This article addresses the design and evaluation of virtual NOx sensors for heavy-duty off-road diesel engines based on static polynomial black box modeling. Three approaches, differing in the chosen sets of regressors, are analyzed regarding their NOx prediction capability. As regressor sets only quantities available on standard production-type electronic control units, features extracted from the in-cylinder pressure trace via singular value decomposition extended by the engine speed as well as geometric values from the pressure trace and heat release curves are utilized, respectively. It is shown that while each of the approaches alone has its drawbacks, a systematic combination, especially of the first two methods, results in high-accuracy NOx models while at the same time keeping the number of regressors low.