One-step-ahead prediction is an evaluation strategy for sequence models where the model predicts the next observation $$x_t$$ using only the actual ground-truth historical observations $$x_{t-1}, \ldots, x_{t-\tau}$$ as inputs. Since the model is always supplied with real historical data rather than its own past predictions, this strategy evaluates the model's immediate next-step accuracy without suffering from error accumulation over time.

Google

A standard linear regression model can serve as a simple baseline for sequence prediction by treating the data as if it satisfies a $$\tau^{\textrm{th}}$$-order Markov condition. The model accepts a fixed-length window of the past $$\tau$$ observations as its input feature vector and predicts the next value in the sequence. The model is trained using standard optimization (such as gradient descent) to minimize a regression loss like mean squared error. While simple, this approach establishes a baseline performance level using a linear combination of past observations before employing more complex sequence models.

Learn Before

Related