We re-examine U.S. Treasury bond return predictability using real-time macroeconomic vintages and news-topic attention measures. We propose a Weighted Group Neural Network (WGNN) that imposes pre-specified macroeconomic groupings to mitigate collinearity across predictors. The loss function adopts volatility weighting, making forecast errors comparable across maturities. For non-overlapping excess bond returns, out-of-sample predictability evidence is weak; for overlapping returns, the WGNN has the highest overall out-of-sample ⁠, exceeding 7%, among all models, and is the only model performing well across all maturities. Text-based predictors improve the WGNN forecasts. The performance decline after removing nonlinear activation is consistent with interaction effects or state dependencies that linear models do not capture. Economic evaluation shows that portfolio gains from better statistical forecasts shrink sharply under realistic leverage constraints and financing costs, especially for short-maturity strategies. Our results point to the roles of real-time measurement, regularization, and trading frictions in assessing bond return predictability.