| dc.description.abstract | Financial technology (FinTech) has emerged as one of the key areas for the application of artificial intelligence (AI), including but not limited to, the prediction of stock
market movements and asset allocation. However, relying solely on stock price forecasting does not guarantee the maximization of investment returns. An investor also needs to
consider asset allocation strategies to either maximize the returns or minimize the losses.
In such a scenario that requires interactions with the environment to reap rewards, reinforcement learning (RL) emerges as an ideal solution. Consequently, in this study, we
propose a strategy for stock investment that employs the Actor-Critic techniques of RL.
To enhance the effectiveness of investment decisions, we employed an AutoEncoder (AE) to learn the features of various technical indicators in stocks, which then aids in making
decisions on stock allocation and return estimation.
However, portfolio management still faces challenges in optimizing allocation strategies and accurately forecasting returns, especially during market volatility. Traditional
strategies often focus on either short-term or long-term investments, leading to a lack of a model in the market that can adapt flexibly to various situations. To address this
problem, we have introduced a novel method that combines reinforcement learning and autoencoders, hoping to fill this gap in the market.
We employed ablation experiments to explore the effects of the dimension of AutoEncoder encoding and the length of historical data on state encoding. The results indicate that by compressing the past 30 days of historical data into five dimensions, the optimal state encoding effect can be achieved. We also discovered that incorporating the
prediction results of the AutoEncoder Predictor can enhance cumulative earnings. Furthermore, we investigated three different investment strategies: RL+AE Predictor, RL
Only, and AE Predictor. Through performance analysis, correlation with the broader market, and error rate analysis, we evaluated the performance of these three strategies in
various market environments.
Experimental results reveal that as a constrained investment strategy, RL+AE Predictor performs the best in maximizing assets and exhibits a stable learning process. Especially during significant market changes, this strategy showcases superior risk resistance and can maintain stable investment returns. Moreover, this strategy has a lower correlation coefficient with the broader market, indicating its independence from market index volatility. In the error rate analysis, the RL+AE Predictor model has a False Positive
Rate (FPR) of 6.46%, which is lower compared to AE Predictor at 38.38% and RL Only at 36.09%. This shows that it outperforms in predicting stock asset allocation, with the lowest error rate.
We validated this method in Taiwan’s stock market environment, conducting experiments with Taiwan stock data from 2019 to 2021 and compared it with the TW50
Index, traditional portfolio theory (Mean-variance optimization, MVO), and Jiang’s
research that uses reinforcement learning Policy Gradient techniques. The experimental results show that the win rate of this study in short-term (3 months), mid-to-long term
(6-9 months), and long-term (1-2 years) investment periods is superior to the benchmarks
TW50, Jiang’s, and MVO, reaching the highest total return in the 12 and 24 month lognterm investment periods. Even in the two-year fixed investment time comparison, starting
at two different investment points, the bull market of 2019 and the bear market of 2020,
the method proposed in this thesis still outperforms the TW50 Index, MVO, and Jiang’s.
In summary, this research provides empirical evidence that a combination of reinforcement learning and autoencoders in portfolio management outperforms the traditional
MVO, TW50 Index, and Jiang’s hybrid deep learning methods in both cumulative return rate and Sharpe ratio. It highlights the potential of AI in complex financial decisions and
points out the need for a more flexible, universal model to bridge the gap between shortterm and long-term investment strategies. These research findings provide significant
reference value for the development and improvement of investment strategies | en_US |