Uncertainty-aware stock selection and portfolio optimization with deep learning-based return prediction
ANNALS OF OPERATIONS RESEARCH, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s10479-026-07448-8
- Dergi Adı: ANNALS OF OPERATIONS RESEARCH
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, ABI/INFORM, Aerospace Database, Applied Science & Technology Source, INSPEC, MathSciNet, Public Affairs Index, zbMATH, Academic Search Ultimate (EBSCO), Business Source Ultimate (EBSCO), Engineering Source (EBSCO), Materials Science & Engineering Collection (ProQuest), Pharma Collection (ProQuest), Technology Collection (ProQuest)
- Kayseri Üniversitesi Adresli: Evet
Özet
Portfolio optimization allocates capital to balance return and risk. Mean-variance optimization (MVO), its foundational framework, is highly sensitive to expected-return estimation errors. Recent studies increasingly apply machine and deep learning to improve return estimation for portfolio construction. However, despite the inherent uncertainty of financial return prediction, most prediction-driven approaches treat forecasts as equally reliable point estimates, overlooking the model's confidence in individual predictions. To address this gap, we propose an uncertainty-aware portfolio-construction framework in which a long short-term memory (LSTM) network with Monte Carlo dropout generates return forecasts and forecast-specific epistemic uncertainty estimates. These estimates guide stock selection through ranking and screening and adjust the expected-return inputs to an uncertainty-aware MVO (UA-MVO). The framework is evaluated through a rolling out-of-sample analysis using 100 large-cap, highly liquid S&P 500 stocks over 2018-2023. Empirical results show that epistemic uncertainty provides a forward-looking signal of forecast reliability, as higher uncertainty is associated with larger subsequent forecast errors across months and volatility levels, while lower-uncertainty forecasts exhibit smaller average errors. Incorporating this information into portfolio construction, UA-MVO generally achieves higher returns, better risk-adjusted performance, and lower downside risk than benchmark allocations. Comparisons with uniform shrinkage, volatility-based adjustment, and robust optimization suggest that the gains arise from forecast-specific reliability rather than common shrinkage, historical volatility, or robust penalties. These findings remain robust across portfolio sizes and parameter settings. Overall, the findings suggest that incorporating forecast-specific model confidence into prediction-driven portfolio decisions can improve portfolio performance relative to approaches that rely solely on point forecasts.