How can I incorporate both general trends and subcategory-specific trends into a model?

Let's say I am predicting factors that affect import volume, for example. There are many industries which have shared features (average size of warehouse). Overall, warehouse size predicts import volume, but it doesn't predict import volume in consumer electronics manufacturing, and only in farms above a certain acreage. There are many features which have useful general trends, but also other features which have relationships that are dependent on industry.

Let's say I am predicting stock price changes based on the congressional trades. There are many companies, and there are shared important features. For example, certain congressmen are much more predictive than others, and high purchase volume is generally predictive of stock price increase. However, there are many feature relationships that are unique to the company. Congressional trading predicts well only for certain companies, and the dollar value of shares purchased is a much strong predictor for defense companies.

In both cases, there are general trends and subcategory-specific trends we want to incorporate in a model. How can this be done best?

My instict is to use a Bayesian model and allow the general trends to form a priori distributions for the subcategory feature effects. However, pure Bayesian models often simply do not predict as well as, say, a neural network or decision tree-based model.



You must log in to answer this question.