Deep Learning for U.S. Bond Yield Forecasting: An Enhanced LSTM–LagLasso Framework

This paper advances a decision-aligned post-processing layer for government bond yield forecasts, turning competent sequence predictions into curve-consistent and economically calibrated outputs with minimal engineering burden. Starting from capacity-fair baselines in the LSTM, GRU and compact transformer families, used only to generate initial point forecasts for five, ten and thirty year maturities at short horizons, we add two model-agnostic stages. A curve consistency projection enforces monotone ordering across maturities and, when warranted, mild convexity while preserving local signal. An asymmetric economic calibration then learns a monotone mapping that down-weights the costlier side of error in basis points and in price space via duration and convexity. Rather than a perfectly linear workflow, we report practical adjustments such as solver choices for the projection and calibration folds for stability. Evaluation considers violation rates, smoothness and decision-weighted loss, and probes weakly coupled transfer from ten year forecasts to five and thirty year using rolling linear links without retraining. Results indicate lower violation rates and reduced economic loss to some extent across horizons, though gains can depend on regimes and may partly reflect calibration rather than new information. Alternative explanations including liquidity frictions or structural breaks remain plausible, and further research is needed on denser tenor grids, portfolio utilities and additional markets.