Introduced on day 37
Exposure bias
The mismatch between training on true history and generating from the model's own outputs, which lets early errors compound in multi-step forecasts or generation.
In detail
One-step teacher-forced validation hides it. Remedies include evaluating in free-running mode per horizon, training direct multi-horizon outputs, and training schemes that expose the model to its own predictions. Recursive forecasts from linear one-step models show the same compounding.