Hashami, Romina and Maldonado, Felipe (2026) NarrativeRL: Narrative-Conditioned Reinforcement Learning for Anticipatory Commodity Hedging. In: ICAIF 2026, the 7th ACM International Conference on AI in Finance, 2026-11-14 - 2026-11-17, Milan. (In Press)
Hashami, Romina and Maldonado, Felipe (2026) NarrativeRL: Narrative-Conditioned Reinforcement Learning for Anticipatory Commodity Hedging. In: ICAIF 2026, the 7th ACM International Conference on AI in Finance, 2026-11-14 - 2026-11-17, Milan. (In Press)
Hashami, Romina and Maldonado, Felipe (2026) NarrativeRL: Narrative-Conditioned Reinforcement Learning for Anticipatory Commodity Hedging. In: ICAIF 2026, the 7th ACM International Conference on AI in Finance, 2026-11-14 - 2026-11-17, Milan. (In Press)
Abstract
Commodity markets react rapidly to evolving geopolitical, macroeconomic, and supply-chain narratives, yet traditional volatility forecasting frameworks often fail to capture the semantic structure of financial news. This paper proposes NarrativeRL, an interpretable generative-AI framework for commodity market risk forecasting and adaptive hedging. Using more than 645,000 Reuters oil-market headlines from 2014–2024, we construct a Narrative Signal Index (NSI) combining attention intensity, tone imbalance, semantic novelty, narrative disagreement, geopolitical risk, and supply-shock signals extracted via sentence-transformer embeddings, topic-aware decomposition, and a large language model interpretation layer, and embed these representations into a reinforcement-learningbased hedging module that adapts hedge ratios under transaction costs and changing market regimes. Empirical results show that narrative-aware representations improve realised volatility forecasting relative to HAR-based benchmarks, most clearly at the 22-day horizon (Diebold–Mariano p = 0.039), with narrative effects becoming larger in high-volatility regimes. In out-of-sample hedging experiments, NarrativeRL attains the highest Sharpe ratio, narrowly ahead of volatility targeting, and reduces CVaR95 by 41% relative to the unhedged position, while volatility-targeting and HAR-based hedges attain lower tail risk at the cost of lower returns. NarrativeRL remains ahead of the unhedged benchmark across all tested transaction-cost settings. These findings suggest that interpretable generative AI can support adaptive financial decision systems by transforming evolving market narratives into deployable risk-monitoring and hedging signals.
| Item Type: | Conference or Workshop Item (Paper) |
|---|---|
| Additional Information: | Published proceedings: _not provided_ |
| Uncontrolled Keywords: | reinforcement learning, financial markets, hedging, large language models, text-based signals, volatility, risk management |
| Divisions: | Faculty of Science and Health Faculty of Science and Health > Mathematics, Statistics and Actuarial Science, School of |
| SWORD Depositor: | Unnamed user with email elements@essex.ac.uk |
| Depositing User: | Unnamed user with email elements@essex.ac.uk |
| Date Deposited: | 12 Nov 2026 13:57 |
| Last Modified: | 14 Nov 2026 02:07 |
| URI: | http://repository.essex.ac.uk/id/eprint/44140 |
Available files
Filename: icaif26-67.pdf
Licence: Creative Commons: Attribution 4.0