Thick versus thin modeling - An important choice

 

You want to get your prediction model right but that is not an easy task. The usual approach is to formulate a single model or specification and then use estimation techniques to generate forecasts. Nevertheless, we know that when we add or combine different forecasts, we will be able to get a better average will lower error or uncertainty. Simply put, we can form a number of models that may represent different views or hypothesis for specification and average to get a better result. This was developed years ago and given the name of thick modeling.  See "Thick Modeling" by the great econometric thinker Clive Granger. A single specification can be viewed as a thin model. A combination of models is a thick modeling approach. 

Thick model has been exploited in data science through several machine learning techniques. For example, a random forest approach can be viewed as a form of thick modeling.  Choosing a single thin specification may throw-out information from other specification which can be useful with forecasting. Instead of having a goal of forming a single specification, it can be helpful to use several specifications to obtain more information. Each specification may have slightly different features or x variables and thus will have slightly different parameter weights which may prove useful in certain regimes or environments. 

Use all information, take an ensemble approach, and form thick models.