Outlook & Discussion

The availability of new machine learning tools improves forecasting quality. Although they might be used as an alternative to classical methods, in case of clustered or panel data it is better to combine them with classical tools that excel in dealing with unit-specificity. This may lead to better forecasts that combine the best of both worlds.

A problem remains in understanding the underlying mechanisms of the forecasts. In classical forecasting methods, a fixed model is estimated with statistical evidence on the contribution of several predictors. This provides great explainability and interpretability, since we can exactly indicate which factors matter and to which degree. Moreover, the process of model building can easily be explained, e.g. as illustrated by the gifs earlier in this playbook. For machine learning random forests methods exist to show the contribution of these predictors, which provides some explainability, but these are not as clear as conventional statistics. Moreover, the random draws used in this procedure may lead to slightly different outcomes, which hampers interpretability. Interpreting the underlying model is difficult anyway in a random forest model.

If there are possible societal issues related to the digital twin, it is possible to add a new set of cards, which present societal issues that digital twins might raise. In this round participants are asked to reflect on the value of the various scenarios for society as a whole. It invites participants to reflect as citizens on the future with the digital twin.

Both the classical and the machine learning modeling approaches are supervised. In the classical model, included variables, model structure and estimation approach are chosen. The estimation approach itself does not require any modeler’s choices. In a random forest the modeler also chooses the input variables. In model structure and approach also many choices have to be made, such as decision criteria, number of random samples drawn, number of leaves and branches in trees, etc. This is important to realize in choosing a forecasting approach. Both classical as well as machine learning approaches require modeler’s choices.