When fitting a non-time series linear regression, I know that the observations need to be independent of one another. But how is this handled when DataRobot fits a linear regression with L2 penalty (Ridge) to a time series data, especially when lagged variables are highly correlated with original ones? Also, is there something I need to do or be thinking about?
DataRobot’s repository contains blueprints that handle correlated features robustly, mainly through the use of regularization. So for a ‘linear regression with L2 penalty (Ridge)’, the L2 penalty itself is handling the effect of correlated features. There are other algorithms in the repository that are very robust to highly correlated features e.g. XGBoost. I would also add that DataRobot optimizes for predictive accuracy, not for assumption/inference checking. Independent observations are only a requirement for the latter. If DataRobot finds that several highly correlated features improve the accuracy of the model on the validation set, then it will include them. You are correct that our TimeSeries product will derive many correlated features (and it does one round of feature reduction behind the scenes). It is often worth doing some feature reduction once Autopilot has finished.