Drift

Drift is a Nowcasting continuous evaluation/deployment library. (also known as walk-forward evaluation)

It supports both univariate and (soon) multivariate time series. It is from the ground-up extensible and lightweight.

Avoid the mistakes people make with time series ML:

ignoring useful features otherwise available in production (value in T-1) accidentally using information that wouldn't otherwise be available at the time of training/evaluation (lookahead bias) It can train models without lookahead bias:

• with expanding window
• with rolling window
• even with a single train/test split, if you really want it It can also help you with creating complex blended models:

Ensembling: (weighted) averaging the predictions of multiple models or pipelines Stacking: feed multiple model's predictions into a model