A concrete example of generating synthetic sequence data uses $$T = 1000$$ time steps. At each step $$t$$, the observed value is computed as $$x_t = \sin(0.01 \cdot t) + \epsilon_t$$, where $$\epsilon_t$$ is drawn from a standard normal distribution and scaled by $$0.2$$, giving the noise a standard deviation of $$0.2$$. The underlying sinusoidal signal provides a smooth, periodic ground truth, while the additive noise makes the prediction task nontrivial. Assuming a $$\tau^{\textrm{th}}$$-order Markov condition with $$\tau = 4$$, the dataset is split so that the first $$600$$ examples (out of $$T - \tau = 996$$) serve as training data, covering approximately one full period of the sine function, while the remaining examples are reserved for evaluation.

Claude

To evaluate sequence models before tackling complex, real-world datasets, practitioners often generate continuous-valued synthetic data. A common approach involves sampling from a known deterministic function, such as a trigonometric signal, and corrupting each observation with additive Gaussian noise. This produces a sequence whose underlying dynamics are fully understood, providing a controlled environment where a model's predictive accuracy can be measured against a known ground truth. The simplicity and reproducibility of such synthetic benchmarks make them an essential first step in validating the behavior of new sequence architectures.

Learn Before

Related