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

Battle Count: 178

Highly relevant to quantitative trading in fixed income. The paper directly addresses duration management strategies, provides actionable trading signals based on yield curve forecasts, and evaluates models through portfolio performance metrics (IR, Omega Ratio, MDD). The proposed evaluation framework combining statistical accuracy with economic relevance is directly applicable to trading strategy development. The findings on NN superiority over traditional models and the specific model recommendations (B31 for U.S., N16 for Europe) provide concrete guidance for practitioners.

Moderate to high complexity. Requires implementing multiple model architectures (DNS, PCA, AFNS, AE, NN), a data augmentation pipeline for European data, BOHB hyperparameter optimization, and a duration management trading strategy. The AFNS model requires Kalman Filter estimation. Training involves two-tier strategy (global re-estimation every 2 years + weekly local updates). Python with keras/tensorflow is used. The 43-model comparison framework adds significant computational overhead.

The paper provides detailed model specifications, hyperparameter tables, and data sources (FRED, ECB, LSEG Datastream). However, no GitHub repository is mentioned. The data augmentation procedure for European data is well-documented. The NN architectures and training procedures are described in detail, but the exact code implementation is not publicly available. LSEG Datastream data requires a license.