Paper: arXiv 2506.20930
Authors: Chi-Sheng Chen, Xinyu Zhang, Ya-Chuan Chen
Abstract
We propose a hybrid quantum-classical reinforcement learning framework for sector rotation in the Taiwan stock market. Our system employs Proximal Policy Optimization (PPO) as the backbone algorithm and integrates both classical architectures (LSTM, Transformer) and quantum-enhanced models (QNN, QRWKV, QASA) as policy and value networks. An automated feature engineering pipeline extracts financial indicators from capital share data to ensure consistent model input across all configurations. Empirical backtesting reveals a key finding: although quantum-enhanced models consistently achieve higher training rewards, they underperform classical models in real-world investment metrics such as cumulative return and Sharpe ratio. This discrepancy highlights a core challenge in applying reinforcement learning to financial domains – namely, the mismatch between proxy reward signals and true investment objectives. Our analysis suggests that current reward designs may incentivize overfitting to short-term volatility rather than optimizing risk-adjusted returns. This issue is compounded by the inherent expressiveness and optimization instability of quantum circuits under Noisy Intermediate-Scale Quantum (NISQ) constraints. We discuss the implications of this reward-performance gap and propose directions for future improvement, including reward shaping, model regularization, and validation-based early stopping. Our work offers a reproducible benchmark and critical insights into the practical challenges of deploying quantum reinforcement learning in real-world finance.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)
Why this score: The paper employs advanced mathematics, including quantum circuit theory, variational quantum algorithms, and complex reinforcement learning formulations, with detailed derivations in the methodology section. Empirical rigor is high due to backtesting on real Taiwan stock market data with consistent feature engineering, though the authors acknowledge NISQ-era limitations and practical challenges in reward design.
Research Flowchart
flowchart TD
A["Research Goal: Sector Rotation in Taiwan Stock Market using Quantum Reinforcement Learning"] --> B{"Methodology: PPO with Hybrid Architectures"}
B --> C["Automated Feature Engineering<br>Capital Share Data"]
C --> D{"Computational Process<br>Policy & Value Networks"}
D --> E["Classical: LSTM/Transformer"]
D --> F["Quantum: QNN/QRWKV/QASA"]
E --> G["Backtesting & Evaluation"]
F --> G
G --> H{"Key Findings/Outcomes"}
H --> I["Quantum: Higher Training Reward"]
H --> J["Classical: Better Real-World Returns & Sharpe Ratio"]
H --> K["Core Challenge: Reward-Performance Gap & NISQ Instability"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a hybrid quantum‑classical reinforcement learning framework for sector rotation in the Taiwan stock market, using PPO with classical (LSTM, Transformer) and quantum‑enhanced (QNN, QRWKV, QASA) policy/value networks.
Source sentence
We propose a hybrid quantum-classical reinforcement learning framework for sector rotation in the Taiwan stock market. Our system employs Proximal Policy Optimization (PPO) as the backbone algorithm and integrates both classical architectures (LSTM, Transformer) and quantum-enhanced models (QNN, QRWKV, QASA) as policy
Assumes 3
- The agent is rewarded based on whether its selected sector appears in the top‑N ranked sectors by future capital share, which is a proxy for investment performance. model
Source sentence
In our environment, the agent is rewarded based on whether its selected sector appears in the top‑N ranked sectors by future capital share.
- Training and testing are performed on historical daily capital share data with a rolling‑window approach, assuming that past data is representative for future decisions. data
Source sentence
A rolling‑window approach is employed to simulate real‑time decision‑making and ensure temporal consistency.
- Quantum circuits operate under Noisy Intermediate‑Scale Quantum (NISQ) constraints, including noise, shallow depth, and barren plateaus. model
Source sentence
the practical application of QRL is currently constrained by the limitations of Noisy Intermediate‑Scale Quantum (NISQ) hardware, including significant noise, shallow circuit depth, and barren plateau phenomena
Data
2007-04 – 2025-06 · Taiwan stock market · equities · daily
- Taiwan stock market sector‑level capital share data fundamentals unknown
Source sentence
We evaluate our framework on real‑world sector‑level financial data from the Taiwan stock market, comprising 2,646 listed companies categorized into 47 industry sectors
Costs
Not modeled
Finds 4
Sharpe 0.88 · annual return 16.9% · total return 125.3% · max drawdown 32.0%
- Quantum‑enhanced models achieve higher final training rewards but lower cumulative return and Sharpe ratio than classical models. performance
Source sentence
while quantum models frequently achieved higher rewards during training, they consistently underperformed classical models on real‑world investment metrics like cumulative return and the Sharpe ratio.
- QASA obtains the highest training reward (2883.30) yet the lowest cumulative return (93.14%) among all models. performance
Source sentence
QASA [13] 2883.30 93.14% 13.49% 19.03% 0.71 -31.73%
- The discrepancy between proxy reward and investment performance suggests overfitting of quantum models to short‑term volatility. negative
Source sentence
This discrepancy highlights a core challenge in applying reinforcement learning to financial domains -- namely, the mismatch between proxy reward signals and true investment objectives.
- Sharpe ratios are similar across models (~0.7‑0.88), indicating return generation as the primary differentiator. performance
Source sentence
The similarity in Sharpe ratios and drawdown across models suggests that the primary differentiator is return generation rather than risk control.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 37 related papers). Coverage is partial, so this is neither replication nor refutation.
3 papers in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (14)
- An analysis of the effect of noise in a heterogeneous agent financial market model (2011)
- Reinforcement learning: A survey (1996)
- Quantum machine learning (2017)
- Quantum reinforcement learning (2008)
- Variational quantum circuits for deep reinforcement learning (2020)
- Quantum architecture search via deep reinforcement learning (2021) — link
- Variational quantum reinforcement learning via evolutionary optimization (2022)
- Proximal policy optimization algorithms (2017) — link
and 6 more
0 of 14 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.