Decision-informed neural network with large language model integration for portfolio optimization
- Document Type
- Publications · Journal Article
- Registered DB
- SCIEScopus
- Journal
- EXPERT SYSTEMS WITH APPLICATIONS334 · #134389 · 2027
- Citation
- Web of Science 0Scopus 0
- ISSN
- 0957-4174
Authors
- Hwang, YoonTaeFirst authorPusan Natl Univ, Inst Environm & Energy, 2 Busandaehak Ro,63 Beon Gil, Busan 46242, South Korea
- Kong, YaxuanUniv Oxford, Eagle House 7,Walton Well Rd, Oxford OX2 6ED, England
- Zohren, StefanCorresponding authorUniv Oxford, Eagle House 7,Walton Well Rd, Oxford OX2 6ED, England
- Lee, YongjaeCorresponding authorUlsan Natl Inst Sci & Technol UNIST, 50 UNIST Gil, Ulsan 44919, South Korea
Abstract
Portfolio optimization can be highly sensitive to estimation errors even when return forecasts are accurate under conventional prediction metrics. We therefore evaluate forecasts through downstream portfolio regret and realized investment performance rather than assuming that a smaller forecasting loss necessarily produces a better allocation. We propose a Decision-informed Neural Network (DINN), an optimization-aware multimodal architecture that couples a frozen language-model encoder with cross-modal attention and a differentiable constrained portfolio layer. The trainable modules are optimized with a hybrid objective that balances point-forecast accuracy and the regret of the induced portfolio. Experiments on the S&P 100 and DOW 30 universes compare DINN with traditional allocation methods, forecasting models, and a decision-focused baseline. Among the tested methods, DINN attains the highest Sharpe ratio and terminal wealth in both universes, although minimum-variance benchmarks perform better on some absolute-risk measures. Controlled ablations indicate complementary gains from contextual representation and decision-focused training.
Keywords
Portfolio optimizationLarge language modelsDecision-focused learningEstimation errorCOVARIANCES
Identifiers
- Scopus EID
- 2-s2.0-105051332424
- WoS UID
- 001887093300001



Login
Back to Publication list