Data-Driven Duration Management -- Term Structure Forecasting Using Machine Learning

Abstract: This paper compares different methods for forecasting the term structure of U.S. and European zero-coupon government bonds using both traditional econometric and Machine Learning (ML) approaches. We compare classical models (e.g., Dynamic Nelson-Siegel (DNS) and Principal Component Analysis (PCA)) with different Neural Network (NN) architectures, including those inspired by the classical models, on the U.S. Treasury market and bonds issued by the European Central Bank (ECB). To enhance predictive performance, macroeconomic variables are incorporated. The findings for both markets are separately analyzed and compared. To this end, we propose a robust model evaluation framework combining statistical accuracy metrics - such as RMSE, MAE, and directional accuracy - with the economic relevance of a quantitative bond trading strategy. Results show that NNs consistently outperform traditional models in both forecasting accuracy and portfolio performance. For the U.S., the most effective approach is a direct-forecasting NN that incorporates DNS factors to reduce the dimensionality of zero-rate data and an Autoencoder (AE) to extract macroeconomic features, while for Europe, the optimal model is a factor-based NN using PCA-derived zero-rate factors without the integration of macroeconomic variables. Overall, the paper demonstrates how combining traditional modeling approaches with modern ML techniques and evaluation can improve yield curve forecasts and support applications in fixed-income portfolio construction.

• The paper demonstrates that NN-based models significantly outperform classical methods in forecasting bond yield curves and managing duration.
• The study utilizes advanced dimensionality reduction techniques like PCA, DNS, and autoencoders with macroeconomic augmentation to optimize model performance.
• Results reveal market-specific optimal configurations, with macroeconomic features proving crucial for U.S. bonds but less impactful for European bonds.

Data-Driven Duration Management: Machine Learning Approaches to Term Structure Forecasting

Introduction

The paper "Data-Driven Duration Management: Term Structure Forecasting Using Machine Learning" (2606.26815) presents a comprehensive methodological investigation of term structure forecasting for U.S. and European zero-coupon government bonds using both classical econometric and modern ML approaches, including advanced dimensionality reduction and neural network (NN) architectures. The primary focus is the efficient and robust modeling of the yield curve for both statistical and practical (i.e., portfolio management) purposes, with particular attention given to macroeconomic input integration, model stability, and the economic value of predictions. The empirical analysis is underpinned by extensive historical datasets, including a novel synthetic extension of pre-2004 European Central Bank (ECB) data via German bonds, enhancing the training horizon available for ML models.

Data, Preprocessing, and Dimensionality Reduction

The study employs weekly zero-rate data for seven key maturities (3m, 6m, 1y, 2y, 3y, 5y, 10y) over several decades for the U.S. and Europe. For U.S. data, long time series are available directly. In contrast, European AAA-rated zero-rate curves are available only from 2004; the authors use OLS-driven regression on German proxies to backcast a synthetic Euro area series to 1992, resulting in close statistical fit ($R^2 > 0.99$ across all maturities). The approach robustly addresses missing values (notably in German short-term maturities) using Dynamic Nelson-Siegel (DNS)-based interpolation—validated empirically to deliver high $R^2$ for all but pre-1994 subsegments.

Figure 1: U.S. zero-rate curve between April 1987 and February 2025.

Figure 2: Euro triple A zero-rate and German zero-rate curves between February 1992 and February 2025, with pre-2004 Euro data artificially extrapolated from German rates.

Dimensionality reduction for the zero-rate curve is addressed via several established and novel techniques. The classical DNS model (level, slope, curvature), PCA, arbitrage-free Nelson-Siegel (AFNS), and nonlinear autoencoder (AE) approaches are all benchmarked. For PCA and DNS, the shape of the term structure factors aligns well with economic intuition.

Figure 3: DNS factor loadings for different times to maturity.

Figure 4: PCA factor loadings for different times to maturity for U.S. (left) and Europe (right).

AEs are also used both for zero-rate curve and macroeconomic variable compression, with network architectures ranging from shallow (7-3-7) to deeper (7-5-3-5-7). Linear AEs recover PCA, while nonlinear AEs can capture higher-order dependencies, though in this empirical setting the main benefit lies in macroeconomic feature extraction.

Figure 5: Single-layered autoencoder architecture used for factor extraction.

Figure 6: Normalized latent factors fitted for U.S. data using DNS, PCA, AE, and AFNS, demonstrating convergence of factor representations.

Model Architectures

Two primary ML model design paradigms are considered: (1) forecast latent factors and reconstruct the zero-rate curve; (2) directly forecast the zero-rate curve or individual rates. Both architectures benefit from macroeconomic augmentation and are implemented using NNs, yielding direct and factor-based forecasting variants for each data reduction approach.

Figure 7: Model architecture for NN-based forecasting of latent factors, with macroeconomic conditioning and zero-rate curve reconstruction.

Figure 8: Model architecture for direct NN-based zero-rate forecasting, bypassing factor reconstruction.

The macroeconomic features (for the U.S.: multiple measures of employment, inflation, industrial production, and credit spreads; for Europe: German proxies) are transformed using annual log growth rates and, where applicable, further compressed by a dedicated AE. The models allow both raw and compressed macro input.

