Ensembles and model delivery for tax compliance

Revenue authorities characteristically have large stores of historic audit data, with outcomes, ready for analysis. The Australian Taxation Office established one of the largest data mining teams in Australia in 2004 as a foundation to becoming a knowledge-based organization. Today, every tax return lodged in Australia is risk assessed by one or more models developed through data mining, generally based on historic data. We observe that any of the traditional modeling approaches, particularly including random forests, generally deliver similar models in terms of accuracy. We take advantage of combining different model types and modeling approaches for risk scoring, and in particular report on recent research that increases the diversity of trees that make up a random forest. We also review, in a practical context, how such models are evaluated and delivered.