BondBERT: What we learn when assigning sentiment in the bond market

Battle Count: 88

Highly relevant to quantitative trading in fixed income markets. The paper demonstrates that domain-specific sentiment models significantly outperform general financial sentiment tools for bond price prediction. The LSTM forecasting framework with sentiment inputs provides a direct pipeline for signal generation. The Diebold-Mariano test confirms statistical significance of improvements. However, the focus on UK sovereign bonds and the use of synthetic labels somewhat limit immediate production deployment. The work is particularly relevant for bond market trading desks, fixed income risk management, and macro-event-driven strategies.

Moderate complexity. Requires: (1) collecting and cleaning ~30,000 news articles, (2) generating sentiment labels via GPT API with careful prompt engineering, (3) fine-tuning a BERT-based transformer with Optuna hyperparameter search, (4) implementing NDI score conversion, (5) building and training an LSTM forecasting model with rolling sentiment features, (6) running event-based correlation and directional accuracy analyses, (7) deploying FinGPT/Instruct-FinGPT with LoRA adapters on GPU. The pipeline is well-documented but involves multiple components and GPU resources for LLM inference.

Preprocessing scripts, labelling prompts, and model fine-tuning configurations are open-sourced on GitHub. However, the bond transactional data is from a proprietary SaaS provider (Propellant.digital), and the news articles were collected via SerpAPI. The GPT-4.1-nano API used for labelling is a paid service. Hyperparameter search via Optuna is documented. The exact prompt for labelling is provided in the paper.