Deep Learning-Based Interval Prediction of Stock Returns and Portfolio Selection: A QRLSTM-Selfattention Approach

2026-04-19

PRE

Wei Rao

Yinze Wu *

Fei Xie

Yiqin Zhang

DOI：10.1080/1540496X.2026.2661865

Original

Original request for help

Save

Abstract

En 中文

Accurately quantifying return uncertainty is critical for risk-aware portfolio construction, yet most equity forecasts remain point estimates. This study proposes a hybrid deep learning-based framework, QRLSTM-SelfAttention, which integrates quantile regression (QR) into Long Short-Term Memory (LSTM) networks equipped with a self-attention mechanism to capture long-range temporal dependencies. Using the SSE 50 dataset from 2011 to 2024, we demonstrate that the proposed model consistently outperforms Kernel Density Estimation (KDE), classical QR, and standard deep learning baselines such as Gated Recurrent Unit (GRU), Multilayer Perceptron (MLP), and LSTM across metrics including prediction interval coverage probability (PICP) and interval width. Furthermore, the economic value is validated through a minimax portfolio strategy based on predicted lower quantiles. The results indicate that portfolios constructed with our model achieve superior risk-adjusted returns and lower maximum drawdowns compared to the market index. These findings highlight that combining attention-enhanced deep learning with QR materially improves interval prediction quality and supports robust decision-making in volatile equity markets.

Keywords:

Equity markets

interval prediction

Long Short-Term Memory

self-attention mechanism

portfolio performance

C45

G11

G17

Deep Learning-Based Interval Prediction of Stock Returns and Portfolio Selection: A QRLSTM-Selfattention Approach

Abstract

Journal

Organization