Your Model's MSE Is Lying to You: Part II

When making multi-step forecasts with autoregressive models, a common but critical mistake is to repeatedly use the model's predicted mean as the input for the next step. This deterministic approach incorrectly treats each prediction as a perfectly certain value, ignoring the very uncertainty the model was trained to capture. As a result, the forecast systematically underestimates total risk because it fails to account for how small errors accumulate and propagate over time. The correct method involves generating many possible future trajectories by sampling from the model's predicted distribution at each step, providing a much more honest and statistically sound representation of long-term uncertainty.