A clothing company keeps its complete product sales history in Amazon S3 and currently uses custom exponential smoothing (ETS) models to forecast demand for existing products. The company needs a method to forecast demand for a new product variation that hasn't sold before. Which approach best meets this requirement?

DeepAR is ideal for forecasting new products because it can leverage data from similar existing products to make predictions, a technique known as "cold start" forecasting. It learns patterns across many related time series, which is crucial when a new product lacks historical sales data. Training another custom ETS model is unsuitable as ETS models require historical data for the specific product being forecasted. K-means clustering is for grouping similar data points, not for time-series forecasting. XGBoost is a powerful supervised learning algorithm, but it's not inherently designed for time-series forecasting, especially for cold-start scenarios where DeepAR excels by learning from a collection of series.