How Hidden Layers Affect Deep Reinforcement Learning for Stock Portfolios
Summary
This study examines how neural network hidden-layer configurations affect reinforcement-learning performance in stock portfolio optimization. Using the FinRL framework, agents receive open, high, low, close, and volume data and seek to improve portfolio outcomes. The experiment covers 45 stocks from Indonesia’s LQ45 index and compares A2C, DDPG, PPO, and TD3 using cumulative and annualized returns, drawdown, Sharpe ratio, and other portfolio measures.
Results vary by algorithm. A2C and DDPG perform best without hidden layers on several reported measures, while PPO improves with one hidden layer and TD3 remains relatively stable. The summary reports specific return and Sharpe figures, but provides little detail on the sample period, validation design, transaction-cost assumptions, or statistical uncertainty. The results suggest architecture should be tuned for each algorithm; they do not establish that a particular configuration will work in other markets or periods.
Key ideas
- The experiment tests how hidden-layer structure affects several deep reinforcement-learning algorithms for portfolio allocation.
- Agents use open, high, low, close, and volume data for 45 Indonesian stocks.
- A2C and DDPG perform strongly without hidden layers, while PPO improves with one hidden layer.
- Adding network depth does not consistently improve results, and the reported findings lack detailed robustness information.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.