forecasted exogenous variables can introduce additional error, and some models impose operational or calibration limitations. section 10.3 section 3.5.4 section 7.2
