Imputation by prediction identifies other data sources that are related to the data that contains missing values, and then uses DDM's own forecasting component to learn to predict the missing values from those other data sources, as though the missing values are part of the test set in a train-test forecasting scenario.

This imputation method is the preferred one if there are other data sources which are sufficiently related to the data with missing values, if the data with missing values can change rapidly (in a non-smooth way), and if there are large numbers of subsequent missing values in the time series of the data.