Time series differentiation¶

Time series differentiation involves computing the differences between consecutive observations in the time series. When it comes to training forecasting models, differentiation offers the advantage of focusing on relative rates of change rather than directly attempting to model the absolute values. Once the predictions have been estimated, this transformation can be easily reversed to restore the values to their original scale.

This methodology is very useful when dealing with time series that exhibit trends, as not all models can capture such trends properly. Among the many machine learning techniques, tree-based models, including decision trees, random forests, and gradient boosting machines (GBMs), stand out for their effectiveness and broad applicability to various domains. Nonetheless, these models are limited in their ability to extrapolate. Their inability to project values outside the observed range during training inevitably results in predicted values that deviate from the underlying trend.

Skforecast, version 0.10.0 or higher, introduces a novel differentiation parameter within its forecasters. This parameter indicates that a differentiation process must be applied before training the model, and this task is performed through the internal use of a new class named skforecast.preprocessing.TimeSeriesDifferentiator. It is important to note that the entire differentiation process is automated and its effects are seamlessly reversed during the prediction phase. This ensures that the resulting forecast values are in the original scale of the time series data.