Affine term structure models (ATSMs) are widely used to analyze the dynamics of interest rates and bond prices. However, traditional ATSMs face several challenges, such as small sample size, computational complexity of incorporating numerous macro-financial factors, and use of latent variables that make economic interpretations difficult. To overcome these limitations, this paper proposes a modified ATSM that incorporates machine learning techniques without breaching traditional economic restrictions. Our model offers several advantages over existing ATSMs. First, by incorporating 123 interpretable factors, including five yield factors and 118 macro-financial factors, and their interactions, we show that our empirical model fits yields and predicts future excess returns with fast computation. The larger the number of macro-financial factors, the better the performance. Second, we use LASSO regression combined with principal component analysis to identify 22 key variables that predict yield curve changes. As a result, we find that indices related to the labor market, long-term interest rates, stock market, and supply and demand explicitly covary with the yield curve dynamics. Third, we illustrate how to use data augmentation technique that enhances the model performance. The main contribution of this paper is to propose a novel machine learning approach to ATSMs that mitigates the long-standing challenges of affine models and improves their empirical performance.