Optimizing portfolio selection through stock ranking and matching: A reinforcement learning approach

•Utilizing Novel Ensemble architecture for stock prediction.•Reinforced Learning stacked on top of LSTM, Deep RankNet, and XGBoost.•Novel Feature Engineering methods.•Novel combination of hyperparameter optimization and deep learning.•Statistical and Risk assessment of portfolios for managers/invest...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Expert systems with applications Ročník 269; s. 126430
Hlavní autor: Alzaman, Chaher
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 15.04.2025
Témata:
Artificial intelligence
Financial markets
Machine learning
Optimization
Predictive analytics
Financial markets
Artificial intelligence
Predictive analytics
Optimization
Machine learning
ISSN:0957-4174
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:•Utilizing Novel Ensemble architecture for stock prediction.•Reinforced Learning stacked on top of LSTM, Deep RankNet, and XGBoost.•Novel Feature Engineering methods.•Novel combination of hyperparameter optimization and deep learning.•Statistical and Risk assessment of portfolios for managers/investors. Predicting asset movements with machine learning (ML) algorithms remains a complex challenge, particularly in selecting optimal models or designing effective ensemble strategies. This study presents a novel methodology that synergizes reinforcement learning (RL) with advanced ML algorithms—LSTM, XGBoost, and Deep RankNet—to improve prediction accuracy and portfolio construction. The approach incorporates hyperparameter optimization, innovative feature engineering, and a comprehensive comparison of algorithmic performance. RL serves a dual role as both an ensemble strategy and a dynamic learning layer, enabling a 15% increase in cumulative returns compared to traditional ensemble techniques. This advancement highlights RL’s capacity to refine predictions and enhance risk assessment by adaptively integrating outputs from diverse algorithms. Beyond demonstrating superior performance, the study provides actionable insights for practitioners seeking to construct effective, risk-sensitive trading portfolios.
ISSN:0957-4174
DOI:10.1016/j.eswa.2025.126430