Adaptive Deep Sequential Learning for Treasury Yield-Curve Forecasting: Integrating Affine Term-Structure Factors and Experience Replay

Accurate out-of-sample yield-curve forecasting requires models that capture complex non-linear temporal dynamics while respecting the cross-sectional economic structure of the term structure. This paper introduces an economically structured deep sequential learning framework for multi-horizon Treasury yield forecasting that unifies multi-head self-attention, temporal convolutions, bidirectional recurrent states, and latent affine term-structure factors. To address the severe optimization degradation caused by strong temporal autocorrelation during gradient descent, we introduce adaptive experience replay dynamics—specifically comparing uniform, prediction-error-based prioritized, and structural yield-entropy sampling regimes—under strict chronological out-of-sample evaluation protocols. Empirical results demonstrate that experience replay mechanisms significantly outperform traditional chronological training regimes. Structurally, entropy-based sampling achieves optimal performance in aggregate yield forecast accuracy and latent factor estimation, while prioritized experience replay excels in absolute point error minimization and conditional volatility tracking. Crucially, incorporating cross-sectional no-arbitrage constraints prevents structural yield-curve inversions and monotonicity violations without compromising predictive precision. Statistically superior point forecasts translate directly into enhanced economic tracking fidelity for duration-based bond portfolio pricing. Overall, the findings demonstrate that unifying structural term-structure constraints with non-uniform historical memory sampling provides a powerful, economically coherent methodology for fixed-income modeling and financial risk management.