Near infrared spectroscopy (NIRS) is widely used to predict the potential nutritive value of feeds. Near infrared spectroscopy calibrations based on a combination of feed and feces spectra may reflect the effective nutritive value. We investigated the usefulness and accuracy of NIRS based on combined spectra from pig feed and feces to predict their chemical composition as well as nutrient digestibility and net energy (NE) content. A total of 62 feeds and 310 freeze-dried feces samples (5 per feed) from 3 in vivo digestibility experiments were used. First, calibrations based on either feed spectra or feces spectra only were developed to predict their chemical composition. Then, calibrations based on combined feed and feces spectra were developed, thereby comparing 3 possibilities: merging, subtracting or averaging spectra. The calibrations were evaluated by cross-validation either by leave-one-feed-out or by splitting the dataset in 4 random groups. Most of the chemical parameters of the feed and the feces could be predicted accurately (residual prediction deviation, RPD ~ 3, R2 > 0.8), but predictions for feed were somewhat less accurate for crude protein (RPD = 2.1), non-starch polysaccharides (RPD = 2.0) and sugar (RPD = 1.0); and for feces, crude fiber (RPD = 2.1), organic matter (RPD = 2.2) and gross energy (RPD = 1.8). Net energy was better estimated using feed spectra than feces spectra, with a standard error of cross validation (SECV) of 0.33 and 0.46 MJ/kg, respectively. The digestibility of nutrients was poorly estimated from both calibrations based on feed or feces spectra alone (RPD ~ 1.5). The combination of spectra resulted in an overall better estimation of the digestibility and NE especially with merging and subtracting; for NE, these combinations resulted in an SECV of 0.26 and 0.27 MJ/kg, respectively. For all calibrations based on feces spectra, either singular or combined, the leave-one-feed-out cross-validation resulted in higher SECV than the validation with 4 random groups. However, the latter cross-validation method may lead to over-optimistic results as the group to be validated may contain spectra which are not independent.