First, because time series forecasting aims to predict the future based on historical data, it usually has a high degree of uncertainty. Unlike other machine learning problems, the test set may differ from the training and validation sets drawn from historical data. Second, real-world time series data often suffer from missing and intermittently high data (such as when most of the time series values are 0). Some time-series tasks may not have available historical data and have cold-start problems, such as forecasting new product sales. Finally, time-series forecasts vary significantly across domains (product sales, web traffic, etc.), granularity (daily, hourly, etc.), historical length, and feature types (Categorical, Numeric, DateTime, etc.).