Clear-air turbulence (CAT) is a significant safety issue for aviation at upper levels in the atmosphere. The current CAT forecasting product, Graphical Turbulence Guidance (GTG) system, uses a global optimization technique to produce a forecast from many different indicators, or diagnostics, of turbulence. Theory suggests that many diagnostics vary in their predictive skill depending on geographic region, but GTG is unable to exploit these regional dependencies due to an insufficient number of timely pilot reports (PIREPs). The In-situ Turbulence Reporting System has provided more plentiful and objective observation data in recent years, enabling a regional approach and the opportunity to use machine learning algorithms such as support vector machines, random forests and logistic regression to improve turbulence forecast accuracy. This paper explores the use of domain knowledge developed from CAT climatologies and diagnostic performance studies, as well as machine learning algorithm performance, to guide regionalization decisions. We compare the relevant performance metrics of the different machine learning algorithms for our domains and discuss how best to choose between or combine them to improve forecasting accuracy.