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

In one sentence

Summary

The paper addresses the critical gap in financial NLP where general sentiment models, trained on equity-centric data, fail to capture the inverse relationship between macroeconomic news and bond prices. While equity markets often rise on economic optimism, bond prices typically fall due to rising yield expectations. The authors introduce BondBERT, a domain-adaptive transformer designed to act as a perception module for financial decision-support agents, specifically targeting fixed-income dynamics where sentiment signals are noisy and effect sizes are small. The research positions BondBERT as a solution to the 'wrong-sign' bias inherent in models like FinBERT when applied to fixed income. Methodologically, the authors compiled and cleaned 30,000 UK bond market articles from 2018–2025, using GPT-4.1-nano with specialized prompts to generate continuous sentiment labels in the [-1, 1] range. BondBERT was created by fine-tuning the FinBERT architecture on this bond-specific dataset, while baselines included ProsusAI/FinBERT, FinGPT (LLaMA-2-7B with LoRA), and Instruct-FinGPT (LLaMA-2-13B). Experiments evaluated sentiment alignment via event-based correlation on sentiment shock days, directional accuracy of next-day returns, and predictive power using an LSTM forecasting model across ten liquid UK sovereign bonds. Metrics included Pearson correlation, normalised RMSE, Information Coefficient (IC), and Diebold-Mariano tests. Results demonstrate that BondBERT consistently produces positive correlations with bond returns, reversing the negative bias seen in FinBERT. It achieved the highest directional accuracy, significantly outperforming the random baseline, while FinGPT and Instruct-FinGPT performed poorly. In LSTM forecasting, BondBERT yielded the lowest normalised RMSE (0.0079 vs 0.0086 for FinBERT) and highest IC (0.80 vs 0.745). Diebold-Mariano tests confirmed statistical significance for several bonds. Limitations include reliance on synthetic GPT-generated labels which may introduce noise or look-ahead bias, and the restriction to liquid UK sovereign bonds, excluding corporate bonds due to illiquidity. The work highlights the necessity of domain-specific adaptation for LLMs in fixed-income markets.

Research problem

General financial sentiment models like FinBERT are trained on equity-centric data and often misinterpret macroeconomic news in bond markets, where economic optimism typically leads to falling bond prices (rising yields). This 'wrong-sign' bias limits their utility for fixed-income analysis. The paper addresses the lack of sentiment tools specifically adapted to the inverse sentiment-price relationship and low-volatility nature of bond markets.

Core contributions

• Introduction of BondBERT, the first transformer-based sentiment model fine-tuned specifically for bond market news.
• Development of a framework for adapting transformers to domains with inverse sentiment-price relationships using synthetic labeling.
• Creation of a curated dataset of 30,000 UK bond market articles (2018-2025) with GPT-4.1-nano generated sentiment labels.
• Comprehensive benchmarking of BondBERT against FinBERT, FinGPT, and Instruct-FinGPT using correlation, directional accuracy, and LSTM forecasting.
• Demonstration that domain-specific adaptation significantly reduces the 'wrong-sign' bias and improves predictive accuracy for sovereign bonds.

Method

• Data Collection: Scraped 45,000 articles via SerpAPI, cleaned for duplicates and relevance, resulting in 30,000 high-quality articles.
• Label Generation: Used GPT-4.1-nano with a specific prompt ('You are a financial analyst...') to generate continuous sentiment scores [-1, 1].
• Model Architecture: Fine-tuned FinBERT (ProsusAI) on bond-specific data to create BondBERT.
• Sentiment Scoring: Converted class probabilities to a Normalised Difference Index (NDI) to amplify polarity and remove neutral class noise.
• Hyperparameter Tuning: Used Optuna to optimize learning rate, batch size, and dropout for BondBERT.
• Baseline Implementation: Deployed FinGPT and Instruct-FinGPT using LLaMA-2 backbones with appropriate adapters/instructions.
• Forecasting: Trained LSTM models using rolling histories of bond prices and daily sentiment scores as inputs.
• Statistical Testing: Applied Diebold-Mariano tests with Newey-West corrections to compare forecast accuracy.

Results & findings

• BondBERT shows consistent positive correlation with bond returns, whereas FinBERT shows negative or neutral correlations on most topics.
• BondBERT is the only model to consistently outperform the 50% random baseline in directional accuracy for next-day bond returns.
• FinGPT and Instruct-FinGPT performed poorly, with directional accuracy well below the random baseline.
• BondBERT achieved the lowest normalised RMSE (0.0079) and highest Information Coefficient (0.80) in LSTM forecasting compared to baselines.
• Diebold-Mariano tests rejected the null hypothesis for several bonds, confirming BondBERT's statistical superiority in forecasting accuracy.
• The study confirms that general financial models suffer from 'wrong-sign' bias in fixed-income contexts due to inverse price-yield dynamics.

Data & experiments

• Dataset: 30,000 cleaned UK bond market articles (2018-2025) sourced via SerpAPI, with sentiment labels generated by GPT-4.1-nano.
• Market Data: Daily transactional prices and log returns for 10 liquid UK sovereign bonds (Jan 2022-Jan 2025) from Propellant.digital.
• Baselines: ProsusAI/FinBERT, FinGPT (LLaMA-2-7B + LoRA), and Instruct-FinGPT (LLaMA-2-13B).
• Evaluation Tasks: Event-based correlation on sentiment shocks, directional accuracy of next-day returns, and LSTM price prediction.
• Metrics: Pearson correlation, Normalised RMSE, Information Coefficient (IC), and Diebold-Mariano statistical tests.
• Experimental Design: Train/validation split on 2018-2021 data; test on unseen 2022-2025 market data to avoid leakage.

Limitations & caveats

• Reliance on GPT-4.1-nano for label generation introduces potential noise and bias.
• Risk of look-ahead bias in synthetic labels, as GPT models may leverage post-event knowledge.
• Analysis restricted to liquid UK sovereign bonds; corporate bonds excluded due to illiquidity.
• Synthetic labels may not fully capture domain expertise compared to human annotation.
• Baselines (FinGPT/Instruct-FinGPT) depend on older LLaMA-2 backbones, limiting comparison with newer models.
• Event-based correlation uses a 7-day rolling window, which may smooth out short-term market reactions.

Curator notes

Discovery status: round3_promoted_seed_for_round4. Source links: A Comparative Analysis of Instruction Fine-Tuning Large Language Models for Financial Text Classification; FinBERT : A Large Language Model for Extracting Information from Financial Text†; FinGPT: Democratizing Internet-scale Data for Financial Large Language Models; Instruct-FinGPT: Financial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models; Large Language Models in equity markets: applications, techniques, and insights