We propose an alternative approach to the linear factor model to estimate and decompose asset risk premia in empirical asset pricing. To resolve the high-dimensional sort difficulty in forming characteristic-based benchmark portfolios, we introduce a benchmark combination model (BCM) that combines multiple basis portfolios as the pricing kernel. With a non-arbitrage objective, our approach minimizes cross-sectional pricing errors and identifies the sources of risk through the combination of basis portfolios. For a 40-year sample for U.S. corporate bonds, we find that BCM outperforms standard factor models in pricing corporate bonds. Second, our study shows that credit ratings, downside risk, and short-term reversal are primary sources of bond risk premia. Finally, incorporating machine learning forecasts into the BCM, we find strong evidence of return predictability.