Network training entails a two-tier schedule: periodic full retraining (every two years) with fresh random initialization—mimicking genuine out-of-sample re-specification—and incremental weekly updates, aligning with practical asset management workflows.

Evaluation Framework

The paper employs a dual-assessment framework integrating both statistical (RMSE, MAE, directional accuracy) and investment-driven measures. For investment performance, model-driven duration management is evaluated via a proprietary allocation rule, adjusting portfolio duration based on forecasted curve changes and constraining duration within [2.5, 7.5], benchmarked against a passive duration = 5 strategy. Portfolio construction utilizes U.S. Treasury and Euro AAA indices, with instruments selected to minimize spread and liquidity risk. Economic relevance is measured through Information Ratio (IR), Omega Ratio, and Maximum Drawdown (MDD), thus capturing both mean-variance-efficiency and downside risk asymmetry.

Results: Empirical Performance and Model Selection

Hyperparameters for AEs and NNs are optimized through Bayesian Optimization and Hyperband (BOHB), with a validation framework partitioning data into pre-2005 training, 2005–2014 validation, and post-2014 test sets, ensuring chronologically plausible generalization.

The systematic model selection procedure filters 43 candidate models (across compression, macro input, and prediction paradigm) using simultaneous thresholds on all performance metrics. Boxplots of the test phase results confirm that NN-based models consistently outperform their econometric counterparts in both predictive and investment terms.

Figure 9: Box plots of U.S. results: RMSE, MAE, directional accuracy, IR, Omega, MDD.

Figure 10: Box plots of Europe results: RMSE, MAE, directional accuracy, IR, Omega, MDD.

For the U.S., the optimal model is a direct zero-rate forecasting NN using AFNS for compression and macroeconomic input compressed by an AE (“B31”). For Europe, the optimal is a factor-based NN using PCA-compressed zero rates and no macroeconomic input (“N16”). Notably, these top models demonstrate not only superior mean performance but also strong phase stability across various periods (falling/rising rates, crisis regimes, volatile environments).

Figure 11: Model B31's duration allocation for the U.S., index yields, and yield changes, highlighting active outperformance versus benchmark.

Figure 12: Model N16’s duration allocation for Europe, index yields, and yield changes, highlighting well-timed allocation decisions.

Further comparisons between classical time-series models (AR, VAR on DNS/PCA factors) and their NN-based analogs show that the former, despite decent RMSE/MAE, are largely inertial—essentially shifting the yield curve rightward in time with little anticipation of inflection points, thus generating weak trading signals.

Figure 13: U.S. 3-month zero-rate predictions using traditional AR/VAR models on PCA factors; forecasts are lagged, exhibiting limited dynamic adaptation.

Figure 14: NN-based one-month-ahead forecasts of the U.S. 3-month zero-rate; notable improvements in dynamic regime adaptation during rising and falling rate regimes.

Discussion and Implications

Several substantive empirical findings are established:

• NNs robustly outperform classical models in both accuracy and economic value metrics, with balanced performance across both direct and factor-based paradigms.
• Optimal model configurations differ by market: For the U.S., macroeconomic augmentation via AE is critical; for Europe, macro factors confer little additional value, possibly reflecting the lower macro sensitivity or higher inertia/sovereign stability of Euro AAA rates.
• Nonlinear feature extraction via AE is valuable primarily for macro variables; for seven-maturity curves, classical PCA/DNS factorization suffices, but larger/more granular maturity panels may benefit more from AE compression.
• Model selection based solely on statistical error (RMSE/MAE) is insufficient; economic significance (e.g., IR, drawdown management) is only revealed in practical allocation outcomes.
• Phase stability and out-of-sample robustness are critical; the best models avoid large negative tail events, with performance gains concentrated in transition or regime-shift periods.

Theoretically, these findings reinforce that generic linear time-series models are inadequate for the nonstationary, nonlinear dynamics now dominant in post-GFC, crisis, and unconventional policy regimes. The flexibility of NN-based approaches enables adaptation to macro-financial nonlinearity and latent regime change. The research further demonstrates that macroeconomic information—if effectively preprocessed—remains valuable (especially in U.S. settings), but the optimal information sets are likely region- and regime-dependent.

Practically, the paper’s unified statistical-economic evaluation protocol provides a template for applied term-structure researchers and fixed-income portfolio managers seeking to avoid overfitting on in-sample benchmarks. The findings have direct relevance for risk management, duration overlay, dynamic hedging, and scenario analysis.

Future Directions

Outstanding questions remain regarding:

• Application to higher-dimensional or cross-market term structure panels (e.g., multi-country, multi-credit-segment curves).
• Model extensions using richer macro data (textual, high-frequency, financial stress indicators).
• More dynamic, volatility-sensitive retraining and ensemble strategies.
• Exploration of alternative ML architectures (e.g., LSTM, transformers) and regularization protocols.

Conclusion

This study comprehensively demonstrates that neural network-based approaches—particularly when coupled with effective dimensionality reduction and macroeconomic feature engineering—substantially improve both the accuracy and actionable economic value of term structure forecasting for U.S. and European government bond markets. These results hold under rigorous validation, robust model selection, and real-world portfolio strategy evaluation frameworks, providing a foundation for future methodological and applied advances in yield curve modeling and fixed income risk management.