Near-infrared spectroscopy (NIRS) has been widely used for feed analyses. Traditionally, partial least squares (PLS) has been used as the statistical method for prediction using spectral data. However, only linear relationships between spectral wavelengths and response variables are explored in PLS models. The objective of this study was to evaluate the ability of artificial neural networks (ANN) to generate accurate predictions of feed components using NIRS data. Feed samples were scanned in a NIRS instrument and analyzed with wet chemistry for: acid detergent fiber (ADF, n = 16,032), neutral detergent fiber (NDF, n = 19,942), ash (n = 11,165), calcium (Ca, n = 50,372), potassium (K, n = 50,438), water soluble carbohydrate (WSC, n = 1,982), mannitol (MAN, n = 2,836), starch (STA, n = 18,694) and butyric acid (BA, n = 12,763). A Bayesian Network approach was implemented to investigate statistical dependences among wavelengths for selection of more informative predictors using the concept of Markov Blanket (MB). The use of MB for variable selection was combined with 2 predictive approaches: PLS and ANN. Prediction quality was assessed by randomly splitting the data set into training and test sets (70% and 30% of the data set, respectively). The ANN and PLS results using all wavelengths (ANN_all and PLS_all) were similar across feed components. The use of MB coupled with ANN (ANN_mb) yielded more accurate and precise predictions compared with MB with PLS (PLS_mb). ANN_mb performed better than the traditional approach (PLS_all) in terms of R² and root mean square error of prediction (RMSEP) for: mannitol (ANN_mb: R² = 0.71, RMSEP = 0.71%, PLS_all: R² = 0.25, RMSEP = 1.7%), WSC (ANN_mb: R² = 0.73, RMSEP = 5.1%, PLS_all: R² = 0.57, RMSEP = 7.3%), Ca (ANN_mb: R² = 0.71, RMSEP = 0.82%, PLS_all: R² = 0.57, RMSEP = 0.92%), BA (ANN_mb: R² = 0.71, RMSEP = 0.42%, PLS_all: R² = 0.64, RMSEP = 0.86%), and similar for: STA (ANN_mb: R² = 0.97, RMSEP = 3.7%, PLS_all: R² = 0.94, RMSEP = 5.0%), ADF (ANN_mb: R² = 0.92, RMSEP = 4.0%, PLS_all: R² = 0.88, RMSEP = 5.0%), and NDF (ANN_mb: R² = 0.90, RMSEP = 4.8%, PLS_all: R² = 0.87, RMSEP = 5.4%). In summary, combining MB with ANN is an effective way to reduce data dimension and increase quality of NIRS-based predictions.